Outage Prediction for Ultra-Reliable Low-Latency Communications in Fast Fading Channels

Page created by Audrey Arnold
 
CONTINUE READING
Traßl et al.

            RESEARCH

        Outage Prediction for Ultra-Reliable Low-Latency
        Communications in Fast Fading Channels
        Andreas Traßl1,2* , Eva Schmitt1 , Tom Hößler1,3 , Lucas Scheuvens1 , Norman Franchi1 , Nick
        Schwarzenberg1 and Gerhard Fettweis1,2,3

            Abstract
              The addition of redundancy is a promising solution to achieve a certain Quality of Service (QoS) for
            ultra-reliable low-latency communications (URLLC) in challenging fast fading scenarios. However, adding more
            and more redundancy to the transmission results in severely increased radio resource consumption. Monitoring
            and prediction of fast fading channels can serve as the foundation of advanced scheduling. By choosing
            suitable resources for transmission, the resource consumption is reduced while maintaining the QoS. In this
            article, we present outage prediction approaches for Rayleigh and Rician fading channels. Appropriate
            performance metrics are introduced to show the suitability for URLLC radio resource scheduling. Outage
            prediction in the Rayleigh fading case can be achieved by adding a threshold comparison to state-of-the-art
            fading prediction approaches. A LOS component estimator is introduced that enables outage prediction in line
            of sight (LOS) scenarios. Extensive simulations have shown that under realistic conditions, effective outage
            probabilities of 10−5 can be achieved while reaching up-state prediction probabilities of more than 90 %. We
            show that the predictor can be tuned to satisfy the desired trade-off between prediction reliability and
            utilizability of the link. This enables our predictor to be used in future scheduling strategies, which achieve the
            challenging QoS of URLLC with fewer required redundancy.
            Keywords: channel prediction; URLLC; radio resource scheduling

        1 Introduction                                                                         could cause damage or even human harm. Latency-
        One of the main pillars of the fifth generation (5G) mo-                               critical mobile connectivity is also required when hu-
        bile broadband standards is ultra-reliable low-latency                                 mans are involved in the control loop [4, 5]. In indus-
        communications (URLLC), which aims to provide ex-                                      try, this is the case, e.g, in teleoperating applications or
        tremely high service availabilities paired with latency                                during installation of machines, where a human could
        values of only a few milliseconds. To realize even more                                train the machine instead of programming it. In the
        ambitious quality of service (QoS) requirements com-                                   future, URLLC might also find its way to consumer
        pared to 5G, URLLC inevitably has to play a key role                                   products for entertainment, for health or even for hu-
        also during research of the sixth generation (6G) mo-                                  man learning.
        bile broadband standards [1, 2].                                                         The major challenges for URLLC are the imperfec-
          The ongoing development of URLLC is driven by a                                      tions of the wireless link, especially the fast fading of
        wide variety of applications. In recent yeas, many of                                  the channel. Due to reflections in the environment,
        these applications were industry-focused, where wire-                                  many copies of the transmit signal arrive at the re-
        less solutions allow for shorter product cycles, more                                  ceiver simultaneously and interfere which each other.
        product individualization and an overall increased flex-                               When the transmitter or the receiver moves, the chan-
        ibility [3]. One major challenge is wireless closed-loop                               nel conditions continuously change since the waves in-
        control as losing packets and the latency of the trans-                                terfere differently at different locations [6]. In the best
        mission might lead to plant instability, which in turn                                 case, all signals constructively add up at the receive
                                                                                               antenna. In the worst case, however, all signals de-
        *
          Correspondence: andreas.trassl@tu-dresden.de                                         structively cancel each other out, effectively leading to
        1
          Vodafone Chair Mobile Communications Systems, Technische Universität
        Dresden, Germany
                                                                                               zero receive power. Situations where the receive power
        2
          Centre for Tactile Internet with Human-in-the-Loop (CeTI)                            is low, so-called outages, have to be avoided for the
        Full list of author information is available at the end of the article                 successful realization of URLLC.

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 2 of 16

          Different from well-established conventional ap-
                                                                                                        10 0
        proaches that cost severe resources either in hard-
        ware (spatial diversity through many antennas and
        signal processing chains) or at the air interface (band-                                       10 -1
        width), the basic concept of this article is to monitor
        the fast fading channel and schedule users only to re-
        sources that are operational. Due to the spatial vari-                                         10 -2
        ation of the channel, a resource that is in outage for
        one user can have perfect channel conditions for an-
                                                                                                       10 -3
        other. Therefore, this approach is expected to realize
        URLLC’s ambitious QoS targets while keeping the re-
        quired additional resource consumption low. This is of                                         10 -4
        utmost importance when considering the scalability of                                                  0            0.1             0.2             0.3             0.4
        a URLLC deployment, i.e., when many devices need
        to be served simultaneously. The resource consump-
        tion becomes even more important when realizing high                                      Figure 1 Performance of outage identification, when relying
                                                                                                  purely on the latest channel observation
        payload applications with latency requirements, e.g.,
        cloud rendered virtual- or augmented reality.
          Rapidly varying channel conditions are challenging
        when monitoring fast fading channels. In a real deploy-                                    • An analysis of the predictor parameters for Ri-
        ment, a monitoring delay τ exists between receiving                                          cian fading is presented. The parameters define
        the last channel observation, scheduling and eventu-                                         the number and periodicity of channel observa-
        ally transmitting the actual payload. Consider the case                                      tions at the input of the predictor.
        that users were scheduled based on the last channel ob-                                    • Performance metrics tailored to the use-case of ra-
        servation. In this case, many actual outages would be                                        dio resource scheduling for URLLC are proposed.
        missed; i.e., the channel is monitored to be operable                                      • Performance evaluation is conducted by means of
        at time t, but non-operable during payload transmis-                                         extensive simulation. Compared to our previous
        sion at t + τ . The performance for Rayleigh fading is                                       works [7, 8], generalized results are provided.
        visualized in Fig. 1, where the effective outage proba-
        bility of such a system Pr(effective outage) is plotted                                2 Methods/Experimental
        against the Doppler frequency normalized monitoring                                    This article aims to study the performance of a novel
                                                                                               outage prediction scheme in Rayleigh and Rician fad-
        delay τ fm for different fading margins F . It is clearly
                                                                                               ing channels. The predictor combines a Wiener filter
        visible that most of the monitoring gain disappears al-
                                                                                               with a threshold comparison for the identification of
        ready for small monitoring delays. In other words, with
                                                                                               future outages. In the Rician fading case, additional
        increasing delay the monitoring quickly becomes point-
                                                                                               line-of-sight (LOS) parameter estimation is employed.
        less and the effective outage probability asymptotically
                                                                                               The focus during design and analysis of the prediction
        converges towards the average link outage probability.                                 scheme is set on a prospective application to URLLC
          A solution to overcome this time delay is to employ                                  radio resource scheduling.
        predictive methods. Although the fast fading chan-                                       The performance of the outage prediction scheme is
        nel conditions change quickly, they still change con-                                  analyzed using analytical statements and Monte-Carlo
        tinuously, which enables predictions. The main con-                                    computer simulation for a practical set of parameters.
        tribution of this article is to describe the design and                                Noisy channel coefficients are generated randomly and
        the performance of an outage predictor. This article is                                fed into the predictor. The predictions are then com-
        based on and extends our previous works [7, 8], where                                  pared with the respective true future value. Evaluation
        outage predictors for Rayleigh and Rician fading were                                  is conducted using classical binary classification anal-
        proposed for the first time. The contributions of this                                 ysis and application-related metrics that are proposed
        article are summarized as follows:                                                     in this work. The number of repetitions is individu-
           • Outage predictors for Rayleigh and Rician fading                                  ally specified during discussion. The underlying data
              channels are described.                                                          of this study is generated from mathematical models,
           • The Rayleigh und Rician fading outage predictors                                  which are completely described in Sec. 4.
              are compared. Their differences are highlighted
              and their area of application is contrasted.                                     3 Related Work
           • Relevant related work regarding fading prediction                                 Estimating the current state and predicting future be-
              is discussed.                                                                    havior of wireless channels has been a research chal-

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 3 of 16

        lenge for many years. As expected, the used estimation                                 to OFDM, [15] also exploits channel prediction for
        methods have been evolving along with the wireless                                     adaptive frequency hopping by adapting transmission
        communications technologies and standards.                                             parameters to the channel conditions at the next cho-
           With the ever-growing demand of higher transmis-                                    sen frequency. Other approaches include the estima-
        sion rates without sacrificing transmission quality in                                 tion of time varying fading parameters with the help
        terms of bit error rate (BER), the availability of chan-                               of Kalman filter variants to enhance the prediction
        nel state information became a necessity. Adaptive                                     performance [17]. In [18], a Kalman filter is used to
        transmission techniques allow a more efficient resource                                directly predict Rayleigh fading. The used state space
        usage, e.g., by choosing the modulation scheme ac-                                     model is based on the SOS model. Recently, opposing
        cording to the current fading conditions [9]. Early ap-                                to statistical methods, machine learning approaches
        proaches were based on the sum-of-sinusoids (SOS)                                      have been proposed for fading channel prediction. For
        modelling. The deterministic fading modelling is based                                 example, in [19] a back propagation neural network is
        on the estimation of Doppler shifts, amplitudes and                                    used to predict I/Q channel coefficients in Rayleigh
        phases of the overlaid sinusoids through ESPRIT- and                                   fading. A recurrent neural-network-based approach is
        MUSIC-based algorithms [10, 11]. This is done un-                                      used in [20] and analyzed in terms of average pre-
        der the assumption that scatterers in the environment                                  diction errors and bit error rates. Those data-driven
        remain constant and the process is dominated by a                                      approaches do not require any modelling or parame-
        few dominant scatterers. The first advances were fol-                                  terization.
        lowed by many variants of auto-regressive (AR) ap-                                       In today’s 5G and future’s 6G mobile broadband
        proaches. These model predicted channel samples as a                                   standard, very high data rates are only one of the en-
        weighted sum of previous samples using the minimum                                     visioned features. For the realization of URLLC, more
        mean square error (MMSE) criterion. This method re-                                    attention needs to be directed towards the reliability
        quires an estimate of the correlation function of the                                  of transmissions rather than solely maximizing spec-
        channel samples. Promising results were obtained, e.g.                                 tral efficiency. Consequently, research on fast fading
        in [9], where a Wiener filter is used for the predic-                                  channel prediction has to perform a paradigm shift
        tion of in-phase and quadrature (I/Q) Rayleigh fading                                  as well. To allow the scheduling system to achieve a
        channels. The predictor is tested using data generated                                 certain QoS, the predictor needs to be built around
        by ray-tracing and real data from vehicular channel                                    suitable reliability measures. For URLLC, the analysis
        measurements. Similar investigations were performed                                    of average prediction errors and BERs is not enough
        in [12, 13]. In these works, an unbiased power predic-                                 anymore. Under the URLLC premise, only very few
        tor is additionally derived based on solely the channel                                investigations have been conducted. In [21, 22] coop-
        power instead of the complex channel gain. The works                                   erative communications schemes, where messages are
        [9, 14, 15] extended the predictor by the use of effi-                                 transmitted over multiple relays, are investigated for
        cient adaptive filtering techniques to update the pre-                                 URLLC. The authors employ fading monitoring and
        dictor coefficients in case of varying long-term channel                               prediction to choose the most suitable relays. It is
        conditions. Motivation for investigating fading predic-                                shown that the coherence time is an insufficient metric
        tion techniques was exclusively the raise of spectral                                  to quantify the reliability of fading prediction methods.
        efficiency.                                                                            This article contributes to fill this gap and provide
           When orthogonal frequency-division multiplexing                                     prediction methods and metrics for general URLLC
        (OFDM) became part of the fourth generation (4G)                                       architectures.
        mobile broadband standard, channel prediction re-
        search was directed towards adaptive bit allocation in                                 4 System Model
        OFDM symbols, aiming again to increase the spectral                                    The overall system model considered in this article is
        efficiency. In [16], the MMSE predictor, the Wiener                                    shown in Fig. 2. It is assumed that the user equipments
        filter predictor, as well as the adaptive methods nor-                                 (UEs) periodically transmit training signals, which are
        malised least mean squares (NLMS) and recursive least                                  used to acquire channel information between the base
        squares (RLS) are derived for a single input single out-                               station (BS) and the individual UE. We assume chan-
        put (SISO) OFDM system. The predictors are com-                                        nel reciprocity such that by measuring the uplink (UL)
        pared in terms of their mean prediction errors. Sim-                                   channel, the downlink (DL) channel can also be rated.
        ilarly, for SISO OFDM systems, a simplified MMSE                                       This is practical after a calibration phase to account
        predictor and its extension to the adaptive least mean                                 for differences in the circuits of transmitter and re-
        squares (LMS) and RLS techniques were proposed                                         ceiver as shown in [23]. Monitoring the uplink channel
        in [14]. Evaluations were conducted in terms of av-                                    is preferred over the downlink since thereby the neces-
        erage prediction error and spectral efficiency. Similar                                sary information is directly available to the scheduler

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                                 Page 4 of 16

                                                                                               The variance of the complex channel coefficient 2σ 2 is
                                  Periodic Training Signal
                                                                                               then determined by the mean power of the channel
                                                                                               ΩNLOS according to 2σ 2 = ΩNLOS .
                 Base Station                                                                    An underlying assumption for the widely assumed

                                                    Fading Channel
                                                                     URLLC UEs
                                                                                               classical Doppler spectrum is that waves arrive solely
            Channel Estimation                                                                 in the horizontal plane with equally distributed an-
                                                                                               gles of arrival. The UE is considered to move with
              Outage Prediction
                                                                                               a constant velocity v in an otherwise static environ-
                    Scheduler                                                                  ment, which results in a maximum Doppler shift fm .
                                                                                               It is well-known that this leads to the classical Doppler
                                                                                               spectrum with its autocovariance function
                                      Scheduling Decision
                                                                                                      rNLOS (t1 , t2 ) = 2σ 2 J0 2πfm (t2 − t1 ) .
                                                                                                                                                
                                                                                                                                                                             (2)
          Figure 2 Overall system model; The outage predictor is
          highlighted and topic of this article                                                Thereby, J0 denotes the zeroth order Bessel function
                                                                                               of the first kind.

        at the BS. The channel estimations are fed into the                                    4.1.2 Rician fading
        outage predictor, whose design and performance will                                    When additionally allowing for a LOS component, the
        be the main topic of this article. For each monitored                                  Rician fading case arises with its I/Q channel coeffi-
        carrier frequency, the predictor calculates for the next                               cient [24]
        possible UL and DL scheduling opportunity if the re-                                                     √
        spective link is operational or not. Based on this infor-                                     h(t) =      2σ · hNLOS (t) + A · hLOS (t)                  .           (3)
        mation, the scheduler allocates resources to the UEs.                                                          √
        As usual, the scheduling decision is transmitted to the                                The NLOS component 2σ · hNLOS (t) is similar to the
        UEs. In the following, the necessary assumptions for                                   Rayleigh fading case described above. The LOS com-
        the fading channel and the communications system are                                   ponent A · hLOS (t) is modeled to be purely determin-
        explained.                                                                             istic following
                                                                                                                                            
        4.1 Fading Channel                                                                        A·hLOS (t) = A·exp j(2πfD,LOS (t−t0 )+ϕ0 ) . (4)
        In this article, we first consider Rayleigh fading with
        classical Doppler spectra before extending our find-                                   In this formula, A is the amplitude, fD,LOS is the
        ings to the Rician fading model. Rayleigh fading can                                   Doppler frequency, ϕ0 is the initial phase of the LOS
        be used to model challenging non-line-of-sight (NLOS)                                  component and t0 is the reference time at which the
        conditions and is often used as a starting point for                                   phase of the LOS component equals the initial phase
        analysis due to its beneficial mathematical properties.                                ϕ0 .
        In the second part of this paper, we extend our findings                                 In Rician fading the K-factor defines the ratio of
        to the Rician fading case by allowing for a LOS com-                                   power in the LOS component ΩLOS over the power in
        ponent. By doing so the question how the presence of                                   the NLOS component ΩNLOS
        a LOS component affects the prediction performance
        is answered.                                                                                           ΩLOS    A2
                                                                                                      K=             =                  .                                    (5)
                                                                                                               ΩNLOS   2σ 2
        4.1.1 Rayleigh fading
        Rayleigh fading assumes that numerous independent                                      Thus, the standard
                                                                                                           √        deviation of the complex NLOS
        multi-path components arrive at the receive antenna                                    component 2σ and the amplitude of the LOS com-
        simultaneously. In this case, the central limit theorem                                ponent A can alternatively be expressed over the K-
        can be applied and therefore the real and imaginary                                    factor and the average power Ω = ΩLOS +ΩNLOS using
        part of the channel coefficient h(t) can be modelled as
                                                                                                      √
                                                                                                                  r                         r
        Gaussian distributed. Thus, the channel coefficient                                                            Ω                         ΩK
                                                                                                       2σ =                ,         A=                      .               (6)
                     √                                                                                                K +1                      K +1
               h(t) = 2σ · hNLOS (t)                 ,                                (1)
                                                                                               For the special case of K = 0, (1) and (3) coincide.
        follows a zero mean complex Gaussian distribution                                      Thus, the more general Rician fading model also in-
        with variance 2σ 2 , since we define hNLOS (t) ∼ CN (0, 1).                            cludes the Rayleigh fading case.

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 5 of 16

                                  |havg |                                                      With knowledge of the sent pilot symbols at the re-
                           -80                                                                 ceiver, the influence of the fading can be observed
           |h(t)| [dBm]

                                                                                     up
                           -90    |hmin |                                                      by estimating the complex channel coefficient h(t).
                                                                                               For this purpose, we use the minimum variance un-
                          -100

                                                                                     outage
                                                                                               biased (MVU) channel estimator [25]
                          -120
                          -130                                                                        ĥ(t) = (pH p)−1 pH y               ,                                  (9)
                                                      t

          Figure 3 Two state fading model
                                                                                               which corresponds the least squares (LS) and maxi-
                                                                                               mum likelihood (ML) channel estimator. Inserting (8)
                                                                                               in (9) yields

        4.1.3 Two State Fading Model                                                                  ĥ(t) = h(t) + n0 (t)           .                                     (10)
          Our predictor is built upon an abstract fading model
        in which the fading is classified in two states up and                                 Thus, under the given assumptions the estimate ĥ(t)
        outage depending on the channel gain. This fading                                      is superimposed by CWGN n0 with variance 2σn2 0 =
        model is depicted in Fig. 3. The respective fading                                     2σn2 (pH p)−1 . When only one pilot is used for channel
        state is based on the relation of the channel gain to                                  estimation, 2σn0 2 = 2σn2 applies. The relationship be-
        a chosen threshold value |hmin |. A different form to                                  tween channel estimation and noise (10) is the starting
        characterize the threshold |hmin | is the fading margin                                point for the derivation of the predictor. The perfor-
        F = |havg |2 /|hmin |2 , which relates the threshold |hmin |                           mance of the predictor will be determined by the SNR
        to the average channel gain |havg |. When the chan-                                    of the channel estimation, SNR = 2σΩ0 2 .
                                                                                                                                                        n
        nel gain is greater than the threshold |h(t)| > |hmin |
        the current fading state is denoted as up. In the up-                                  5 Problem Statement
        state the signal/noise ratio (SNR) at the receiver is                                  Prediction of the channel state involves so-called bi-
        high enough for an URLLC application to be work-                                       nary classification.Usually, the results of the classifica-
        ing satisfactory. Packet errors are still possible in the                              tion problem are simply called positive and negative.
        up-state, but the probability for an error is low and                                  Since the outage predictor aims at predicting outages,
        long error bursts are rare. Analogously, an outage oc-                                 this article defines the up-state prediction as the neg-
        curs if the channel gain is below the threshold value                                  ative and the outage prediction as the positive classi-
        |h(t)| < |hmin |. In outage, the SNR is usually too low                                fication result. In binary classification, four potential
        for successful decoding, leading to high probabilities                                 outcomes exist. Apart from true positive and true neg-
        of packet errors and long error bursts. Following these                                ative (correct classification), also the two error types
        considerations, a URLLC service is expected to work                                    false positive and false negative prevail. In the context
        satisfactory in the up-state and fail in the outage state.                             of outage prediction these four outcomes represent the
                                                                                               following:
        4.2 Communications System and Channel Estimation                                          • true positive – detection of a future outage
        For the communications system we assume that the                                          • true negative – detection that an outage will not
        transmission between the UE and the BS is affected                                          occur
        only by the fading of the wireless channel, resulting                                     • false positive – miss that an outage will not occur
        in the complex channel coefficient h(t), and complex                                      • false negative – miss of a future outage
        white Gaussian noise (CWGN) n(t) with variance 2σn2 .                                     The evaluation of such binary classification problems
        Hence, the transmit signal x(t) and the receive signal                                 is well known and various metrics for different applica-
        y(t) are related by                                                                    tions are available. Three important metrics are sum-
                                                                                               marized from [26], where a good overview is given. An
                                                                                               intuitive metric for evaluation of a classifier is its ac-
                          y(t) = x(t) · h(t) + n(t)       .                           (7)
                                                                                               curacy, which is defined as
        To acquire information about the wireless channel, a                                                                  TP + TN
        column vector p consisting of P pilot symbols is trans-                                       accuracy =                                             .              (11)
                                                                                                                         TP + FP + TN + FN
        mitted for the estimation of h(t). Thus, when taking
        (7) into account, this leads to                                                        In this formula TP is the number of true positives and
                                                                                               TN is the number of true negatives. Similarly, FP is
                          y = p · h(t) + n .                                          (8)      the number of false positives and FN is the number of

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 6 of 16

        false negatives. An accuracy of 0.8 means that 20 % of
        the prediction results are wrong. However, by investi-                                                            Channel Estimation
        gating the accuracy metric alone it is impossible to say
        if these errors are false positives or false negatives. The                                                            I/Q Prediction
        metric is of limited benefit for URLLC outage predic-
        tion since the error types have a different impact on the                                        Comparison with                         Prediction Error
        reliability of the wireless link. In the context of QoS-                                         Threshold |h0min |                         Analysis
        focused URLLC, false negative classifications have a
        much higher impact than false positives as they can be                                                                                      Future
        a cause for transmission errors which ultimately lowers                                         Outage Prediction
                                                                                                                                               Outage Probability
        the reliability of the wireless communications system.
        This emphasizes the importance of choosing the right
                                                                                                  Figure 4 Structure of the outage predictor for Rayleigh fading
        metric. Investigating the individual probabilities for
        correct or wrong classification allows for more explicit
        statements. Two common metrics to fully describe the
        classifier are the probability of false alarm (also known                              to assess the utilizability of the resource, i.e., how
        as the false positive rate)                                                            often the observed link can be used for URLLC
                                                                                               traffic of a specific UE. The more false alarms oc-
                                             FP
               Pr(false alarm) =                                                    (12)       cur, the lower Pr(predicted up) gets. For example,
                                           TN + FP                                             Pr(predicted up) = 0.8 indicates that the observed
        and the probabability of detection (also known as the                                  link can be considered for URLLC traffic in 80 % of
        true positive rate)                                                                    the time, whereas in the remaining 20 % the link will
                                                                                               not be assigned to that particular UE for transmission.
                                          TP                                                     The ultimate goal for the predictor is to maximize
               Pr(detection) =                             .                        (13)
                                        FN + TP                                                Pr(predicted up) and minimize Pr(effective outage)
                                                                                               given a certain prediction horizon tp .
           The above metrics are well suited to investigate bi-
        nary test results. However, they do not provide in-
        tuitive interpretation in the context of URLLC radio                                   6 Outage Prediction
        resource scheduling. For example, in the case of very                                  In the following, we describe the structure of the out-
        high Rician K-factors even Pr(detection) = 0 (all out-                                 age predictor. We first address the Rayleigh fading case
        ages are missed) could be acceptable as outages are                                    and pursue to the more complex Rician fading after-
        already very rare. Therefore, Pr(detection) has only                                   wards.
        little qualitative meaning if the predictor performs well
        enough for scheduling purposes. Here, we propose two                                   6.1 Rayleigh Fading Prediction
        new metrics: the compound probability for an up-state                                  The outage predictor for the Rayleigh fading case is
        prediction, but the channel being truly in outage                                      shown in Fig. 4 and its structure is briefly explained
                                                        FN                                     here before providing mathematical details.
              Pr(effective outage) =                               (14)                           The starting point for outage prediction is a history
                                                 TP + FP + TN + FN
                                                                                               of channel estimations which is collected at the input
        and the average probability to have an up-state pre-                                   of the outage predictor. Afterwards, I/Q channel coef-
        diction on the monitored link                                                          ficients need to be predicted from the available channel
                                                                                               estimations by an appropriate prediction technique.
                                                  FN + TN
              Pr(predicted up) =                                                 . (15)        In order to obtain a binary prediction for the chan-
                                             TP + FP + TN + FN                                 nel state (up or outage), the predicted I/Q channel
        First, Pr(effective outage) covers the risk of fatal fail-                             coefficient is compared with a threshold |h0min |. Sub-
        ures due to prediction errors. Thus, this metric en-                                   sequently, an outage prediction is available which can
        ables statements about the reliability of the system.                                  be used, e.g., for scheduling purposes. For the Rayleigh
        For example, Pr(effective outage) = 10−3 indicates                                     fading case, the exact distribution of the I/Q predic-
        that on average 1 in 1000 predictions will result in                                   tion error is known at the Wiener filter output. Thus,
        an outage. Here, we assume that a (perfect) sched-                                     additionally to the outage prediction also the probabil-
        uler can prevent any predicted outage, since the de-                                   ity for a future outage can be calculated for future time
        sign of the scheduler is beyond the scope of this ar-                                  instants. In the next sections a detailed description of
        ticle. Second, Pr(predicted up) is defined in a way                                    each block in Fig. 4 is provided.

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 7 of 16

        6.1.1 I/Q Prediction
        Based on the history of channel estimations, predic-
        tions of I/Q channel coefficients need to be calcu-
        lated. For this purpose, a well investigated Wiener-
        filter-based approach is employed [9, 12]. It was shown
        that the Wiener filter has a promising performance
        not only under the Rayleigh fading assumption, but
        also for real fading channels where empirical covari-
        ances need to be utilized. In contrast to machine learn-
        ing based approaches, analytical statements about the
        prediction error can be derived, which allows calcu-
        lation of future outage probabilities in the Rayleigh
        fading case. For these reasons, the Wiener filter was
        preferred over other available fading prediction tech-
        niques. Nevertheless, if only the outage prediction is of
        interest, other I/Q prediction techniques can be eas-
        ily incorporated into the proposed framework as well
        and replace the Wiener filter. The key statements to
        implement the Wiener filter for the Rayleigh case are
        summarized from [12].                                                                     Figure 5 Outage prediction concept; A threshold for the
                                                                                                  prediction |h0min | different from the outage threshold |hmin | is
           For a prediction horizon tp , the prediction of the I/Q                                introduced to tune the prediction uncertainty
        channel coefficient

               ĥ(t + tp | t) = ϕθ                                                  (16)       not in measured fading channels, since the autocovari-
                                                                                               ance is directly estimated in this case anyway. There-
        is the output of a finite impulse response (FIR) filter                                fore, we do not introduce estimators for these parame-
        with coefficients θ. The observation vector                                            ters and instead assume that they are known through-
                                                                                               out the article. Using the outage predictor for mea-
                                                                                               sured fading channels is beyond the scope of this arti-
                                                                    
            ϕ = ĥ(t) ĥ(t − ∆t)                ... ĥ(t − (M − 1)∆t) (17)
                                                                                               cle and instead left for future work.
        contains M past channel estimations with a fixed time
                                                                                               6.1.2 Comparison with Threshold
        between the observations ∆t. The filter coefficients
                                                                                               Since we are interested if an up-state or an outage will
                                                                                               occur, the predicted I/Q channel coefficient is com-
               θ = R−1
                    NLOS r NLOS                                                     (18)       pared with the threshold |h0min |. Our idea is to choose
                                                                                               a different threshold value for the predicted channel co-
        are calculated from the cross-covariance between chan-                                 efficient |h0min | and not the threshold in the two state
        nel coefficient and observations r NLOS and the autoco-                                fading model |hmin |. This idea is depicted in Fig. 5.
        variance matrix RNLOS of the observations according                                      By using the threshold |h0min | for the prediction, we
        to                                                                                     are able to adjust the trade-off between the effective
                                                                                               outage probability and the probability for an up-state
                 [r NLOS ]j = 2σ 2 J0 2πfm (tp + (j − 1)∆t) ,
                                                                                              prediction as discussed in Sec. 5. The objective is to get
                                                                                               a more conservative predictor, such that falsely pre-
                                                                                    (19)       dicted up-states are rare, which allows the predictor
                                          2
                                       2σ J0 (2πfm |j − i|∆t),               i 6= j            to be used for URLLC scheduling. In return, falsely
               [RNLOS ]ij =                                                           .        predicted outages occur more frequently, which the
                                       2σ 2 + 2σn2 0 ,                       i=j
                                                                                    (20)       scheduler has to deal with. Numerical evaluation of
                                                                                               this trade-off is presented in Sec. 7.3.
          To design the Wiener filter, knowledge about the                                     6.1.3 Prediction Error Analysis
        variance 2σ 2 , the maximum Doppler frequency fm and                                   For the given assumptions in Sec. 4 it can be shown
        the noise variance 2σn20 is needed. However, knowledge                                 that the prediction error
        of these parameters is only required in a model-based
        analysis, which we concentrate on in this article, and                                        e(t) = h(t + tp ) − ĥ(t + tp | t)                                    (21)

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 8 of 16

                                                                                                                          Channel Estimation

                                                                                                                                                     Parameter
                                                                                                                                                     Estimation

                                                                                                         Subtraction of the                 Â, fˆD,LOS , ϕ̂0
                                                                                                         LOS Component

                                                                                                           I/Q Prediction

                                                                                                          Addition of the                   Â, fˆD,LOS , ϕ̂0
                                                                                                         LOS Component

          Figure 6 Illustration of the integration area to calculate the                                 Comparison with
          future outage probability; The red circle marks the integration                                Threshold |h0min |
          area and has radius |hmin |

                                                                                                        Outage Prediction

        follows a zero mean complex Gaussian distribution
                                                                                                  Figure 7 Structure of the outage predictor for Rician fading

                                    2σe2
                                           
               e(t) ∼ CN 0,                    .                                    (22)

        This originates from the fact that both h(t) and                                       Here fe(t) (x, y) is the zero mean bivariate Gaussian
        ĥ(t+tp | t) are zero mean complex Gaussian distributed                                probability density with variance σe2 for both dimen-
        and therefore also their difference follows a zero mean                                sions I and Q. As there is no closed-form solution avail-
        complex Gaussian distribution. Similarly, the filter-                                  able for this integral, it must be evaluated numerically.
        ing operation in (16) does not change the distribu-
        tion type, as scaled and summed zero mean complex                                      6.2 Rician Fading Prediction
        Gaussian random variables are again zero mean com-                                     We now extend the outage predictor for the more gen-
        plex Gaussian. Thus, the distribution of (22) is com-                                  eral Rician fading, where not only NLOS fading, but
        pletely parameterized by the variance of the prediction                                also a LOS component is present. The structure of the
        error [12]                                                                             outage predictor for the Rician fading case is presented
                                                                                               in Fig. 7.
             2σe2 = IE |e(t)|2 = 1 − r T
                                           −1                                                  Rician fading has a non-zero I/Q mean generated
                                       NLOS RNLOS r NLOS                         . (23)
                                                                                               by the LOS-component, which is incompatible with a
          Knowing the distribution of the prediction error,                                    Wiener filter prediction. Therefore, the strategy when
        a predicted channel coefficient value can now be                                       dealing with non-zero mean processes is to subtract
        associated with the probability of outage. We de-                                      the mean before filtering and adding the mean back
        note the probability for a future outage given a                                       again at the Wiener filter output [27]. As the time
        certain predicted channel coefficient ĥ(t + tp | t) as                                varying LOS-component can hardly be assumed to be
        Pr(future outage). As illustrated in Fig. 6, a future                                  known, the outage predictor in the Rician fading case
        outage occurs when the prediction error e(t) lies in                                   has to employ estimators for the LOS parameters A,
        the complex plane within an area S of a circle around                                  fD,LOS and ϕ0 as first step. The estimated LOS pa-
        −ĥ(t + tp | t) with radius |hmin |. This is because the                               rameters lead to the full description of the LOS com-
                                                                                               ponent at time t. After subtracting it from the history
        sum of predicted channel coefficient ĥ(t + tp | t) and
                                                                                               of channel estimations, the filter coefficients are cal-
        prediction error e(t) is the true value of the future
                                                                                               culated and the NLOS component can be predicted
        fading (rearranged version of (21)). Consequently, for
                                                                                               equal to the Rayleigh fading case. In parallel, the LOS
        a prediction error within the area S the true chan-
                                                                                               parameters are used to calculate a prediction of the
        nel coefficient lies within the outage region. Therefore,
                                                                                               LOS component at time t + tp which can then be
        Pr(future outage) is determined by a double integral
                                                                                               added to the Wiener filter output. This leads to a pre-
        over the area S according to
                                                                                               dicted I/Q channel sample, which can be thresholded
                                    Z                                                          against |h0min | to obtain an outage prediction. All steps
             Pr(future outage) =       fe(t) (x, y) dS .    (24)                               from the LOS parameter estimation to the comparison
                                                   S

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 9 of 16

        with the threshold, are repeated when a new prediction                                 consists of N values and is sampled at a discrete sam-
        needs to be calculated.                                                                pling period ∆t. Furthermore, a vector of exponential
          When comparing the outage predictor in Fig. 7 with                                   terms
        Fig. 4, it can be seen that for Rician fading no outage                                          
        probability is calculated. This is due to the fact, that                                      e = exp −j(2πfD,LOS (N − 1)∆t)             ...
        it is not possible to analytically calculate the error dis-
                                                                                                                                                 
                                                                                                                             exp −j(2πfD,LOS ∆t) 1                          (27)
        tribution of the introduced LOS parameter estimation.
        The result is that also the distribution of the prediction                             is part of the estimator. Since the true value of the LOS
        error is unknown and outage probabilities cannot be                                    Doppler frequency fD,LOS could be located between
        calculated. The same problem arises when measured                                      the bins of the periodogram, the frequency estimation
        fading is predicted and the assumptions about the dis-                                 can be greatly improved by interpolation as shown in
        tributions do not hold anymore. In the rest of this                                    [29]. The authors propose an iterative approach, which
        section, we explain the individual predictor elements                                  we also employ in this article to refine the frequency
        shown in Fig. 7 in detail.                                                             estimate (25). As suggested by the authors, we also
                                                                                               use two iterations.
        6.2.1 Parameter Estimation                                                                The ML estimates for the remaining parameters can
        Different from the Rayleigh fading case, a time vary-                                  then be calculated by inserting the frequency estimate
        ing LOS component is present and the parameters A,                                     in (27) (we denote this vector as ê in the following)
        fD,NLOS , ϕ0 need to be estimated from a history of                                    and using
        channel estimations.
          When considering (10) as an estimation problem,                                                       ϕ0 êH
                                                                                                        Â =              ,                                                 (28)
        where 2σ 2 , A and fD,LOS are the unknown parameters,                                                 êêH
        both √the CWGN n0 (t) and the random NLOS compo-                                                         (          )
                                                                                                                   ϕ0 êH
        nent 2σ·hNLOS (t) act as noise. We neglect√ the tempo-                                        ϕ̂0 = arg                         .                                   (29)
        ral correlation of the NLOS component 2σ · hNLOS (t)                                                        êêH
        for parameter estimation, since the complexity of the
        derivation is lower and the resulting estimators still                                   The estimator (29) yields a phase estimate of the last
        perform very well in our outage prediction use-case.                                   element in (26), which is preferable from a prediction
        With both the CWGN and the NLOS component be-                                          point of view. Combining the estimates fD,LOS , Â and
        ing complex Gaussian distributed the sum is complex                                    ϕ̂0 gives an estimate of the LOS component
        Gaussian, too, and can be combined into a single vari-
                                                                                                 · ĥLOS (t) = ·exp j(2π fˆD,LOS (t−t0 )+ ϕ̂0 ) . (30)
                                                                                                                                                   
        able.
          This leads to the standard problem of estimating the
                                                                                               6.2.2 I/Q Prediction and Comparison with Threshold
        parameters of a complex sinusoid in CWGN, which can
                                                                                               With the available parameter estimations, a prediction
        be tackled using a ML estimation approach as shown
                                                                                               of the I/Q channel coefficients in the Rician fading case
        in [25]. For the special case of noiseless Rician fading
                                                                                               can be performed. Since we are again using a Wiener
        the desired ML estimators were derived in [28]. In the
                                                                                               Filter which relies on the input to be zero mean, a
        following, we adapt the estimators from [28] and em-
                                                                                               prediction of future I/Q channel samples
        ploy an optimization to the frequency estimation.
          A ML estimation of the frequency                                                                                        
                                                                                                   ĥ(t + tp | t) = ϕ − Â · ĥLOS θ + Â · ĥLOS (t + tp ) (31)
                                                          2
                                                              !
                                                 ϕ0 eT                                         consists of the estimated LOS component at prediction
               fˆD,LOS = −arg max                                                   (25)
                                                  eeH                                          time  · ĥLOS (t + tp ) added to the FIR filter output
                                                                                                                
                                                                                                ϕ − Â · ĥLOS θ, with the observation vector of the
        is found by maximizing the periodogram with respect                                    Wiener filter ϕ being adjusted for the estimate of the
        to fD,LOS . Since the periodogram is the square of a dis-                              LOS component vector
        crete Fourier transform (DFT), a practical implemen-                                                           
        tation would utilize the fast Fourier transform (FFT)                                         Â · ĥLOS = Â · ĥLOS (t0 ) ĥLOS (t0 − ∆t)
        algorithm. The observation vector of the LOS estima-
                                                                                                                                                         
                                                                                                                               ... ĥLOS (t0 − (M − 1)∆t) (32)
        tion
                                                                                               at the input of the filter. Since after subtraction of the
            ϕ0 = ĥ(t − (N − 1)∆t)
                                                                      
                                                   ... ĥ(t − ∆t) ĥ(t) (26)                   LOS component only the NLOS fading remains, the

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                            Page 10 of 16

        Table 1 Values for numerical evaluation
                 Parameter                                    Value                            Throughout the whole section, the normalized predic-
                 Rician K factor                              0, 5, 10                         tion horizon tp fm is arbitrary set to 0.1. The results of
                 mean channel estimation SNR                  20 dB, 10 dB                     this section were generated by means of computer sim-
                 fading margin F                              10 dB
                                                                                               ulation according to the Monte-Carlo approach. For
                                                                                               each point 4 × 105 predicted fading samples were com-
                                                                                               pared with the respective ideal future fading value.
        filter coefficient can be calculated in the same way as
        in the Rayleigh fading case, shown in Sec. 6.1.1. Also,                                7.2.1 Sampling Period
        the comparison with the threshold is no different from                                 In Fig. 8, the mean squared error (MSE) of the I/Q
        the Rayleigh fading case in Sec. 6.1.2.                                                prediction is plotted against the normalized sampling
                                                                                               period ∆tfm for a fixed history length of the Wiener
        7 Results and Discussion                                                               filter and the LOS estimator. Although the MSE is
        In this section, the performance of the outage predictor                               not suitable to describe the outage prediction perfor-
        for the Rayleigh and the Rician fading case is analyzed                                mance directly, it can be used as a performance in-
        numerically. The scenario and the chosen parameters                                    dicator. Generally speaking, the higher the error of
        for numerical evaluation is described in Sec. 7.1. After                               the I/Q prediction, the worse the outage prediction
        investigating the influence of the predictor parameters                                performance after comparing the predicted I/Q chan-
        in Sec. 7.2, the performance evaluation of the outage                                  nel coefficient with the threshold |h0min |. When look-
        prediction schemes is conducted in Sec. 7.3.                                           ing at the curves in Fig. 8, it can be seen that very
                                                                                               small as well as very large sampling periods do not
        7.1 Scenarios                                                                          perform well for the investigated SNRs and K factors.
        The numerical evaluation is conducted for selected                                     For a fixed history length, very small sampling periods
        numerical values. In the following, three different K-                                 ∆t result in the observation not spanning enough to
        factors are investigated to reflect the case of a strong                               capture the continuous variation of the fading. Simi-
        LOS component (K = 10), a medium LOS component                                         larly, very large sampling periods ∆t, which are greater
        (K = 5) and no LOS component (K = 0, Rayleigh fad-                                     than the coherence time, lead to uncorrelated observa-
        ing case). The performance is shown for two different                                  tions. According to the popular rule of thumb from
        mean channel estimation SNRs of 20 dB and 10 dB. In                                    [6], the coherence time tcoh can be approximated as
        both cases, the fading margin is set to F = 10 dB. All                                 tcoh = 0.423
                                                                                                        fm , which is close to the point in Fig. 8 where
        plots in the following sections are generalized by using                               the MSE begins to rise steeply. In all cases, a long
        times that are normalized to the maximum Doppler                                       plateau of the MSE can be observed, where a wide
        frequency fm . This allows the results to be used for                                  range of sampling periods ∆t perform almost equally
        evaluation of various applications without the need of                                 well. In case of a small SNR of 10 dB and especially for
        re-simulation. To put the provided normalized plots                                    Rayleigh fading (K = 0), a clear optimum for the sam-
        into perspective, the example of a remote-controlled                                   pling period arises. For practical systems, a sampling
        automated guided vehicle (AGV) in an industrial cam-                                   period between this optimum and the coherence time
        pus network is considered. For this use-case we assume                                 could be chosen. The choice of the sampling period is
        a constant relative velocity v = 0.8 m/s and a carrier                                 a trade off: For scheduling purposes, large sampling
        frequency fc = 3.75 GHz. According to fm = fcc·v ,                                     periods unavoidably lead to large prediction horizons,
        where c is the speed of light, this yields a maximum                                   which generally lead to worse prediction performance
        Doppler shift of 10 Hz. In the Rician fading case, the                                 than short-term predictions. On the other hand, small
        LOS Doppler frequency fD,LOS and the starting phase                                    sampling periods imply that training signals need to
        of the LOS component ϕ0 were varied randomly.                                          be sent more frequently. This, however, has a nega-
                                                                                               tive impact on the efficiency of the communications
        7.2 Parameter Analysis                                                                 system and the number of users which can be allo-
        Before being able to investigate the performance of                                    cated to send these training signals. For numerical
        the outage prediction schemes, the predictor requires                                  evaluation, we settled on a normalized sampling pe-
        parameterization. The Wiener filter is parameterized                                   riod of ∆tfm = 0.05. This equals a sampling period of
        by its history length M and its sampling period ∆t.                                    ∆t = 5 ms in the AGV use case described in Sec. 7.1.
        Furthermore, the history length N of the LOS estima-
        tor needs to be specified. In the following, an analy-                                 7.2.2 Wiener Filter History Length
        sis of these parameters is conducted for the scenario                                  In Fig. 9, the Wiener filter history length M is in-
        described in Sec. 7.1 to obtain a satisfactory con-                                    vestigated. Throughout all curves, an increase of the
        figuration for the following performance evaluation.                                   history length M results in a reduction of the MSE.

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                            Page 11 of 16

                 0.2                                                                                     0.2

                0.15                                                                                   0.15

                 0.1                                                                                     0.1

                0.05                                                                                   0.05

                    0                                                                                       0
                        0       0.1         0.2        0.3         0.4        0.5                               0            10              20              30              40

          Figure 8 Influence of the sampling period ∆t on the MSE of                              Figure 9 Influence of the Wiener filter history length M on
          the prediction for F = 10 dB, M = 25, N = 128 and                                       the MSE of the prediction for F = 10 dB, ∆tfm = 0.05 and
          tp fm = 0.1                                                                             N = 128

                                                                                                         0.2
        This is intuitive as adding more information to the
        estimation will not lead to degradation. However, the
        performance gain lowers with increasing M , which is                                           0.15
        also intuitive as recent samples carry more information
        about the future channel state compared to outdated
        samples. However, in the case of a very low K factor                                             0.1
        of K = 0, high values of M still improve the perfor-
        mance. For higher K factors, as the LOS component
        dominates the NLOS component, the curves become                                                0.05
        more flat even for small Wiener filter history lengths
        M . Since the Wiener filter is responsible for the NLOS
                                                                                                            0
        prediction, the choice of M becomes less relevant for                                                   0                 50                100                 150
        these high K factors. Throughout our numerical eval-
        uation, we settle for a Wiener filter history length of
        M = 25.                                                                                   Figure 10 Influence of the LOS estimation history length N
                                                                                                  on the MSE of the prediction for F = 10 dB, ∆tfm = 0.05
        7.2.3 LOS Estimation History Length                                                       and M = 25
        Finally, in Fig. 10 the MSE is plotted for different his-
        tory lengths of the LOS estimator N . Similar to the
        history length of the Wiener filter M , high values of                                 7.3.1 Rayleigh Fading Prediction
        N are beneficial for the performance of the predictor.                                   We first investigate the performance of the outage
        Again, the steepness of the curves decreases with rising                               prediction, which is one of the two predicto outputs.
        N , thus, after a certain value the increase of N does
                                                                                               The shown values originate from Monte-Carlo simu-
        not lead to a significant performance increase anymore.
                                                                                               lations. In our simulations, 2 · 108 I/Q channel co-
        Due to complexity minimization, N should be kept as
                                                                                               efficients were fed into the outage predictor for each
        low as possible as the calculation of the FFT is re-
                                                                                               prediction horizon and compared with the true future
        quired in (25). For numerical evaluation of the Rician
                                                                                               fading state. To put this into perspective, at a sam-
        fading case, we settled on N = 128.
                                                                                               pling period of 5 ms, which is used in the AGV sce-
        7.3 Numerical Evaluation                                                               nario, this equals 10.6 days of consecutive fading. Two
        With the scenario from Sec. 7.1 and the predictor pa-                                  examples for classical metrics to investigate our bi-
        rameters from Sec. 7.2, the prediction schemes can be                                  nary classification problem were introduced in Sec. 5
        evaluated numerically. In the following, we conduct                                    and are plotted in Fig. 11 for different prediction hori-
        the evaluation for the Rayleigh fading outage predic-                                  zons. In Fig. 11(a) the accuracy of the outage predic-
        tor and the Rician fading outage predictor separately                                  tor is plotted against different threshold values for the
        as they have different feature sets.                                                   prediction h0min . One can see that an optimal accu-

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                              Page 12 of 16

                     1                                                                                    1

                  0.8                                                                                0.8

                  0.6                                                                                0.6

                  0.4                                                                                0.4

                  0.2                                                                                0.2

                     0
                                                                                                          0
                         0                  0.5                    1                    1.5                          0.2          0.4          0.6          0.8            1

                                            (a) Accuracy                                               (b) receiver operating characteristic (ROC) curve

          Figure 11 Analysis of the outage prediction performance using common binary classification metrics for 20 dB SNR, F = 10 dB,
          ∆tfm = 0.05 and M = 25

                    1                                                                                 1

                  0.8                                                                              0.8

                  0.6                                                                              0.6

                  0.4                                                                              0.4

                  0.2                                                                              0.2

                    0                                                                                 0
                         10-6     10-5      10-4       10-3      10-2      10-1       100                 10-6     10-5       10-4      10-3      10-2      10-1       100

                                           (a) 20 dB SNR                                                                    (b) 10 dB SNR

          Figure 12 Analysis of the outage prediction performance for 20 dB SNR, F = 10 dB, ∆tfm = 0.05 and M = 25

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                            Page 13 of 16

                 6                                                                             to evaluate the percentage of time the observed link
                                                                                               can be utilized for URLLC traffic of a specific UE. The
                                                                                               performance curves with these metrics are plotted in
                                                                                               Fig. 12. Similar to the ROC, each line spans differ-
                 4                                                                             ent operating points, which can be adjusted by vary-
                                                                                               ing the threshold |h0min |. A prominent point in these
                                                                                               curves is |h0min | = 0, where the channel is predicted as
                 2                                                                             up 100 % of the time and the effective outage proba-
                                                                                               bility equals the average outage probability. The av-
                                                                                               erage outage probability for Rayleigh fading can be
                                                                                               calculated using Pr(outage) = 1 − exp(−1/F ) [30]. In
                 0                                                                             Fig. 12(a) the results for a mean channel estimation
                  -1              -0.5              0              0.5               1         SNR of 20 dB are shown and discussed using the AGV
                                                                                               scenario with fm = 10 Hz. If, for instance, a prediction
                                                                                               horizon of tp = 5 ms is needed to overcome the delay
          Figure 13 Validation of the prediction error distribution                            τ between monitoring and payload and an effective
                                                                                               outage probability Pr(effective outage) = 10−3 is tar-
                                                                                               geted, the link can be used approximately 82 % of the
        racy arises for small prediction horizons tp when the                                  time for URLLC traffic. If a higher prediction horizon
        threshold for prediction h0min is chosen near the actual                               of tp = 10 ms is required and the same effective outage
        outage threshold hmin . This might appear appealing                                    probability of Pr(effective outage) = 10−3 is targeted,
        for the choice of h0min , however, the optimum and the                                 the link can only be utilized approximately 76 % of the
        metric as a whole have only little practical relevance                                 time. In Fig. 12(b), a lower mean channel estimation
        for the outage predictor. The metric combines both                                     SNR of 10 dB is shown. The lower SNR leads to a worse
        error types (false positives and false negatives) within                               overall performance, e.g., when using our previous ex-
        a single number. However, false positives are consid-                                  ample with tp = 5 ms and Pr(effective outage) = 10−3 ,
        erably more critical for an URLLC scheduler as they                                    the probability of having a predicted up link is only
        are the defining parameter for the QoS. Even worse,                                    62 % instead of 82 %. An increase of the maximum
        as we tune the outage predictor to be more conser-                                     Doppler frequency, resulting for example from a higher
        vative by increasing h0min , the number of false posi-                                 carrier frequency fc or an increasing movement speed
        tives is orders of magnitude smaller than the number                                   v, will reduce the achievable prediction horizon tp for
        of false negatives. Therefore, the accuracy of the pre-                                a set of Pr(effective outage) and Pr(predicted up). As
        dictor almost only reflects false negatives, which is the                              a result it can be concluded, that the proposed pre-
        dominating error type in this case. A more informa-                                    diction approach is unsuitable for realizing URLLC
        tive performance evaluation of the predictor can be
                                                                                               services in future mmWave and Terahertz communi-
        done by studying a ROC, which is shown in Fig. 11(b).
                                                                                               cations systems.
        In this performance figure the probability of detection
                                                                                                 Additional to the prediction of outages, the Rayleigh
        Pr(detection) is plotted against the probability for a
                                                                                               fading outage predictor is able to calculate future out-
        false alarm Pr(false alarm). The threshold value for
                                                                                               age probabilities under the assumptions discussed in
        the prediction |h0min | now shifts the operating point
        in the ROC. We can learn from the ROC curve that,                                      Sec. 6.1. Basis for the calculation is (22), which states
        when increasing |h0min | we get a more conservative pre-                               that the real and imaginary parts of the prediction er-
        dictor so that outages are more likely to be detected                                  ror follow a zero mean Gaussian distribution. The vari-
        at the cost of more false alarms. However, we are still                                ance of the zero mean Gaussian distribution was cal-
        not able to make qualitative statements about the ex-                                  culated in (23). These findings are validated in Fig. 13,
        pected scheduling performance since we are not able                                    where an empirical estimate (a normalized histogram)
        to evaluate which combination of Pr(detection) and                                     of the probability density is compared with the analyt-
        Pr(false alarm) is acceptable in the context of URLLC                                  ical calculation for different prediction horizons. One
        radio resource scheduling.                                                             can see that the empirical histograms fit very well be-
          Therefore, following our discussion in Sec. 5, we use                                neath the calculated probability densities.
        Pr(effective outage) instead of Pr(false alarm). This                                    With the known error distribution, the probability
        metric shows the effective outage probability of a per-                                for a future outage can be calculated over the dou-
        fect scheduler and thus can be used to qualitatively                                   ble integral (24). An analysis of this integral reveals
        evaluate the risk of fatal failures due to prediction er-                              that the probability for a future outage only depends
        rors. Instead of Pr(detection) we use Pr(predicted up)                                 on the amplitude of the predicted fading |ĥ(t + tp |t)|.

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
Traßl et al.                                                                                                                                            Page 14 of 16

                                                                                               pronounced, however, the curves are still close to their
                10 0
                                                                                               ideal counterparts.
               10 -1                                                                             When comparing the same prediction horizons for
                                                                                               different K factors, one can observe that the out-
               10 -2                                                                           age predictor performs better at high K factors. For
               10 -3                                                                           example, when in the AGV scenario a prediction
                                                                                               horizon tp = 10 ms is utilized in case of K = 0
               10 -4                                                                           and the effective outage target probability is set to
                                                                                               Pr(effective outage) = 10−5 , the observed link is only
               10 -5                                                                           predicted as up with a 61 % probability. However, for
               10 -6                                                                           K = 5 the same prediction horizon and effective prob-
                                                                                               ability is achieved while the channel is predicted as
                       0             0.5              1              1.5              2
                                                                                               up with a much higher probability of 92 % and for
                                                                                               K = 10 even 99 % is reached. A reason for that is the
                                                                                               decrease of randomness in the fading for increasing
          Figure 14 Future outage probability depending from the
          prediction amplitude for 20 dB SNR, F = 10 dB,
                                                                                               K factors. The randomness originates from the NLOS
          ∆tfm = 0.05 and M = 25                                                               component, whose impact is reduced when a strong
                                                                                               LOS component is present. Ultimately, the determin-
                                                                                               istic LOS component is easier to predict resulting in a
                                                                                               better outage prediction performance.
        The resulting future outage probabilities are plotted
                                                                                                 In Fig. 16, an overall worse performance can be ob-
        in Fig. 14. Using the AGV scenario with fm = 10 Hz,
                                                                                               served compared to Fig. 15 due to the lower SNR.
        if a channel coefficient with an absolute value of 0.6
                                                                                               While high K factors of K = {5, 10} still show a
        is predicted at a prediction horizon of tp = 5 ms, the
                                                                                               promising performance, for K = 0 a small prediction
        channel state would be in the outage region with a
                                                                                               horizon of 5 ms achieves Pr(effective outage) = 10−5
        probability of 10−4 . If, instead, the distance of the pre-
                                                                                               only with a predicted up probability of 26 %. To allevi-
        dicted value from the origin is 0.7, the probability is
                                                                                               ate this behaviour to some extent, the number of pilot
        2 × 10−7 and therefore approximately three orders of
                                                                                               symbols P can be increased, though leads to a reduced
        magnitude smaller. For higher prediction horizons tp ,
                                                                                               spectral efficiency. However, analyzing this trade-off is
        the predicted channel coefficient has to be farther away
                                                                                               out of scope of this article and will be left for future
        from the origin to achieve the same outage probabil-
                                                                                               work.
        ities. This is due to the fact that the variance of the
        error distribution is higher.
                                                                                               8 Conclusion
        7.3.2 Rician Fading Prediction                                                         In view of reducing radio resource consumption for
        The performance curves of the outage predictor de-                                     ultra-reliable wireless communication links, monitor-
        scribed in Sec. 6.2 for the more general Rician fading                                 ing the fast fading channel and taking measures based
        case are shown as solid lines in Fig. 15 for a SNR of                                  on the predicted fading state is a promising strategy.
        20 dB and in Fig. 16 for a SNR of 10 dB. Each line is                                  This article provided outage prediction schemes for
        based on 2 × 108 predicted I/Q channel coefficients. To                                Rayleigh and Rician fading and introduced suitable
        investigate the influence of the LOS estimation, which                                 metrics that describe their performance.
        is the novel component compared to the Rayleigh fad-                                     For the Rayleigh fading case and especially for small
        ing case, we also show the case of ideal LOS parameter                                 prediction horizons, the presented predictor features
        estimation as dashed lines, where ideal estimates are                                  a low missed outage probability while simultaneously
        used for subtraction and prediction of the LOS com-                                    not rigorously denying the current channel. For the Ri-
        ponent and only the NLOS fading is realistically pre-                                  cian fading case, i.e., with a LOS component present, a
        dicted using the Wiener filter. Therefore, the dashed                                  LOS estimator is utilized. Compared to the case where
        lines for K = 0 correspond to the performance curves                                   the LOS component is known, the LOS estimator only
        discussed in Fig. 12. The performance loss which can                                   degrades prediction performance minorly. Generally in
        be observed between solid and dashed lines originates                                  the presence of a LOS component, the outage predic-
        from the imperfections of the introduced LOS estima-                                   tion performance is improved to the Rayleigh fading
        tor. Overall, the performance loss is the smallest when                                case. Evidently, the LOS estimation comes at the cost
        tp is small and the SNR is high. For large prediction                                  of increased complexity, mainly due to the calculation
        horizons and a small SNR the performance loss is more                                  of the FFT as part of the LOS Doppler estimation.

This is a post-peer-review, pre-copyedit version of an article published in EURASIP Journal on Wireless Communications and Networking. The final authenticated version is available online at:
http://doi.org/10.1186/s13638-021-01964-w
You can also read