Spectral-Based Proactive Blockage Detection for Sub-THz Communications

Human body blockage is one of the main reasons prohibiting the adoption of modern millimeter wave (mmWave, 30–70 GHz) and future sub-terahertz (sub-THz, 100–300 GHz) cellular systems. To utilize system-level blockage avoidance solutions, such as multiconnectivity, one must be able to detect blockage events prior to their occurrence. To this end, recent approaches have utilized machine learning algorithms operating in the time domain. In our letter, we advocate for the use of spectral representation, where the gap between pre-blockage and non-blockage periods is much wider and directly measurable. By utilizing this idea, we developed a simple proactive blockage detection algorithm for indoor deployment of sub-THz systems and evaluated it using blockage detection probability, mean time to blockage, and false alarm rate. Our results obtained using measured traces of blockage events at 156 GHz show that it is capable of detecting a blockage event at least 50 ms prior to its occurrence. Compared to other approaches, it is characterized by a false alarm rate of less than one events/s making it robust even in fading environments.

Spectral-Based Proactive Blockage Detection for Sub-THz Communications Fariha Zhinuk, Anna Gaydamaka , Dmitri Moltchanov , and Yevgeni Koucheryavy Abstract-Human body blockage is one of the main reasons prohibiting the adoption of modern millimeter wave (mmWave, 30-70 GHz) and future sub-terahertz (sub-THz, 100-300 GHz) cellular systems.To utilize system-level blockage avoidance solutions, such as multiconnectivity, one must be able to detect blockage events prior to their occurrence.To this end, recent approaches have utilized machine learning algorithms operating in the time domain.In our letter, we advocate for the use of spectral representation, where the gap between pre-blockage and non-blockage periods is much wider and directly measurable.By utilizing this idea, we developed a simple proactive blockage detection algorithm for indoor deployment of sub-THz systems and evaluated it using blockage detection probability, mean time to blockage, and false alarm rate.Our results obtained using measured traces of blockage events at 156 GHz show that it is capable of detecting a blockage event at least 50 ms prior to its occurrence.Compared to other approaches, it is characterized by a false alarm rate of less than one events/s making it robust even in fading environments.

I. INTRODUCTION
H UMAN body blockage remains one of the biggest obstacles to the adoption of modern millimeter wave (mmWave, 30-70 GHz) 5G New Radio (NR) and future subterahertz (sub-THz, 100-300 GHz) 6G systems.Such blockage events lead to substantial drops in the received signal strength, often resulting in loss of connectivity [1], [2], thus negatively affecting the quality of user experience.
Several attempts have been made to alleviate the impact of human body blockage in mmWave/sub-THz systems.At the system level, 3GPP multiconnectivity function has been proposed [3].However, as shown in [4] and [5], there are still noticeable connectivity losses even when the number of simultaneously supported links by user equipment (UE) is high.On top of this, multiconnectivity requires dense base station (BS) deployments and real-time blockage detection algorithms.To provide cost-efficient densification, 3GPP proposed an integrated access and backhaul (IAB) system design [6].By utilizing low-cost relay nodes, IAB brings the access points closer to the UEs alleviating the impact of blockage.Although these system-level solutions do not fully The authors are with the Faculty of Information Technology and Communication Sciences, Tampere University, 33100 Tampere, Finland (e-mail: anna.gaydamaka@tuni.fi).
Digital Object Identifier 10.1109/LCOMM.2024.3397743 eliminate the problem of occasional connectivity losses caused by human blockages, they allow to significantly decrease its impact.However, all of them require proactive blockage detection techniques that are capable of determining blockage events before their occurrence.Detection techniques capable of identifying blockage prior to its occurrence have received little attention so far, as most of the approaches are capable of determining the blockage only after it occurs, see [2] and references therein.Although, we utilize the measurements from [2], we design a principally different, proactive blockage detection technique.This approach is particularly appealing to system designers because it enables them to predict the blockage time instant prior to its occurrence.
The authors in [7] utilized cross-correlation between beam states to train a deep neural network (DNN) model for predictive blockage detection.The method suggested in [8] also utilizes DNN, but is location-dependent, requiring a long training phase.The sensing capabilities of mmWave/sub-THz bands for blockage detection were suggested in [9].However, this approach requires dynamic cooperation between UEs in real-time, which might not be feasible in practice.Finally, the authors in [10] and [11] observed the so-called "pre-blockage signatures" that occurred in mmWave/sub-THz channel before blockage events, which we also observed in our experiments.Subsequently, by operating in the time domain they utilized a recurrent neural networks for the proactive detection of blockage events.In summary, the proactive blockage detection approaches are mostly probabilistic.They attempted to detect the blockage in the time domain, where the signs of blockage were not easily measurable.To this end, they relied upon machine learning (ML) that requires long training phases.
The aim of this letter is to propose a proactive blockage detection technique for indoor deployments of sub-THz systems.We utilize the measurements of the blockage process at 156 GHz confirming the presence of pre-blockage signatures.We adopt the spectral representation of the received signal by using the short-time Fourier transform and show that these signatures are heavily pronounced and easily measurable in the spectral domain.Using it, we design a threshold-based proactive blockage detection algorithm.We assess the performance of the proposed approach and compare it with the alternatives reported in the literature using the blockage detection probability, mean time to blockage, and false alarm rate.
The main contributions of our work are: • utilization of the spectral domain for proactive blockage detection, where the difference between non-blockage and pre-blockage states measured using the summed short-time Fourier transform reaches two orders of magnitude;

TABLE I EXPERIMENTAL SETUP CONFIGURATIONS
a set of its parameters that maximizes the mean time to blockage and minimizes the false alarm rate; • results showing that the proposed algorithm is capable of detecting the blockage at least 50 ms prior to its occurrence with a false alarm rate of less than one event/s.The letter is organized as follows.The measurement data are presented in Section II.We design the proposed algorithm in Section III and assess its performance in Section IV. Conclusions are drawn in the last section.

II. MEASUREMENTS AND HYPOTHESIS
In this section, we first describe the measurement campaign.We then introduce and illustrate our main hypotheses.

A. Measurement Campaign
We utilize the measurements available in [2].To produce empirical data on human body blockage attenuation, the authors conducted a large-scale measurement campaign.They used a THz source (Tx) operating at a carrier frequency of 156 GHz with an emitted power 90 mW.Both the Tx and receiver (Rx) were equipped with pyramidal horn antennas.The Tx and Rx half-power beamwidths (HPBW) were 10 • with a gain of 25 dB.Tx and Rx antennas were aligned at all time instants during the experiments.The time constant for the amplifier was set to 30 µs.The channel sampling resolution was ∆ = 50 µs.
The measurement campaign was carried out in an empty hall with length 7.5 m, width 4.  I, where d ′ = x − d is the blocker-to-Rx distance.The experiments were repeated 30 times for each configuration.For a more detailed description of the statistical data we refer to [2].

B. Main Hypothesis
Fig. 1 shows the original measured signal behavior during the blockage and the smoothed one using the exponentially weighted moving average (EWMA) algorithm with constant γ = 0.005.Here, EWMA filtering is utilized to smooth fast signal variations due to fading.Parameter γ in EWMA is responsible for the degree of smoothing.Lower values of EWMA lead to a higher degree of smoothing.One of the most interesting features is the presence of signal oscillations just prior to the blockage.This phenomenon can be explained as follows.Short-distance sub-THz wireless channels are vulnerable to diffraction effects that appear in the channel just prior to the blockage event [2].This fact explains both the existence of ripples in the non-blockage state immediately preceding the blockage event and their appearance during it.This behavior was observed in all the experiments.
These oscillations in the time series of the received signal strength can be used for proactive blockage detection.They last for approximately 100-200 ms which is sufficiently long not only to detect the blockage before it occurs but also to take actions to avoid it, e.g., reroute the active session to the back-up link by utilizing the multiconnectivity functionality [5], [12].We also note that similar behavior was reported in [10] and [11] for mmWave band.

III. PROACTIVE BLOCKAGE DETECTION ALGORITHM
In this section, we describe the spectral properties of the non-blockage, pre-blockage, and blockage periods.We then specify the proposed algorithm.

A. Spectral Properties of the Received Signal
Time domain variations of the received signal prior to the blockage are inherently limited in amplitude to 1-2% of those in the non-blockage state.Thus, pattern recognition techniques using ML models can be used to detect blockages.Alternatively, one may utilize the spectral representation by applying the short-time Fourier transform (STFT) in the following form where S(m, ω) represents the STFT magnitude at time instant m and the frequency index Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.  is the n-th sample of the window function.Note that instead of STFT, one may utilize any transform that converts a signal from the time domain to the frequency domain.
Fig. 2 shows the STFT for non-blockage, pre-blockage, and blockage periods computed for W = 500 samples for a trace smoothed using EWMA with γ = 0.005.We observe that the non-blockage period is characterized by the lowest power.Significantly more power is observed during the pre-blockage period due to the oscillating signal behavior.Finally, the highest power is observed during the blockage period, where even the smoothed signal fluctuates significantly.
Table II lists the summed STFT powers for the nonblockage, pre-blockage, and blockage periods for 30 experiments with x = 7 m, h = 1.35 m, d = d ′ = 3.5 m.It can be observed that the summed STFT power for the non-blockage and pre-blockage periods differs by at least an order of magnitude.Thus, simple threshold-based algorithms can be effectively utilized for proactive blockage detection.

B. Proposed Algorithm
The proposed algorithm can be divided into three stages: (1) initialization, (2) active, and (3) blockage recovery (see Algorithm 1).The first and last stages are the service phases, which allow the algorithm to be properly initialized for the active phase.Similar to other approaches, this algorithm can only detect blockage during the active stage.We also note that the algorithm can be implemented on both the UE and BS.For certainty, we assume that the UE runs the algorithm.
1) Initialization Stage: The aim of the initialization phase is to detect whether the UE is currently in a non-blockage state, as this information is not available a priori.To this end, we employ the following procedure knowing that the minimal summed STFT power corresponds to the non-blockage state.Once the UE is initialized, it begins gathering channel samples and smoothing them using EWMA statistics with parameter γ.After collecting W samples, we estimate the current summed STFT power S(t, ω) at time instant t.Then, we collect another W samples and check what happens at (t + W ): (i) if the summed STFT power estimated over (t + W ) is smaller, then we started in the blockage state at t and now switched to the non-blockage state, (ii) alternatively, if the power increases, we move from the pre-blockage period to the blockage state or from the non-blockage state to the pre-blockage period.In the former case, the algorithm proceeds to the active stage (line 5).Otherwise, to avoid starting monitoring the channel in the blockage period, we skip the 95-th quantile of the blockage time t b , reported in [2], to be approximately 400 ms, and execute the initialization stage again (line 7).
2) Active Stage: In the active stage UE continues to collect channel samples and smooth them using EWMA.Once l samples are collected, UE calculates the current value of the summed STFT power S cur (line 10).The blockage detection threshold S th is set then (line 11).If S cur > S nb + S th , the blockage is signaled.The UE performs actions to avoid blockage and proceeds to the blockage recovery stage (line 14).Otherwise, the UE skips l channel samples and repeats the procedure (line 13).
3) Blockage Recovery Stage: The goal of the recovery stage is to ensure that monitoring is not performed during the blockage stage.To this end, the UE skips the 95-th quantile of the blockage time reported in [2] and then enters the initialization stage (line 16).Note that this stage may not be required if the UE changes the network association point in response to a blockage event.In this case, the algorithm proceeds directly to the initialization stage.

C. Parameterization
The proposed algorithm is characterized by four main parameters: (i) EWMA smoothing constant γ, (ii) blockage detection threshold S th , (iii) STFT window length W , (iv) shift length l.The former is responsible for smoothing the channel variations caused by e.g., fast fading and thus widening the gap between the non-blockage and preblockage periods.The blockage detection threshold affects the sensitivity of the algorithm to blockage events and false alarm rate.The STFT window length W provides a trade-off between the robustness of the algorithm and the amount of time the channel remains unattended.Finally, the shift length l is aimed at decreasing the complexity invoking the STFT computation only for every l samples.In the following section, we investigate the impact of these parameters on algorithm performance.

IV. NUMERICAL ASSESSMENT
In this section, we illustrate the performance of the proposed algorithm.To this end, we utilize three metrics: (i) blockage detection probability, P B , (ii) mean time to blockage, E[D], and (iii) false alarm rate, R F .First, we start with parameter selection, provide a trade-off between the considered metrics, and then compare the performance with rival algorithms.
1) Performance Assessment: Fig. 3 illustrates the blockage detection probability, mean time to blockage, and false alarm rate as functions of the detection threshold multiplier, k, specifying the blockage detection threshold via where S nb and S b are the values of the STFT for the non-blockage and blockage periods, respectively, k is the threshold multiplier, and N is the granularity parameter which is set to 10.Here, the graphs are illustrated for different values of the EWMA smoothing constant γ, while the window size is W = 500 samples with a shift length of l = 20 samples.By analyzing the presented results, it can be observed that the blockage detection probability remains perfect for all the considered values of the threshold multiplier k and EWMA smoothing constants, except for γ = 0.0005 and γ = 0.1 corresponding to extreme and light smoothing, respectively.However, as expected, there is a trade-off between the mean time to blockage and false alarm rate.Specifically, for small values of the threshold multiplier k the proposed approach allows to detect blockage very early as E[D] > 150 ms.However, it is sensitive to variations in the computed STFT values, leading to a false alarm rate higher than 15 events/s for all the considered values of the EWMA smoothing threshold.When k increases, the mean time to blockage begins to decrease, which is also associated with a rapid decrease in the false alarm rate.Notably, even for large values of k there is a significant mean time to blockage of approximately 50 ms while the false alarm rate is lower than 1 event/s.We also note that very small values of γ representing a high level of smoothing lead to slightly higher R F and are thus not recommended.For the channel sampling time of ∆ = 50 µs, we recommend utilizing γ ∈ (0.005 − 0.01) and k ∈ (0.4 − 0.8).
We now investigate the response of the algorithm to the window length W and shift length l, for fixed values of the threshold multiplier k = 5 and EWMA smoothing constant γ = 0.005.We specifically note that for all the considered values of W and l the blockage detection probability is strictly unity and is thus not shown here.As can be observed, the increase in the window size, W , utilized to estimate the current value of the STFT, leads to a smaller mean time to blockage.However, even for large values of W it still plateaus at 50 ms.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.Similarly to the threshold multiplier k, the increase in W leads to a decrease in the false alarm rate.The impact of the shift length, l, is noticeable only for small values of W and mainly affects the false alarm rate.In general, the best trade-off between the mean time to blockage and the false alarm rate is ensured by utilizing l ∈ (20 − 60) and W ∈ (400 − 600).
2) Comparison With Other Algorithms: We now compare the performance of the proposed approach to alternative proactive blockage detection methods available in the literature that do not require the use of external sources of information.To this end, Fig. 3(a) illustrates a comparison with the three ML methods presented in [10] and [11].They do not depend on the threshold multiplier k and therefore remain constant.Out of three, Gated Recurrent Units (GRU) show the best performance characterized by 85% chance that the future interval of 0.8 seconds will contain a blockage event.Other methods are characterized by worse performance.
3) Complexity Assessment: In terms of computational complexity, the proposed algorithm requires limited computational effort.Specifically, the algorithm performs the following operations: exponential smoothing of the input data according to EWMA and STFT calculations.The computational complexity of EWMA is O(1), as it depends solely on the previous value when calculating the next one.The complexity of STFT is O(W log(W )), where W represents the window size.Hence, the overall complexity of the algorithm is O(W log(W )).
Other studies that propose proactive blockage detection algorithms utilized Convolutional Neural Networks or Recurrent Neural Networks [8], [10], [13].Specifically, the latter utilized in [8] and [10] is characterized by O(Lm 2 + Lmd + Lm) complexity, where n is the length of the input sequence, m is the number of recurrent units, d is the dimensionality of the input data, and L is the number of layers in the network.
Regarding the spatial complexity, the proposed algorithm requires only 500 samples (W = 500).ML algorithms typically require datasets ranging from 20000 to 10 5 samples for training purposes.

V. CONCLUSION
In this letter, we propose a proactive blockage detection algorithm for indoor deployment of sub-THz systems.Instead of operating in the time domain, where ML-based pattern detection algorithms are required, we advocate the use of the spectral domain, where the gap between the received signal strength in blockage and pre-blockage states reaches two orders of magnitude.Using this representation, we developed a simple approach capable of proactively detecting blockage events.
Our numerical results show that the proposed approach keeps the blockage detection probability at unity while ensuring that it is detected at least 50 ms prior to the blockage event.More importantly, the false alarm rate is less than one events/s for the careful choice of the input parameters.The proposed approach is recommended for conventional and IAB-based sub-THz cellular deployments multiconnectivity.However, further assessment is required in outdoor deployments.
5 m, and height 3 m.The Tx-to-Rx distance was set to 3, 5, and 7 m.The authors considered a blocker crossing the line-of-sight (LoS) line at a standard walking speed of 3.5 km/h.They utilized multiple Tx-to-blocker distances, denoted by d, for each Tx-to-Rx distance, denoted by x: (i) x = 7 m: d = 1.5, 2.5, 5.5 m, (ii) x = 5 m: d = 1.5, 2.5 m, and (iii) x = 3 m: d = 1.5 m.Finally, two Tx and Rx heights were considered: h = 1.35 m, corresponding to the LoS blockage by a chest, and h = 1.65 m, corresponding to the head blockage.All the considered configurations are summarized in Table

Fig. 1 .
Fig. 1.Typical behavior of the signal during the blockage event.

Fig. 3 .
Fig. 3. Algorithm performance as a function of threshold multiplier, k.

Fig. 4 .
Fig. 4. Algorithm performance as a function of window, W .

TABLE II STFTS
FOR NON-BLOCKAGE, PRE-BLOCKAGE, AND BLOCKAGE PERIODS