Microwave Tomography Data Deconstruct of Spatially Diverse C-Band Scatter Components Using Clustering Algorithms

Communication signals that propagate through free space are subject to multi-path interference due to scattering by various objects in the propagation channel. The effect is especially severe in complex situations in dense urban environments. To investigate the problem, a typical multi-static detection scenario is reconstructed under controlled laboratory conditions, from which suitable data sets are created. Data-driven models are then employed in EDGE computing platforms to profile the scatter centers based on the subjective manner in which they affect the signals. These have been interpreted primarily based on clustering algorithm (CA) operations– using a select suite of pre-processing models that effectively tame the variations in the C-band spatial-temporal data. A subset of the data of interest could then be subjected to an optional, compute-intensive machine learning (ML) approach. The relative advantages of the proposed method vis-a-vis an array of conventional schemes are highlighted, while also considering its carbon friendly attribute. Given the more significant association of the data to antenna radiation patterns, estimation of the latter can now be performed free of any anechoic chamber set up in a time and cost agnostic manner. The benefit of this work would lie in the realm of mid-band 5G-NR (and the future 6G) cellular communication systems deployment, where optimizing the distributed antenna location attributes on time and cost-constrained scales becomes imperative before any large-scale deployment.

Scattering-induced multipath signals are closely identified 95 with the concept of indirect far-field measurement, wherein 96 the complexities in RF testing would scale directly with 97 the properties of individual scatterers. All such interactions 98 result in a marked departure from the plane wavefront def-99 inition integral to the measurements conducted at a length-100 ier LoS separation classified as the Fraunhofer distance 101 [17]. Extracting accurate multipath components from channel 102 measurements is an essential and distinguishing factor in 103 simultaneous localization and mapping (SLAM) solutions to 104 multipath-assisted positioning [18], [19], [20], [21] and is 105 based on the operating frequency and bandwidth.

107
The frequency band considered here for evaluation is signif-108 icant in the context of the 5G new radio (5G-NR) wireless 109 technology in digital cellular networks. Of the two bands 110 allocated for this service, frequency range 1 (FR1) constitutes 111 the mid-band 5G (and the future 6G) in the sub-6 GHz band, 112 considered the ''waterfront property'' of the RF spectrum 113 for 5G. FR1 offers the furthest reach in terms of negotiating 114 terrestrial obstacles than FR2 (the mm-wave band) [22], 115 hence the best compromise between RF coverage and RF 116 bandwidth. This mid-band spectrum balances coverage and 117 capacity characteristics, establishing itself as a vital player in 118 the global rollout of 5-6G services.

119
A 5-6G network will have to incorporate extensive test-120 ing that optimizes its performance in an eventual deploy-121 ment, based on propagation analysis and coverage mapping. 122 This testing process shall also ensure peak performance in 123 the related systems while arraying against other negative 124 attributes such as RF interference, noise and distortions. The 125 3GPPs vision of pervasive 5G communication includes the 126 usage of space-based satellite and airborne high-altitude plat-127 form station (HAPS) network nodes [23]. Site and channel 128 characterizations for the eventual location of antennas are an 129 elaborate exercise in employing an extensive array of indoor 130 and outdoor test range equipment. 131 According to the 3GPP standard, carrier aggregation can 132 be activated for each cell group [24]. The 5G deployments 133 in FR1 combine multiple LTE carriers with one NR carrier. 134 A vast majority of these networks worldwide use the TDD 135 mode. As most frequency bands worldwide are FDD and are 136 used by LTE, the first 5G NR network deployments would 137 take advantage of these underutilized TDD frequency bands. 138 Hence, the first generation of 5G modems and, subsequently, 139 the first generation of 5G mobile devices (MD) only support 140 the TDD mode for FR1. Not all service providers own spec-141 trum licenses within a TDD band. To take advantage of 5G 142 with optimized quality of service, to lower latencies and to 143 further address the new market verticals (e.g., automotive and 144 industrial), a network operator must transition to standalone 145 (SA) mode [25], in which the 5G RAN is connected to the 146 5G core network across a series of intermediate steps in the 147 deployment. The optimum path an operator follows is based 148 5G applications, placing ML deployments close to where ms 205 or µs decisions are implemented will be essential. 206 Multi-access EDGE computing is a significant develop-207 ment in-network functionality where data is processed locally 208 at the edge of the network-close to users and devices-209 to circumvent congestion. A key benefit is the significantly 210 reduced inference delay. EDGE computers are a new genera-211 tion of high-performance, modular compute servers that are: 212 -Optimized for ML workloads, such as the highly facile 213 PALE models considered in this work.

214
-Designed throughout for rugged environments.

215
-Scalable by upgrading internal components, such as 216 advanced NVMe flash memory, PCIe 4.0, and high-217 bandwidth networking that can potentially add additional 218 individual units. 219 The telecommunication network disaggregation that even-220 tually aims to provide competent resources at the EDGE has 221 an unintended yet lucrative aside to it in enabling a joint 222 integrated multi-access computing solution at these EDGE 223 locations. With the number of such components requiring 224 deployment and management across geographic areas pro-225 jected to grow exponentially, the communications services 226 providers are looking to transform their networks as capable 227 infrastructure to monetize 5-6G. 228 The current work discusses using certain pre-processing, 229 data-driven algorithms in 5-6G radio network design and 230 optimization. It explores possibilities for applying these tech-231 niques to current 5G (and the future 6G) air interfaces, mainly 232 in the deployment or test phases, to generate immediate 233 improvements to systems. A forecast has been made on 234 the emerging 3GPP AI/ML framework to shape the devel-235 opment of radio access technologies over the longer term. 236 In particular, the work references federated learning [30] -237 a process that creates a collective model from data sourced 238 across distributed learning nodes. Results from the study 239 shall demonstrate the ability of lightweight-compute-device-240 based ML models to contribute unique data and efficiencies 241 to the training process and in on-device inference generation 242 for rapid performance improvements. With these rugged-243 designed EDGE platforms, quick deployment of ML models 244 for real-time ''transportable'' applications is realized. 245 Recent advancements in computation and the advent of 246 highly efficient clustering algorithms have made it possible 247 to model the propagation channels using an entirely data-248 driven approach. This task has been attempted using a select 249 band of pre-processing algorithm-lean engines (PALE) that 250 find immediate acceptance in embedded EDGE computing 251 platforms such as MDs and other portable devices. Such 252 lightweight, compact and rugged computing paradigms bring 253 proximity computation to the data sources to enable trade-254 offs independent, robust, time-sensitive, and acquire-process 255 ability in the information processing resources. The result 256 is an agile decision-making ability regarding the discrimi-257 nation of scatterer-defined signatures present in the propaga-258 tion channel. Based on such time-critical operations, a first-259 order effectiveness of these on-the-fly type algorithms, when 260 spacing. The RF probes can be used to derive details about the 297 paths between any such pair of points and, hence, determine   2) The work projects the PALE pre-processing pipeline 320 approach as a viable alternative to the mainstream 321 and resources-demanding ML/DL algorithms, given its 322 tremendous ability to tame the variations in C-band 323 spatial-temporal big data, resulting in a much dimin-324 ished relative carbon footprint for the proposed activity. 325 3) Identification of distributed EDGE acquisition and 326 computation resources, much-touted in recent times 327 as a ubiquitous technology tool -in a first-of-a-kind 328 quantitative way, as the most suitable solution in place 329 of conventional time-and cost-extensive facilities such 330 as the anechoic chamber for location and operation of 331 mobile and base stations in dense, urban environments. 332 The article is organized in the following manner: a report 333 on the prior art study and an extensive discussion on the 334 experimental setup and hardware specifications are made, 335 leading to the details about data acquisition and specifi-336 cations. A description of the topical pre-processing algo-337 rithms that adopt various clustering techniques to analyze the 338 spatial-temporal data is given. The salient interpretations of 339 prominent scatter signatures are then obtained based on the 340 progressive understanding gained from deploying the special 341 algorithm on the data set, presented in extensive results and 342 discussions. The benefits of the proposed scheme are laid 343 out based on analyses involving a wider gamut of parameters 344 identified to rate its performance against the conventional 345 methods and include the relative carbon footprint during 346 computation. The paper concludes on a poignant note by 347 identifying potential impacts that this research work shall 348 have across a wide gamut of 5-6G technologies related to 349 antennas and RF system design, deployment and operations. 350 frequency range 10 MHz to 6 GHz, with a frequency step 392 less than 1 kHz. The Frequency accuracy is based on a GPS 393 disciplined oven-controlled crystal oscillator (OCXO). The 394 USRP has a total of 2 Tx and 2 Rx channels, and can realize 395 a maximum instantaneous real-time bandwidth of 160 MHz. 396 The maximum I/Q sample rate is 200 MS/s, and the 16-bit 397 DAC has a spurious-free dynamic range (SFDR) of 80 dB. 398 On the Tx side the maximum output power (Pout) can range 399 from 50-100 mW, while the Rx side has a maximum input 400 power (Pin) of -15 dBm and a noise figure of 5-7 dB. The 401 USRP device has a large Xilinx Kintex-7 (410T) FPGA in 402 a half-1U rack-mountable form factor. The Kintex-7 FPGA 403 is a reconfigurable LabVIEW FPGA target that incorporates 404 DSP48 coprocessing for high-rate, low-latency applications, 405 with a flexible hardware architecture and the LabVIEW uni-406 fied design flow [37].

407
The NI PXIe-8374 is the PXIe-MXI Express Interface Kit 408 for connecting the two USRP hardware units to the PXIe-409 1082 chassis. The latter has 8 reconfigurable add-on slots and 410 features a high-bandwidth backplane to meet a wide range 411 of high-performance test and measurement application needs. 412 The overall system is controlled using a PXIe controller that 413 operates in the LabView environment. The NI PXIe-8880 PCI 414 eXtensions for instrumentation (PXI) Express/CompactPCI 415 Express embedded computer [38] is a high bandwidth sys-416 tem controller that integrates standard I/O features in a sin-417 gle unit by using state-of-the-art packaging. Combining an 418 NI PXIe-8880 embedded controller with a PXI Express-419 compatible chassis, such as the NI PXIe-1082 has resulted 420 in a fully PC-compatible computer in a compact, rugged 421 package. By taking advantage of PCI Express technology 422 in the backplane, PXI Express increases the available PXI 423 bandwidth from up to 132 MB/s to up to 48 GB/s for a more 424 than 60x improvement in bandwidth. The standard I/O on 425 each module includes one DisplayPort 1.2 video port, four 426 high-speed USB 2.0 ports, two high-speed USB 3.0 ports, a 427 PCI-based GPIB controller, two Gigabit Ethernet connectors, 428 a reset button, and PXI Express triggers. The NI PXIe-8880 is 429 a modular PC in a PXI Express 3U-size form factor, and has 430 an Octa-Core Intel R Xeon R E5-2618L v3 processor, triple 431 channel DDR4, 1866 MHz memory, all the standard I/O, and 432 an integrated sold-state drive.

433
NI's hardware and software work together to facilitate 434 the PXI Express communications system. The LabVIEW 435 application development environment combines with leading 436 hardware drivers such as NI-DAQmx to provide exceptional 437 control of the NI USRP hardware. LabVIEW is a power-438 ful and easy-to-use graphical programming environment for 439 acquiring data from several different instruments interfaced 440 via the various ports in the PXIe-8880, enabling a meaningful 441 convertion of the acquired data into vital results using pow-442 erful data analysis routines.

443
The TX-RX system design has been integrated as an SDR, 444 with the application layer defined in the Here, the radiation resistance on the transmit side is 468 R = 50 . The power radiated by the transmitter antenna 469 would then be: where g t = 20 dB is the gain of the transmitting side blocks. Besides this, the other functional blocks support the 499 primary ones by deploying different validation metrics such 500 as RX frequency validation, RX bit error rate (BER) vali-501 dation (derive robust inferences from the eye diagram), and 502 radiation pattern correlation coefficient. A qualitative confir-503 mation is obtained from the radiation pattern plot and the 504 received frequency histogram. The carrier signal frequency 505 corresponds to the resonant frequency of the transmit and 506 receiver antennas. The USRP hardware defines a minimum 507 sampling rate of 390-kilo samples per second (ksps); a more 508 reasonable sampling rate of 400 ksps has been considered 509 during the data acquisition phase.

510
The sampling rate has to be set per the theoretical definition 511 of the Nyquist criterion. For a message signal bandwidth of 512 200 kHz to be detected alias-free, a sampling rate of 400 ksps 513 shall suffice. The sampling is done at a much higher rate to 514 improve frequency resolution and signal-to-noise ratio [44]. 515 Oversampling by a factor of 2 has been applied here [45], and 516 this results in a message signal frequency that is a quarter 517 of the sampling rate, i.e., 100 kHz. The USRP hardware 518 allows the transmit and the receive gain to be varied from 519 0 dB to 31.5 dB, corresponding to a transmit power range of 520 50-100 mW. The transmit and receive gain parameters were 521 chosen as 10 dB based on careful observations during signal 522 acquisition. The experiment requires 10 GB of storage size 523 without discarding the data, further reduced to 2 GB. Storing 524 20k samples allows for a frequency resolution as low as 525 400k/20k = 20 Hz [46]. The read and store process consumes 526 2 s, referred to as the dwell time -the halt duration in 527 the transmit antenna orientation at a specific angle. After 528 the fixed dwell time lapses, the transmit antenna is rotated 529 through a further 0.697 • about a fixed vertical axis to the 530 next angular position. Although it is desirable to decrease the 531 dwell time, optimizing the total duration of data acquisition 532 (to accommodate any potential variability in the ambient 533 parameters), lowering below 2 s impacts the quality of radia-534 tion patterns recorded. The quality aspect was inferred from 535 the signal-to-noise ratio (SNR) estimated during raw data 536 processing. The primary data was acquired at f c = 5.45 GHz, 537 and a periodic switching of the receiver to 2.45 GHz was 538 done to obtain data as part of a frequency diversity analysis at 539 angles of 0 • , 90 • , 180 • , and 270 • . The read-and-store process 540 was repeated for the remaining angles until the transmit 541 antenna completed a full rotation.

542
C. PORTING THE DATA 543 MATLAB R2019a was used for data processing and anal-544 ysis in this research work. The five independent data sets 545 employed in this analysis each had a size of 1 GB and 546 comprised 4096 rows and 20,000 columns. The EDGE com-547 puting hardware for data processing was the R-Pi SBC with 548 a Quad Core 1.2 GHz ARM Cortex-A53 64-bit processor 549 with 1 GB RAM; the Broadcom VideoCore IV at 250 MHz 550 GPU; and Debian 9 version of the Linux kernel 4.19 based 551 R-Pi operating system.
Here, P t is determined from equation (2) The first term on the right is the experimental received 590 power value in dBi, and the second corresponds to the refer-591 ence power value obtained from the antenna manufacturer's 592 datasheet. The variation observed in the residual values is 593 primarily attributed to the influence of scatterers present in 594 the propagation channel. Polar plots, like the one shown in Figure 4 for the case 596 of a 60 • off-boresight orientation in RX-2, depict the angu-597 lar distributions of all such residuals. Five more residual 598 plots were generated in a polar reference frame, eventually 599 revealing the angular distributions and characteristics of the 600 scattering centers in terms of variability in the magnitudes of 601 such residuals.  In all subsequent analyses, approach (1) was chosen to pre-617 clude any loss of information in the experimental data.

619
An exercise classifying scatter centers is defined fundamen-620 tally based on their interaction with the C-band signals. The 621 scattering centers' primary classification is either an absorb-622 ing or a reflecting type. While the former would diminish 623 VOLUME 10, 2022 where j is the number of objects in cluster, the object values in Euclidean distances for an n-dimensional Euclidean space are 656 defined as [54]: Silhouette analysis is used to understand the degree of simi-675 larity among the various object in a cluster, i.e. cohesiveness, 676 compared with those in other clusters indicative of divergence 677 [55]. While most performance evaluation methods need a 678 training set, the Silhouette analysis does not require a training 679 set to evaluate the clustering results making it more appro-680 priate as a clustering task in the present study. The values 681 of such metrics in the silhouette analysis range from −1 to 682 +1, with −1 indicating the worst possible clusters and +1 683 indicating the best possible clusters. The silhouette width 684 for every data point in the cluster is calculated using the 685 expression: Given a cluster, a i is the average of distances within the 688 clusters, and b i is the minimum nearest cluster distance, i.e., 689 the distance to the point in the adjacent cluster. The greater an 690 element's S-value (more positive), the higher its likelihood of 691 being clustered in the correct group. Elements with negative 692 S are more likely to be clustered in wrong groups, and the 693 ensemble average within a dissimilarity is more significant 694 than between dissimilarities. Only those clusters with average 695 silhouette scores greater than a certain threshold are consid-696 ered to have exhibited strong structures that are discernible in 697 the input data [55]. Silhouette analysis can also be used to determine the opti-700 mum number of clusters by iterating the K-means clustering 701 analytics process for different numbers of clusters, k, say 702 from 1 to 30. The variation of k with the average silhouette 703 value can then be studied, and the value of k that results in the 704 highest average silhouette value can be considered the most 705 optimum clustering outcome.

706
One observation about this approach is that the centroids 707 are initialized randomly for each value of k. The most opti-708 mum cluster is not considered for the computation of the 709 silhouette value, and the resulting average silhouette value 710 is not the best possible value for that number of clusters: 711 even if clustering is repeated for the exact value of k, 712 such as for n = 5, the resulting clusters are different each 713 time. The method employed has been to iterate clustering 714 multiple times for each value of k, say 20, and find the 715 most optimum cluster by selecting the cluster with the low-716 est Euclidean distance [56]. Each time the clustering pro-717 cess iterates, the centroids are randomized within the range 718 of the dataset, ensuring that all possible combinations get 719 explored. While this iterative approach is not guaranteed to 720 deliver the clustering that globally minimizes within-cluster 721 variation, the algorithm nonetheless ensures that the most 722 optimum cluster for that particular value of k is always 723 selected, from which the average silhouette values can be 724 computed.
The mean of the variable pair is denoted by and, and

785
The correlation coefficient among the k clusters in each of 786 the angular datasets (taken separately) vs the k' clusters of the 787 boresight dataset would result in a (k ×k') matrix. Each cell in 788 the matrix corresponds to a correlation coefficient value that 789 is indicative of the degree of linearity that exists between a 790 pair of clusters: e.g., the value in the fifth row, second column 791 cell, corresponds to the correlation coefficient between the 792 fifth cluster in the angular dataset and the second cluster in 793 the boresight dataset.

794
A histogram plot would have to be generated to visualize 795 the distribution in the correlation coefficients with a careful 796 choice of the bin size to enhance the resolution of the his-797 togram. The correlation coefficient values that fall within the 798 bins were compiled, and the histogram was plotted for each 799 dataset.

801
Box plots are suitable for visualizing the range and level of 802 data and highlighting outliers, often used in explanatory data 803 analysis. The ''box'' shows the inter-quartile range (IQR), 804 with the ''whiskers'' representing the outliers. A generic 805 box plot is a standard way of representing data based on 806 a 5-number summary [61]. This 5-number summary derived 807 out of the box plot, such as the one shown in Figure 5, consists 808 of a minimum (Min); the first quartile (Q1: a quarter of the 809 data is less than this value); the median (m: half of the data is 810 greater than this value); the third quartile (Q3: a quarter of the 811 data is greater than this value); and the maximum value (Max) 812 of the given data set. With outlier data points up to a distance 813 of 1.5 referenced to the respective quartile, box plots give a 814 fair indication when comparing the distributions of several 815 data groups since they summarize the center and spread of the 816 data effectively. The size and degree of symmetry (skewness) 817 of a box plot are directly reflective of the diversity in the data 818 values: the closer the values, the smaller the box size, and vice 819 versa. 820 The data would be quantitatively analyzed, in terms of its 821 variability in received power and angular distribution, based 822 VOLUME 10, 2022 on two approaches employed in generating the box plots.   The presence of random noise in the neighborhood of the 856 experimental space thereby gets detected, e.g. interference 857 from MDs. All such data points were discarded from the 858 further pre-processing stages. Beyond this point, by choosing 859 the number of clusters to be 3, the average silhouette value 860 for the clusters was determined to be 0.81544, which is a 861 clear indication of robust clustering in the data set. Since only 862 the power domain was given as the input to the clustering 863 algorithm, the algorithm splits the data set into three partitions 864 in the respective domain. Such distinct groupings as outcomes 865 of residuals analysis are interpreted as qualitatively repre-866 senting the characteristics of scatterers in the propagation 867 channel. The following observations have been made, referenced in 869 Figure 6: ations. This trend indicates that these residual types 874 do not constitute a definitive signature left by the 875 scatterers of concern in the propagation channel. Their 876 low-profile and ubiquitous presence is discounted from 877 any further active study.  which, in this case, is very unlikely given the random nature of data sets, the correlation coefficients shall be very nearly 0.

1016
The value corroborates the well-known understanding that

1063
Inferences from kernel analytics and its associated methods 1064 are severely influenced by the cluster size and centroid def-1065 inition, while histogram analysis is affected by bin sizes. 1066 Hence, for a conclusive inference, a refinement that uses box 1067 plots in quantitative analytics has been incorporated into this 1068 work. The data was analyzed quantitatively by generating 1069 two different box plots. Every row in the experimental data 1070 series was divided into 19 separate blocks, each consisting 1071 of 1024 data points; the power spectrum analysis helped 1072 determine the power values.

1073
Two different data processing approaches were followed: 1074 1) To understand the variance of power values over 1075 each step, a box plot was generated with the data 1076 corresponding to the exact angles. With each angle 1077 comprising eight snapshots, the combined yield is 1078 19 × 8 = 152 power values for each angle. The 1079 VOLUME 10, 2022 resulting box plots contained eight boxes, and each box position (angle step) in the TX: RX-1 data (Type-2) and 1102 exhibit lower-order variability. Presence of outliers has been 1103 indicated with a '+' symbols in each case. The box plots cor-1104 responding to the various other RX-2 positions also closely 1105 resemble the Type-2 case across all five experimental data 1106 sets, indicating that there has been a negligible effect due to 1107 the known scatterers on the propagating signal in a majority 1108 of cases. The range of the boxes is also relatively low, as can 1109 be inferred from the ordinate values in this plot.

1110
In the select instances that conform to Type-1, the scat-1111 terers' preferential location and location-specific attributes 1112 are evidenced by the reliable signature of a tell-tale scatter-1113 ing phenomenon noticed at specific angular ranges in the 1114 spatial distributions. Potential candidates include electrically 1115 conducting and dielectric types: such as the window grills 1116 and cupboards for the former and doors for the latter. In the 1117 practical field studies, these shall translate into a diverse 1118 assembly of composite and metallic civil structures, vehicles, 1119 and organic constituents within the cell served by FR1. The 1120 snapshots were taken at the exact spatial locations and within 1121 a minimal time interval as a further proof. Yet, the data sets 1122 exhibit significant variations in the received power. A further step in the analysis was to consolidate and accen-1124 tuate the widespread trends in several such box plots. The 1125 specific attributes of the boxes, derived from the several plots 1126 similar to the one shown in Figure 10, could then be related 1127 to the non-LoS arrival paths of RF signals that correspond 1128 to the physical characteristics of the prominent scatterers in 1129 the propagation channel. The results of this consolidation 1130 effort become immediately apparent, as has been inferred 1131 from Figure 11, in that any relative variation is insignificant 1132 at positions with no scatterers (Type-2). At the same time, 1133 significant variabilities are also equally evident at the exact 1134 locations of the scatter centers (Type-1), whose locations have 1135 been identified in terms of their spatial positions. In this 1136 final segment of the study, specifics in the box plots were 1137 directly ascribed to the diversity of the RX data values; i.e., 1138 the more tightly clustered these values are, the smaller the 1139 box size, and vice versa. Similar plots were also generated for all the five experimental data sets corresponding to the 1141 various off-boresight angles θ. Inferences drawn about the scatterer profiles from all such similar plots were starkly 1143 consistent among the different data sets considered in this 1144 study. All of them indeed exhibit a trend identifiable from 1145 the representative plot of Figure 11.

1146
In keeping with the principle goal of the study reported 1147 here, every prominent feature in the several plots generated 1148 (similar to Figure 11) directly impacts the type of interaction 1149 that the C-band signal had suffered since leaving the TX. Complex modeling or RT calculations, using such param-1199 eters, have been the conventional methods to arrive at 1200 distortion-free and efficient communication pathways in 1201 the MIMO channel that require further implementations of 1202 corrective, control and reconfiguration steps, contributing 1203 to sizable latency and inefficiency. The PALE-on-EDGE 1204 scheme proposed here has sidestepped such tasks with highly 1205 encouraging outcomes. The link attributes among BSs, and 1206 dynamic MD-based EDGE units deployed on-field could 1207 now be defined primarily as a data-driven, on-the-fly sig-1208 nal acquisition and processing (and, by extension, a con-1209 trol/reconfiguration) activity. The significant advantage here 1210 would be the inference-delay-free classification of the chan-1211 nel attributes even while the communication services are 1212 fully operational, performed in effect as an algorithmic 1213 pre-processing task.

1214
Inferences derived from the distributions in the bar charts 1215 and the box plots have been instrumental in defining the prop-1216 erties of scatterers in more precise terms and with enhanced 1217 definiteness about their positions and characteristics. Such 1218 a quantitative analysis has primarily derived salient bene-1219 fits from the outcomes of the many intensely data-driven 1220 approaches pursued during this study. The K-means clus-1221 tering scheme was invoked to characterize the scatterers as 1222 part of an elaborate exercise to identify and deploy a suitable 1223 clustering strategy for the spatial-temporal data set. The latter 1224 have been progressively facilitating (and refining) an eventual 1225 lead onto the conclusive stages in scatterer profiling.

1226
To effectively project the advantages in the current work, 1227 vis-a-vis the previous related attempts at discerning the char-1228 acteristics of the scattering centers in the propagation chan-1229 nel, a detailed comparison has been made. The few highly 1230 topical evaluation parameters that had been considered in this 1231 exercise are: 1248 Several poignant attributes from such a comparison study 1249 have been made evident, and these are detailed in Tables 3-5. 1250 In a significant number of instances, a marked advantage has 1251 VOLUME 10, 2022 been noticeable with the employ of the distributed PALE-on-1252 EDGE implementations.  gap is projected to widen, as the duration of activity increases, in this work, distinctly enabling its functionality without 1272 resorting to any of the attended carbon offset remedies.

1273
In the case of the PALE-on-EDGE implementation, the 1274 actual ''edge'' is the completeness in the features it offers, 1275 in the form of a latency-free signal acquisition and inference-1276 delay tolerant data analysis, that the cloud-based systems 1277 distinctly lack. With such a test arrangement, robust esti-1278 mates on antenna placement and features in the topol-1279 ogy can now be derived with enhanced thoroughness 1280 and relatively less effort, with no requirements for any 1281 further meaningful classification expected from the ML 1282 algorithms.

1283
The proposed method would drastically enhance the effi-1284 ciency of system-wide deployments of 5-6G technologies 1285 across an otherwise complex urban landscape. Complexity 1286 notwithstanding, the results obtained here are agnostic to 1287 the multitudes of scattering centers of varying scattering 1288 coefficients as against the precise knowledge of the types 1289 and locations in their limited numbers considered in the 1290 diffuse scattering models and RT approaches [15], [16]. 1291 The latter are conspicuous in their deployment of sophis-1292 ticated test and measurement devices and simulation soft-1293 ware that would prove disadvantageous and unbound on 1294 time and cost scales in the context of dynamic 5-6G field 1295 conditions.

1296
There would be multiple such conducive arrangements 1297 in the actual field scenarios, and data from several among 1298 such receivers could be harnessed. All such data can then be 1299 subjected to an in situ analysis and interpretation, using the 1300 PALE model identified with this study and deployed on mul-1301 tiple MD-based EDGE platforms. The real-world, distributed 1302 large-scale profiling exercise would significantly weigh in 1303 favor of a collective effort made by the otherwise compu-1304 tationally light-weight EDGE platforms. A stark contrast 1305 now emerges to the HPC-defined ML platforms that have 1306 immense computing requirements (upwards of a 100-fold 1307 increase in the hardware cost [67]) and relative, unavoidable 1308 delays in retrieving the processed results; these had not been 1309 considered in this work.

1310
The design and development of 5-6G systems is a massive 1311 undertaking. C-band 5G (and the future 6G) networks would 1312 require a different level of attention in terms of planning, 1313 deployment, and maintenance. To build and iterate till the 1314 system becomes operational is time-consuming and expen-1315 sive. Time-to-market and network quality will depend on the 1316 of tests and measurements during the complete life cycle 1317 of the network. Complete virtual prototyping and simulation 1318 can substantially reduce costs and accelerate this design and 1319 deployment process.

1320
The 5G technologies can create a revolution in operational 1321 flexibility, and the sixth-generation (6G) will be known for 1322 using AI to capitalize on this flexibility [69], [70], [71]. 1323 In 6G, intelligent services will span from cloud data cen-1324 ters to end devices and IoT devices, e.g., self-driving cars, 1325 drones, and auto-robots. It is of prime importance to design 1326 ultra-low latency, ultra-low power and low-cost inference 1327 broader gamut of pre-processing schemes has convincingly 1383 demonstrated its capability to profile the scatter centers. The 1384 task had been accomplished without resorting to an intensive 1385 ML algorithm-based classification regime that is computa-1386 tionally demanding. The proposed method, with its two parts, 1387 viz., acquisition and analysis, is also in stark contrast to 1388 the conventional analytical model-driven schemes that have 1389 been extensively used to characterize various scatterers in the 1390 C-band signal propagating channel. The distinctive advantage 1391 here is the demonstrated ability of the PALE algorithms 1392 suite to closely track the evident patterns that the scatterers 1393 create during their interaction with the RF signals. A concur-1394 rent means has been identified to qualitatively characterize 1395 and quantitatively determine the influence of the scattering 1396 centers on the propagating signal. The derived interpreta-1397 tions have been the location, material attributes, morphology, 1398 and kinematics of the genuine scatterers. Such parameters 1399 have been identified even at the first-order processing of the 1400 tracked time series data.

1401
The results obtained in this work shall find significant 1402 relevance in situations that involve the intentional incor-1403 poration of technologies that shall help evade the detec-1404 tion of targets of interest in aerospace situational aware-1405 ness operations by being employed as potent electronic 1406 counter-countermeasures. The proposed method has the sig-1407 nificant advantage of sensing and tracking such targets from 1408 the signatures of change in the reference signal in the 1409 wake of their interaction with the hostile constituents in 1410 the propagation channel. The results obtained are primarily 1411 based on the outcome of pre-processing a spatial-temporal 1412 data set.

1413
AI/ML is evolving rapidly. The MD industry must take 1414 advantage of this technology and apply it to the mobile net-1415 work architecture from the core to the RAN, the radio inter-1416 face, and the end-user device itself. The federated learning 1417 models proposed in this work leverage 5-6G connectivity, on-1418 device learning and inference techniques to take ML closer 1419 to the real-time processing needed for air interface optimiza-1420 tion and superior end-user experience. Device-based learning 1421 enhances 5-6G QoS. The prospect of an adaptive, ML-native 1422 air interface, for example, could generate a radically simpler 1423 radio that generates unprecedented gains in efficiency and 1424 performance.

1425
As a possible extension of the work reported here, a test 1426 system can be devised to perform in a manner that would 1427 constrain its acquisition parameters dynamically. Such a sys-1428 tem would function with the precise understanding of scat-1429 ter signatures obtained from the trends and patterns in the 1430 acquired data (primarily antenna radiation patterns) fed back 1431 in real-time. With such a closed-loop scheme, the malicious 1432 components in a sensitive RF signal detection procedure can 1433 now be subjected to adaptive excision, a task currently being 1434 pursued at an exorbitant cost and time in use of facilities 1435 such as the anechoic chamber. The profiling of antennas 1436 with unknown radiation characteristics will become a wholly 1437 computationally defined task that will prove cost-effective. 1438 of noise detection, with its progressive exposure to additional

1743
He has been a Faculty Member with the Department of Electronics and 1744 Communication Engineering, Amrita University, since 2007. Since 2015, 1745 he had been a PI of a funded project, and a Co-PI of two other funded 1746 projects, one of which was in ''Design and Evaluation of a DRFM Mitigation 1747 System'' the academic research grant funded by NI, USA. He has published 1748 around 60 articles till date. His research interests include applying ML/AI 1749 to condition monitoring, advanced driver assistance systems, and financial 1750 markets. 1751 S. I. HARUN was born in Coimbatore, Tamil Nadu, India. He received 1752 the B.Tech. degree from the Department of Electronics and Communication 1753 Engineering, ASE-Coimbatore, Amrita University, in 2020, and the master's 1754 degree from Nanyang Technological University, Singapore. He did the final 1755 year project activity in the related area to the one reported in this article.