A Primer on Large Intelligent Surface (LIS) for Wireless Sensing in an Industrial Setting
Cristian J. Vaca-Rubio, Pablo Ramirez-Espinosa, Robin Jess Williams, Kimmo Kansanen, Zheng-Hua Tan, Elisabeth de Carvalho, Petar Popovski
AA Primer on Large Intelligent Surface (LIS) forWireless Sensing in an Industrial Setting
Cristian J. Vaca-Rubio ∗ , Pablo Ramirez-Espinosa ∗ , Robin Jess Williams ∗ , Kimmo Kansanen † ,Zheng-Hua Tan ‡ , Elisabeth de Carvalho ∗ and Petar Popovski ∗∗ Department of Electronic Systems, Connectivity Section (CNT) Aalborg University, Denmark † Norwegian University of Science and Technology, Trondheim, Norway ‡ Department of Electronic Systems, Signal and Information Processing Section (SIP) Aalborg University, Denmarkemail: {cjvr, pres, rjw}@es.aau.dk, [email protected], {zt, edc, petarp}@es.aau.dk
Abstract —One of the beyond-5G developments that is oftenhighlighted is the integration of wireless communication and radiosensing. This paper addresses the potential of communication-sensing integration of Large Intelligent Surfaces (LIS) in anexemplary Industry 4.0 scenario. Besides the potential for highthroughput and efficient multiplexing of wireless links, a LIS canoffer a high-resolution rendering of the propagation environment.This is because, in an indoor setting, it can be placed in proximityto the sensed phenomena, while the high resolution is offeredby densely spaced tiny antennas deployed over a large area. Bytreating a LIS as a radio image of the environment, we developsensing techniques that leverage the tools of image processingand computer vision combined with machine learning. We testthese methods for a scenario where we need to detect whetheran industrial robot deviates from a predefined route. The resultsshow that the LIS-based sensing offers high precision and has ahigh application potential in indoor industrial environments.
I. I
NTRODUCTION
Massive multiple-input multiple-output (MIMO) is a funda-mental technology in the 5th generation of wireless networks(5G), with the addition of a large number of antennas perbase station as its key feature [1]. Looking towards post-5G,researchers are defining a new generation of base stations thatare equipped with an even larger number of antennas, givingraise to the concept of large intelligent surface (LIS). Formally,a LIS designates a large continuous electromagnetic surfaceable to transmit and receive radio waves [2], which can beeasily integrated into the propagation environment, e.g., placedon walls. In practice, a LIS is composed of a collection ofclosely spaced tiny antenna elements. Whilst the performanceof LIS in communications has received considerably attentionrecently [3]–[5], the potential of these devices could go be-yond communications applications, e.g., environment sensing.Indeed, such large surfaces contain many antennas that can beused as sensors of the environment based on the channel stateinformation (CSI).Sensing strategies have been widely addressed in the litera-ture in different ways, being the measured magnitude the main
This project has received funding from the European Unionâ ˘A ´Zs Horizon2020 research and innovation programme under the Marie Sklodowska-Curiegrant agreement No 813999.This work has been submitted to the IEEE for possible publication. Copyrightmay be transferred without notice, after which this version may no longer beaccessible. distinction between them. Thus, we can find sensing methodsbased on the communication signals produced by active users[6], Doppler shifts [7], radio tomographic images obtained fromthe received signal strength [8], [9] or radar-like sensing [10],[11], to mention but a few of relevant examples. Interestingly,whilst some of these sensing techniques resort solely on theamplitude — equivalently, power — of the receive signals[8], [9], in those cases where sensing small scale variationsis needed, the full CSI —i.e., amplitude and phase of theimpinging signals — is required [10], [11].On a related note, machine learning (ML) based approachesare gaining popularity in the context of massive MIMO systems,providing suitable solutions to optimization problems [12], [13].Due to the large dimensions of the system in extra-large arrays,it is crucial to use deep learning to exploit complex patternsof information dependency between the transmitted signals.The popularization of LIS as a natural next step from massiveMIMO gives rise to larger arrays and more degrees of freedom,providing huge amounts of data which can feed ML algorithms.Hence, deep learning arises as a potential solution to exploit theperformance of LIS.In this work, we aim to pave the way to the combined useof both deep learning algorithms and the aforementioned largesurfaces, exploring, for first time in the literature, the potentialof such a joint solution to sense the propagation environment.Specifically, the contribution of this work is twofold: • We propose an image-based sensing technique based onthe received signal power at each antenna element of aLIS. These power samples are processed to generate anhigh resolution image of the propagation environmentthat can be used to feed computer vision algorithms tosense large-scale events. • A computer vision algorithm, based on transfer learn-ing and support vector machine (SVM), is defined toprocess the radio images generated by the LIS in orderto detect anomalies over a predefined robot route.The performance of the proposed solution is tested inan indoor industrial scenario, where the impact of the arrayaperture and the inter-antenna distance is thoroughly evaluated.We show that both larger apertures and smaller separations a r X i v : . [ ee ss . SP ] J un etween the LIS elements render higher resolution images,improving the performance of the system.II. P ROBLEM FORMULATION
We consider an industrial scenario where a robot is supposedto follow a fixed route, and assume that, due to arbitraryreasons, it might deviate from the predefined route and followan alternative (undesired) trajectory. Hence, our goal is, basedon the sensing signal transmitted by the target device, beingable to detect whether the robot is following the correct routeor not.In order to perform the anomalous route detection, weassume that a LIS (i.e., a large array of closely spaced antennas),is placed in the scenario. Therefore, the sensing problemreduces to determine, from the received signal at each of theLIS elements, if the transmission has been made from a pointat the desired route or from anomalous ones. Due to the factthat, in general, acquiring an accurate CSI is a non-trivialtask, we resort on the received signal amplitude (equivalently,power), which may lead to simpler systems implementations.To understand the necessity of large arrays and ML techniquesto tackle this problem, let consider the following preliminaryexample:Assume that we have two points, p and p , belonging tothe desired and the anomalous route, respectively. Then, thereceived complex signal vector at the array of M antennas from p and p is given by y k = h k + n k , k = , , (1)where h k is the complex noiseless received signal (channel) and n k ∼ CN M ( , σ I ) represents the noise vector. To verify that p and p are actually points belonging to different trajectories,we could try to perform an hypothesis testing based on theeuclidean distance between the received signal amplitudes, i.e., M (cid:107)| y | − | y |(cid:107) = M M (cid:213) i = | y , i | + | y , i | − Re {| y , i || y , i |} , (2)with y k , i denoting the elements of y k . If we consider that M issufficiently large, then the law of large numbers holds and (2)is rewritten as M (cid:107)| y | − | y |(cid:107) = M M (cid:213) i = (cid:2) | h , i | + | h , i | (cid:3) + σ − M M (cid:213) i = | y , i | M (cid:213) i = | y , i | . (3)In (3), the first term is the sum of the average power ofthe channels, whilst the second term represents the equivalentnoise, which completely depends on the channel realizations.If we would perform an hypothesis testing in order to establisha certain threshold that determines if the two points are indifferent routes, then the variance of the error term woulddetermine the probability of failure. Note that, to obtain anoptimum estimator, we would need to know all the possiblestates of the channels for each path. Moreover, even in themost simple case, i.e., assuming a pure line-of-sight (LoS) propagation, we would still be unable to distinguish if the twopoints are in different trajectories or at distinct positions of thesame route.The use of a very large number of antennas arises as apossible solution to mitigate the effect of the noise term. Forthe sake of illustration, Fig. 2 depicts the variance of the noiseterm as a function of the number of antenna elements in a pureLoS propagation environment, showing clearly how the variancetends to zero as the number of antennas increases.However, although the use of a large number of elements inthe LIS may reduce the noise variance, in a realistic environ-ment, the complexity of the propagation paths is considerable,and the theoretical analysis becomes cumbersome and site-dependent. Hence, in order to gain insight into how the propaga-tion paths between different positions translate into differencesin the received signals, we have to resort on machine learningalgorithms. This, together with the use of LIS, can providethe necessary information about the propagation environmentin order to perform the anomalous route detection.III. H OLOGRAPHIC SENSING
A hologram is a recorded interference pattern as a result ofconstructive and destructive combinations of the superimposedlight-wavefronts, i.e., a photographic recording of a light field[14]. In a wireless context, a LIS could be described as a struc-ture which uses electromagnetic signals impinging in a deter-mined scatterer in order to obtain a profile of the environment.That is, we can use the received signal power received at eachof the multiple elements of the LIS to obtain a high resolutionimage of the propagation environment. Using this approach, thecomplexity of the multipath propagation is reduced to usinginformation represented as an image. This provides a twofoldbenefit: i) the massive number of elements that composed theLIS leads to an accurate environment sensing, and ii) it allowsthe use of computer vision algorithms and image processingtechniques to deal with the resulting images.As an illustrative example, Fig. 1 shows the holographic im-ages obtained from different propagation environments. Specif-ically, Figs. 1a and 1b correspond to a LoS propagation (noscatterers), whilst Figs. 1c and 1d were obtained from anindustrial scenario with a rich scattering. Note that, in thecase in which different scatterers are placed, their position andshapes are captured by the LIS and represented in the image.Thanks to the large aperture offered by the surface, we areable to reconstruct a feature map (image) that describes whatis occurring in space, based on the information acquired fromthe radio propagation environment. To the best of the authors’knowledge, this is the first time that imaged-based sensing isproposed in the literature.IV. M ACHINE LEARNING FOR HOLOGRAPHIC SENSING
A. Model description
We here propose the use of a machine learning model toperform the anomalous route classification task, based on theholographic images obtained at the LIS. In our consideredproblem, the training data is obtained by measuring the received a) LoS, noiseless. (b) LoS, noisy. (c) Real scenario, noiseless. (d) Real scenario, noisy.Fig. 1. Holographic images for LOS and Industry scenarios. The noisy holographic images are shown for a SNR of γ = − dB.
100 200 300 400 500 60000 . . . . · − M E rr o r t e r m v a r i a n ce Fig. 2. Equivalent noise term variance in terms of M in a LoS scenario. power at certain temporal instants while the target device ismoving along the route. Therefore, the amount of data availabledepends on the length of the route and the sampling period,which may lead in some cases to an insufficient dataset tocompletely train the model. A solution to this issue is makinguse of transfer learning [15]. Among the available strategies forthis matter, we will use feature representation.One of the main requirements for transfer learning is thepresence of models that perform well on already defined tasks.These models are usually shared in the form of a large numberof parameters/weights the model achieved while being trainedto a stable state [16]. The famous deep learning Python library,Keras [17], provides an easy way to reuse some of this popularmodels. Due to our large number of features (the RGB values ofeach pixel in the holographic image), and the training data con-straints, we propose the use of a SVM binary classifier, whichhas been proved to perform correctly in the aforementionedconditions [18].The model is detailed in Fig. 3, and consists in two clearlydifferentiated parts: i) feature extractor by using VGG19, and ii) SVM binary classifier. The first diagram corresponds to theoriginal VGG19 architecture. In order to perform the featureextraction, we remove the last fully connected layer (FC)that performs the classification for the purpose of VGG19and modify it for our specific classification task (anomaly/notanomaly in robot’s route). We note that the architecture has beenfrozen for our case, i.e., the weights and biases in VGG19 arenot re-trained along the process. Instead, they are re-used togenerate the features that will be later used to feed the SVM binary classifier. Hence, only the SVM will be trained by usingthe extracted features.
B. Dataset format
The dataset is obtained by sampling the received signalpower at each element of the LIS while the robot moves alongthe trajectory, containing then T samples (holographic imagesnapshots of received power). At each sample, the input is animage represented by a matrix with n c = channels (RGB)of size n w = and n h = pixels. Our data can then bedenoted as { X ( i ) , y ( i ) } Ti = , where X ( i ) is the i -th n w × n h inputfeatures matrix and y ( i ) is the corresponding desired output labelassociated to the image (anomaly or not anomaly).Once the feature extraction is performed, the output is n c = channels of size n w = and n h = pixels. Since SVMworks with vectors, the data is reshaped into an input featurevector formed by × × = features, being now ourdata { x ( i ) , y ( i ) } Ti = , where x ( i ) is the i -th n -dimensional traininginput features vector (being n = ), x ( i ) j is the value of the j -th feature, and y ( i ) is the corresponding desired output labelvector. V. M ODEL VALIDATION
In order to validate the proposed method, we carried outan extensive set of simulations to analyze the performance ofthe classification algorithm based on holographic images. Toproperly obtain the received power values, we make use of a raytracing software, so we can capture the effects of the multipathpropagation in the most reliable way. Specifically, we considerA
LTAIR F EKO W INPROP [19].
A. Simulated scenario
The baseline set-up is described in Fig. 4a. A typical smallsize industrial scenario of size 484 m , where the target robot(highlighted in red colour) follows a horizontal fixed routeparallel to the bottom wall, in which the LIS is deployed. Thedistance between the LIS and the desired trajectory is . m ,and three anomalous routes are considered, with a separationof . m within them, as detailed in Fig. 4b.For these routes, we simulate in the ray tracing software T time steps, which corresponds to different positions of the robotin both the correct and anomalous routes. Then, a holographic , G , B
224 pixels p i x e l s R , G , B
224 pixels p i x e l s Fig. 3. VGG19 and proposed feature extractors.(a) Use case scenario.
LIS
50 cm50 cm50 cmCorrect routeAnomalous route (b) Correct robot route (blue) vsanomalous routes (orange).Fig. 4. Simulated scenario. image snapshot of the measurements is taken at every T . Themost relevant parameters used for simulation are summarizedin Table I. TABLE I. P
ARAMETERS
Frequency(GHz) TxPower(dBm) Nraypaths Antennatype Surface widthdimension (m) Surface heightdimension (m) AntennaSpacing (cm) Propagationmodel3.5 20 20 Omni 22 8 λ / λ / λ Free Space
In our simulations, we set T = , and thus the datasetis composed of 365 radio propagation snapshots containingimages of anomaly and not anomaly situations, as describedin Section IV-B. The dataset is then split into a 80% trainingset and 20% for the test set. During the training phase, thehyperparameter C is tuned to prevent overfitting by controllingthe balance between bias and variance in the SVM model. Theoptimum value used is C = . , which was identified byusing a 5-fold cross-validation strategy [20]. B. Received power and noise modeling
The complex electric field arriving at the i -th antennaelement at sample time t , (cid:101) E i ( t ) , can be regarded as the su- perposition of each path, i.e., (cid:101) E i ( t ) = N r (cid:213) n = (cid:101) E i , n ( t ) = N r (cid:213) n = E i , n ( t ) e j φ i , n ( t ) , (4)where N r is the number of paths and (cid:101) E i , n ( t ) is the complexelectric field at i -th antenna from n -th path, with amplitude E i , n ( t ) and phase φ i , n ( t ) . From (4), and assuming isotropicantennas, the complex signal at the output of the i -th elementis therefore given by (cid:101) V i ( t ) = (cid:115) λ Z i π Z (cid:101) E i ( t ) + n i ( t ) , (5)with λ the wavelength, Z = π the free space impedance, Z i the antenna impedance, and n i ( t ) is complex Gaussian noisewith zero mean and variance σ . For simplicity, we consider Z i = ∀ i . Thus, the power W i ( t ) = | (cid:101) V i | is used at eachtemporal instant t to generate the holographic image. Finally,in order to test the system performance under distinct noiseconditions, the average signal-to-noise ratio (SNR), γ , is definedas γ (cid:44) λ π Z MT σ T (cid:213) t = M (cid:213) i = | (cid:101) E i ( t )| , (6)where M denotes the number of antenna elements in the LIS. C. Performance metrics
To evaluate the prediction effectiveness of our proposedmethod, we resort on common performance metrics that arewidely used in the related literature. Concretely, we are focusingon the F1-Score which is a metric based on the Precision andRecall metrics [21] and is described as: • Positive F1-Score ( PF ) and Negative F1-Score( N F )as the harmonic mean of precision and recall: PF = · PP · RPPP + RP , N F = · PN · RNPN + RN . (7)Where PP and RP stand for Precision and Recall of the positiveclass (anomaly) while PN and RN stand for Precision andRecall of the negative class (not anomalous situation). P o s i t i v e F - S c o r e ( % ) =20 dB /22 Fig. 5. PF score in terms of the array aperture for distinct spaces. VI. N
UMERICAL RESULTS AND D ISCUSSION
We now analyze the system performance in terms of thedifferent involved parameters. Specifically, we focus on theimpact of the array aperture, the distance between the distinctelements of the LIS, and the relation between them.Generally, in the considered industrial setup, it would bemore desirable to avoid undetected anomalies (which mayindicate some error in the robot or some external issue in thepredefined trajectory) than obtaining a false positive. Hence, allthe figures in this section shows the algorithm performance interms of the PF metric. Nevertheless, for reader’s convenience,the N F metric is also provided in tables for all the cases underconsideration. A. Impact of spacing and aperture
To evaluate the impact of both spacing and aperture, weconsider linear arrays, extracted at the middle height of thewall. We assume a fixed γ for all the cases ( γ = dB), anddifferent spacings with respect to the wavelength ( λ / , λ and λ ). Note that the majority of the variations of the pattern inthe holographic images are contained in the horizontal axis,being of interest to analyze this information. The performanceresults for the distinct configurations are depicted in Fig. 5 andsummarized in Table II. As observed, the spacing of λ —which is far from the concept of LIS — is presenting reallyinaccurate results showing that the resolution obtained for thatspacing is quite poor. Hence, it is concluded that the quick vari-ations along the surface provide important information to theclassifier performance. Here, the effect of antenna densificationfor a given aperture is highlighted and it can be seen that thelowest spacing leads to the best results. B. LIS and horizontal array comparison
To show the potential of LIS for sensing, a comparisonbetween using the LIS (understanding by LIS the use of thewhole wall) and several horizontal array configurations takenat middle height of the LIS (128, 64 and 8 antennas) areconsidered for the best spacing case ( λ / ). The results are TABLE II. R
ESULTS S UMMARY APERTURE VS SPACING
Apertures (m)10.88 5.44 2.72 1.36 0.68 λ PF (%) 71 65 64 61 50 NF (%) 71 69 62 57 50 λ PF (%) 63 61 58 51 49 NF (%) 37 66 51 50 46 λ PF (%) 60 57 51 50 46 NF (%) 60 51 50 49 42 plotted in Fig. 6 in terms of γ , and summarized in TableIII. Overall, we observe that system performance is worsenedas the noise is increased or the aperture is decreased, whichare in fact expected results. Regarding Fig. 6, we not onlyobserve that the use of a LIS leads to a considerable betterperformance, but also that it is more robust against noise.This is coherent with the massive MIMO effect highlightedin Section II, where the variance of the resulting noise termreduces as we increase the number of antennas. Moreover, theLIS system robustness to noise in the low γ = − dB scenarioshows the potential of using a full holographic image whichcaptures accurately the macroscopic pattern of the propagationenvironment based on low power sensing signals (which maybe interesting for autonomous devices where energy efficiencyis a desired requirement). As a final remark, it is clear thata bigger aperture and extra dimension (vertical axis of LIS)is adding insightful information of the sensed environment,something expected taking into account the impact of theaperture described previously.
20 0 -20 -30 -40 (dB)5060708090 P o s i t i v e F - S c o r e ( % ) /2 Industry Scenario LIS128 antennas64 antennas8 antennas
Fig. 6. PF score in terms of γ for distinct array configurations.TABLE III. R ESULTS S UMMARY
LIS
VS ARRAY CONFIGURATIONS γ (dB)20 0 -20 -30 -40LIS PF (%) 96 96 93 85 58 NF (%) 96 96 93 85 54Array 128 PF (%) 72 72 71 69 49 NF (%) 73 67 74 65 47Array 64 PF (%) 71 65 64 52 47 NF (%) 71 70 62 51 43Array 8 PF (%) 57 54 50 50 43 NF (%) 48 47 51 51 42 . LIS aperture comparisons In this case LIS with different apertures instead of lineararrays have been evaluated. Once again, the spacing is fixed to λ / . The results are shown in Fig. 7 and summarized in TableIV.
20 0 -20 -30 -40 (dB)5060708090100 P o s i t i v e F - S c o r e ( % ) /2 Industry Scenario LIS Entire WallLIS 128x128LIS 64x64LIS 32x32LIS 16x16LIS 8x8
Fig. 7. Different LISs apertures comparison
Looking at Fig. 7, we can see a counterintuitive behav-ior at a first glance. Focusing for instance in the interval γ ∈ [− dB , dB ] , it is observed that the performance ofthe 64x64 and 128x128 LIS is higher than the LIS whoseaperture is the entire wall. This interesting behavior may berelated to the SNR distribution at each of the antenna elements.According to Section V-B, the noise samples added at eachelement are equally distributed, and their power is calculatedwith respect to the average SNR in both time and space.However, as the LIS dimensions become larger, the differencesin the average received power at each antenna become non-negligible. Hence, the antennas variance of the SNR per elementshould increase with the LIS aperture, making that, although inpractise we are increasing the number of elements to improvethe image resolution, due to the presence of these low SNRat the edges, we may be actually losing information instead oftaking advantage of the aperture increment. TABLE IV. R
ESULTS S UMMARY
LIS
S APERTURES COMPARISON γ
20 0 -20 -30 -40LIS EntireWall PF
96 96 93 85 58 NF
96 96 93 85 54LIS128x128 PF
99 99 99 96 51 NF
99 99 99 96 53LIS64x64 PF
100 100 100 87 46 NF
100 100 100 86 52LIS32x32 PF
90 90 85 63 53 NF
91 91 87 60 49LIS16x16 PF
72 72 70 59 56 NF
67 65 67 58 54LIS8x8 PF
64 67 64 58 49 NF
59 62 59 51 47
VII. C
ONCLUSIONS
We have shown the potential of LIS for sensing the environ-ment, being able to provide high resolution radio images of thepropagation environment that can be processed by existing andversatile solutions in the context of computer vision algorithms.This sensing technique, which we consider appropriate to referto as holographic sensing, arises as a robust solution to capturethe large scale events of a target scenario, with the inherentadvantage that the received signal phase does not need tobe estimated. The combined used of both LIS and machinelearning algorithms may be potentially used in the contextof cognitive radio and multiuser massive MIMO as a supporttechnology to enhance the performance of these systems.R
EFERENCES[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong,and J. C. Zhang, “What will 5g be?”
IEEE J. Sel. Areas Commun. ,vol. 32, no. 6, pp. 1065–1082, 2014.[2] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potentialof data transmission with large intelligent surfaces,”
IEEE Trans. SignalProcess , vol. 66, no. 10, pp. 2746–2758, 2018.[3] E. Basar, “Transmission through large intelligent surfaces: A new frontierin wireless communications,” in
EuCNC: Eur. Conf. Netw. Commun.
IEEE, 2019, pp. 112–117.[4] M.-A. Badiu and J. P. Coon, “Communication through a large reflectingsurface with phase errors,”
IEEE Wireless Commun. Lett , 2019.[5] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, andR. Zhang, “Wireless communications through reconfigurable intelligentsurfaces,”
IEEE Access , vol. 7, pp. 116 753–116 773, 2019.[6] H. Wang, D. Zhang, Y. Wang, J. Ma, Y. Wang, and S. Li, “Rt-fall: A real-time and contactless fall detection system with commodity WiFi devices,”
IEEE Trans. Mobile Comput. , vol. 16, no. 2, pp. 511–526, 2016.[7] Q. Pu, S. Gupta, S. Gollakota, and S. Patel, “Whole-home gesturerecognition using wireless signals,” in
Proc. 19th Annual Inter. Conf.Mobile Comput. & Netw. , 2013, pp. 27–38.[8] Y. Zhao, N. Patwari, J. M. Phillips, and S. Venkatasubramanian, “Radiotomographic imaging and tracking of stationary and moving people viakernel distance,” in . IEEE, 2013, pp. 229–240.[9] J. Wilson and N. Patwari, “Radio tomographic imaging with wirelessnetworks,”
IEEE Trans. Mobile Comput. , vol. 9, no. 5, pp. 621–632,2010.[10] M. Zhao, T. Li, M. Abu Alsheikh, Y. Tian, H. Zhao, A. Torralba, andD. Katabi, “Through-wall human pose estimation using radio signals,” in
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2018, pp. 7356–7365.[11] F. Adib, Z. Kabelac, H. Mao, D. Katabi, and R. C. Miller, “Real-timebreath monitoring using wireless signals,” in
Proc. 20th Annual Inter.Conf. Mobile Comput. Netw. , 2014, pp. 261–262.[12] J. Joung, “Machine learning-based antenna selection in wireless commu-nications,”
IEEE Commun. Lett. , vol. 20, no. 11, pp. 2241–2244, 2016.[13] O. T. Demir and E. Bjornson, “Channel estimation in massive MIMOunder hardware non-linearities: Bayesian methods versus deep learning,”
IEEE O. J. Commun. Soc. , vol. 1, pp. 109–124, 2020.[14] R. Syms, “Practical volume holography clarendon,”
Oxford , vol. 19902,p. 125, 1990.[15] S. J. Pan and Q. Yang, “A survey on transfer learning,”
IEEE Trans.Knowl. Data Eng. , vol. 22, no. 10, pp. 1345–1359, 2009.[16] D. Sarkar, R. Bali, and T. Ghosh,
Hands-On Transfer Learning withPython: Implement advanced deep learning and neural network modelsusing TensorFlow and Keras . Packt Publishing Ltd, 2018.[17] F. Chollet et al. , “Keras,” https://keras.io, 2015.[18] C. M. Bishop,
Pattern recognition and machine learning