Multi-Label Wireless Interference Identification with Convolutional Neural Networks
MMulti-Label Wireless Interference Identificationwith Convolutional Neural Networks
Sergej Grunau, Dimitri Block, Uwe Meier inIT - Institute Industrial ITOWL University of Applied SciencesLemgo, Germany { sergej.grunau, dimitri.block, uwe.meier } @hs-owl.de Abstract —The steadily growing use of license-free frequencybands requires reliable coexistence management and thereforeproper wireless interference identification (WII). In this work,we propose a WII approach based upon a deep convolutionalneural network (CNN) which classifies multiple IEEE 802.15.1,IEEE 802.11 b/g and IEEE 802.15.4 interfering signals in thepresence of a utilized signal. The generated multi-label datasetcontains frequency- and time-limited sensing snapshots withthe bandwidth of 10 MHz and duration of 12.8 µs, respectively.Each snapshot combines one utilized signal with up to multipleinterfering signals.The approach shows promising results for same-technology in-terference with a classification accuracy of approximately 100 %for IEEE 802.15.1 and IEEE 802.15.4 signals. For IEEE 802.11b/g signals the accuracy increases for cross-technology interfer-ence with at least 90 %.
I. I
NTRODUCTION
Artificial neural networks and especially convolutional neu-ral networks (CNNs) achieved excellent results for differentbenchmarks in recent years [1], [2], [3]. Neural networksachieve the best performance, e. g. for character recognitionof the mixed National Institute of Standards and Technology(MNIST) database [4]. The results achieved by the CNNsfrom Cires¸an et al. [1] are comparable to human performance.Therefore, a growing number of research fields apply CNNsas classification systems.One of these research fields is wireless interference iden-tification (WII) for coexistence management of license-freeradio bands such as the 2.4 GHz industrial, scientific andmedical (ISM) band. Such bands are shared between incom-patible heterogeneous wireless wireless comunication systems(WCSs). In industrial environments, typically standardizedwireless technologies (WTs) within the 2.4 GHz ISM bandare wide-band high-rate IEEE 802.11b/g/n, narrow-band low-rate IEEE 802.15.4-based WirelessHART and ISA 100.11a,and IEEE 802.15.1-related PNO WSAN-FA and Bluetooth.Additionally, the radio band is shared with many proprietaryWTs which target specific application requirements such asindustrial WLAN (iWLAN) from Siemens AG which is de-rived from IEEE 802.11, FHSS-based Trusted Wireless fromPhoenix Contact and IEEE 802.15.1-based WISA from ABBGroup.Heterogeneous temporal and spectral medium utilization ofshared radio bands result in interferences. Any interference | X ( f ) || X ( f ) || X ( f ) || X ( f ) | f / MHz | X ( f ) | interfering signal utilized signal Fig. 1. Exemplary cross-technology interference snapshots with a utilizedsignal based on the WT IEEE 802.15.1 and an increasing number of interferingsignals based on the WTs IEEE 802.15.4 and IEEE 802.11 b/g can cause packet loss and transmission latency for industrialWCSs. Both effects have to be mitigated for real-time mediumrequirements. The IEC 62657-2 norm [5] for industrial WCSsrecommends an active coexistence management for reliablemedium utilization. Therefore, it advises the utilization of(i) manual, (ii) automatic non-cooperative or (iii) automaticcooperative coexistence management. The first approach isthe most inefficient one, due to time-consuming configurationeffort. The automatic approaches (ii) and (iii) enable efficientself-reconfiguration without manual intervention and radio-specific expertise. Automatic cooperative coexistence manage-ment (iii) requires a control channel, i. e. a logical commoncommunication connection between each coexisting WCSto enable deterministic medium access. For legacy WCSswithout control channel, the non-cooperative approach (ii) isrecommended. Non-cooperative coexistence management ap-proaches are aware of coexisting WCSs based on independentWII and mitigation.In our previous work [6], we proposed a WII approach upondeep CNNs. It classifies interference signals on a per sensingsnapshot basis. In this paper, we extend the approach totarget coexistence management for an utilized WCS which is a r X i v : . [ c s . C V ] A p r nterfered by non-cooperative WCSs. Therefore, the proposedapproach is capable of identification of multiple interferingsignals in the presence of the utilized signal as illustrated inFig. 1.To face realistic WCS capabilities the approach is limitedto a sensing bandwidth of 10 MHz and a sensing snapshot islimited to a duration of 12.8 µs, which results in 128 in-phaseand quadrature (IQ) samples. The evaluation is performed withthe standardized WTs IEEE 802.11b/g, IEEE 802.15.4 andIEEE 802.15.1, which are sharing the 2.4 GHz ISM band. Intotal 19 different variants of modulation types and symbol ratesare utilized. Thereby, the WII approach has to differ between15 allocated frequency channels of the WTs. Additionally, thesensing snapshots contain the dominant signal of the utilizedWCS which aggravates WII even more. Hence, the interferingsignal is acquired in the presence of the utilized signal. Hence,each sensing snapshot superimposes one utilized signal withmultiple interfering signals.The paper is structured as follows. In the next chapter II,the related work is discussed. Then, in chapter III, the gener-ated dataset is explained. Chapter IV aims the CNN design.Chapter V shows the performance of the proposed approach.Finally, chapter VI concludes the paper and suggests futurework. Feature MappingOutputOutput OutputInput Input InputRule-based system Kernelclassification system Representation learning systemFeature MappingHeuristicFunction FeaturesKernel Functions
Fig. 3. Classification system types with engineering (white) and self-optimizing (yellow) processing units
II. R
ELATED W ORK
Classification systems can be divided into (i) rule-based,(ii) kernel-based and (iii) representation learning ones asillustrated in Fig. 3. Rule-based systems are mostly engineeredimplementations of heuristic functions to result in approximatesolutions based on problem domain knowledge. In contrast,kernel classification systems self-optimize the weighting andassignment of pre-engineered features which are also based onproblem domain knowledge. The finally classification systemtype is based on representation learning. Thereby, also thefeature extraction is performed by self-optimization. Hence,representation learning systems eliminate any problem domain knowledge requirement and result in a full self-optimizationclassification system.A well-known sub-type is called deep neural network(DNN), which is a multi-layer neural network for higher-orderfeature extraction. Another sub-type is a CNN, which utilizesconvolution-related feature extraction layers. Hence, a deepCNN classifies with higher-order convolution-related features.The self-optimization processing units in classification sys-tems require a training phase. For the training phase, thesesystems require a set of input data, which contain specificobjects, and the output labels of the object classes. Thereby,the data type, objects, and the labels depend on the applicationdomain. Hence, classification systems assign input data objectsto the label of the corresponding class. With multiple classes,classification problems can be divided into (i) single- and (ii)multi-label ones. In the former, only one out of several objectsis present in the input data. Multi-label classification permitsmore than one object within the input data.Within the problem domain of WII, an input data item is asensing snapshot. Further, a contained object is a superimposedtransmitted signal of a specific WCS. The particular frequencychannel and WT of the transmitted signal express the objectclasses and therefore also the labels. The transmitted signalsare distorted by the radio channel, e. g. attenuation and additivenoise. Additionally, superposition of multiple signals impairsthe classification task. Hence, WII in the presence of a utilizedsignal raises the multi-label requirement.
A. Neuro-Fuzzy Signal Classifier
The first system is called neuro-fuzzy signal classifier(NFSC). The NFSC is a rule-based system that was imple-mented and tested in [7]. The NFSC classifies frequency chan-nels of IEEE 802.11 and IEEE 802.15.1 signals in six differentindustrial scenarios. The NFSC consists of six layers. In thesesix layers, the NFSC extracts feature from the input data.The extracted features are the center frequency, bandwidth,spectral pulse shape, time behavior, and spectral hoppingbehavior of the signals. Then, the features are assigned to thecorresponding WT and frequency channel class. The NFSCcan detect multiple signals in the input data and can, therefore,handle a multi-label classification problem. Nevertheless, themanual engineered heuristic functions lead to sub-optimalclassification accuracies.
B. Convolutional Neural Network Classification
O’Shea et al. [8] introduce an approach based on CNNs.To recognize signal modulation types instead of WT or fre-quency channels. The approach differentiates between elevenmodulation types. Thereby, the input data only contains thesignal, and therefore it is a single-label classification problem.To train these system, a dataset with 96,000 snapshots wasused. For radio channel distortion, a variable-strength whiteGaussian noise was added to the dataset. So, it consist ofsnapshots with a signal-to-noise ratio (SNR) of -20 dB to20 dB. They show that CNN-based self-optimizing systemsoutperform rule-based systems and also other DNN-based X ( f ) || X ( f ) | f / MHz | X ( f ) | f / MHz 2425 2430 f / MHz 2425 2430 f / MHz 2425 2430 f / MHzIEEE 802.15.1 IEEE 802.11 IEEE 802.14.4 Fig. 2. Snapshot examples from the single-label dataset of each class of the wireless technologies IEEE 802.15.1, IEEE 802.11 b/g and IEEE 802.15.4 systems. CNN-based systems achieve the best classificationaccuracies of greater than 90 % for a SNR of at least -2 dB.In our previous work, Schmidt et al. [6] transferred themodulation recognition approach from O’Shea et al. to theproblem domain of WII. The dataset was created using avector signal generator (VSG) for signal generation and areal time spectrum analyzer (RSA) for data acquisition. Thesensing bandwidth was limited to a bandwidth of 10 MHz andduration of 12.8 µs. The dataset includes signals from the WTsIEEE 802.11 b/g, IEEE 802.15.1, and IEEE 802.15.4 withthere in-band frequency channels. They are divided into fifteenclasses.Schmidt et al. evaluated two different network architecturesof CNNs. The first architecture was adapted from O’Sheaet al. [8], while the latter one was a reduced variant toavoid overfitting. However the classification accuracy of thefirst architecture is higher during the training phase, thereis hardly any difference in the validation phase. It resultsin classification accuracies of at least 95 % at SNR of -5 dB. Additionally, the CNNs outperforms the NFSC regardingclassification accuracy. It shows a processing gain of 5.32 dBand a classification accuracy improvement of 8.19 %. But theCNNs is limited due to its single-label capability only toclassify one signal from each sensing snapshot.III. M
ULTI -L ABEL D ATASET G ENERATION
The multi-label WII has to detect interference signals inthe presence of a utilized signal of the WTs IEEE 802.15.1,IEEE 802.11 b/g, and IEEE 802.15.4. To approach real WCScapabilities, a limited sensing bandwidth of 10 MHz wasassumed. Hence, eight simultaneous operating instances arerequired for loss-less WII within the 2.4 GHz band. Eachsensing band covers ten, three and two frequency channels ofthe WTs IEEE 802.15.1, IEEE 802.11 b/g, and IEEE 802.15.4,respectively. Thereby, the frequency channels of the WTIEEE 802.11 are only partially within the sensing bandwidth.The training and validation dataset for the multi-label WIIwas derived from the one of Schmidt et al. [6] as illustratedin Fig 4. They generated a synthetic dataset D single with VSGstimulation and RSA recording. D single consists of severalsnapshots x single ,i and labels Y single ,i . Each snapshot x single ,i is represented by complex 128 IQ samples and contains one x US x multi, j x IS,0 x IS, N -1 / N Fig. 4. Multi-label snapshot generation based on single-label snapshots of autilized signal x US and a weighted sum of multiple interfering signals x IS , i of fifteen different classes. The classes represent the ten,three and two frequency channels of the WT IEEE 802.15.1,IEEE 802.11 b/g, and IEEE 802.15.4, respectively. Addition-ally, D single utilized additive white Gaussian noise (AWGN)such that the SNR varies between -20 dB up to 20 dB with astep size of 2 dB. In total, D single contains 225,225 snapshots,with 715 snapshots per class and SNR combination.The multi-label dataset D multi requires snapshots with sev-eral classes. Thereby, one class is the utilized signal, andthe remainders are interfering signals. Since it is not likelythat all fifteen classes occur in a snapshot simultaneously,it is limited to the utilized signal an up to six interferingsignals. Additionally, despite the varying number of interferingsignals the signal-to-interference ratio (SIR) remains constant.Hence, the results depend only on the amount of interferingsignals. Furthermore, classification of an increasing number ofinterfering signals with a fixed SIR is more challenging.Fig. 4 shows the multi-label generation signal flow withthe combination of the single-label snapshots. Each resultingsnapshot x multi ,j contains a single utilized signal and N interfering signals. Hence, the corresponding label is the unionof the labels of the contained classes. For input single-labelsnapshots with a SNR of 20 dB are used. Furthermore, theinterfering signals were weighted with the factor /N tokeep the SIR constant with the value one. Resulting same-and cross-technology interference (STI, CTI) snapshots areillustrated in Fig. 5 and Fig. 1, respectivly.The entire dataset D multi consists of 450,000 snapshots andlabels. These are divided into 360,000 snapshots for trainingand 90,000 snapshots for validation purposes. X ( f ) || X ( f ) || X ( f ) || X ( f ) | f / MHz | X ( f ) | interfering signal utilized signal Fig. 5. Exemplary same-technology interference snapshots with a utilizedsignal and an increasing number of interfering signals with same WTIEEE 802.15.1
IV. N
EURAL N ETWORK D ESIGN
The CNN utilizes some pre-processing for the input data.Schmidt et al. [6] and Danev et al. [9] have shown that theclassification of the frequency representation of radio signalsincreases the classification accuracy. Therefore, the snapshotsare transformed with the fast fourier transform (FFT). Then,the resulting 128 complex values have been translated into a × matrix with the extracted real and imaginary parts.Thereby, real values are in the first column and the imaginaryvalues in the second one.The CNN output is a vector with 15 elements with thevalue range [0 , . Thereby, each element represents a class.For validation, a threshold is applied to result in binary output. A. Network Architecture
The network architecture of the CNN is shown in Tab. I.It is derived from the CNN architectures of Schmidt et al.[6] and O’Shea et al. [8]. Thereby, Schmidt et al. haveused a softmax activation function at the output of the lastlayer for optimal single-label classification. However, softmaxactivation function is not suitable for a multi-label classifica-tion problem. Therefore, it has been replaced by a sigmoidactivation function. The sigmoid activation function enablesthe independent output calculation for each a class.
B. Network Training
The CNN was trained in 200 epochs. The Adam optimiza-tion was used with the standard default parameters and alearning rate of 0.001 [10]. As a cost function, the binarycross entropy was used as which is the optimal choice forsigmoid output activation functions. The batch size of 256 hasbeen adjusted to the limitations of the computing platform.Additionally, no hyperparameter optimization was applied.
TABLE ICNN A
RCHITECTURE
Layer type Input size Parameters Act. fct.
Convolutional × × filter kernel Rectifiedlayer feature maps linearConvolutional × × × filter kernel Rectifiedlayer feature mapsDropout 60 % linearDense layer , × neurons RectifiedDropout 60 % linearDense layer × neurons Sigmoid C. Implementation Aspects
The CNN was implemented in the programming languagePython with the libraries Keras [11] and Tensorflow [12]. Ahigh end platform with an Intel XENON E5-1660 v3 centralprocessing unit (CPU), 16 GB RAM and a Nvidia GTX 960graphics processing unit (GPU) was used. During the trainingprocess a CNN, an epoch took 390 s, resulting in a durationof 21.6 h for training. V. R
ESULTS
For evaluation, we use a metric called true positiv rate(TPR), which is proposed by Godbole and Sarawagi [13].TPR evaluates the outcome of a single class. Thereby, all dataitems which contain the distinct class are considered. TPRexpresses the proportion of the actual correct classified ones,as illustrated in Fig.6.
Probability ValuesClassificationThreshold Belong to classDo not belong to class
TPFPTNFN
Fig. 6. Evaluation metric true positiv rate
T P R = T P/ ( T P + F N ) expresses the proportion of the correct classified data items out of all itemsbelonging to the certain class It is important to note, that WII targets the classification ofthe interfering signals. The known utilized signal is, therefore,an unwanted one and therefore distorts the snapshots.
A. Single-Label Classification
First, the proposed multi-label WII approach was comparedwith single-label WII one from Schmidt et al. [6]. Therefore,the same single-label validation dataset from Schmidt et al.was utilized. The resulting averaged TPR is shown in Fig. 7with the varying SNR. Thereby, the proposed approach onlyiffers for a SNR below -8 dB. Below, the multi-label WIIapproach is up to 10.14 % worse and subtracts a processinggain of up to 1.1 dB. Hence, the single-label WII approachfrom Schmidt et al. [6] results in slightly better performancewith the assumption of single-label data, and therefore withoutany utilized signal.
20 16 12 8 4 0 4 8 12 16 20SNR in dB0.00.20.40.60.81.0 T P R Single-Label-WIIMulti-Label-WII
Fig. 7. Single-label classification comparison of the proposed multi-label WIIwith the single-label one from Schmidt et al. [6]
B. Same-Technology Interference Classification
The results of the multi-label WII approach with STI areshown in Fig. 8. Thereby, the snapshots contain a utilizedsignal and a varying number of interfering signals of the sameWT. T P R IEEE 802.15.1IEEE 802.11 b/gIEEE 802.15.4
Fig. 8. Same-technology interference WII with the mean TPR (solid line)and the 10% / 90% percentile interval (transparent area)
The classification of STI with the narrow-band WTsIEEE 802.15.1 and IEEE 802.15.4 approach a TPR of one.The optimal behavior results from the limited spectral over-lapping of the superimposed signals. Another reason is thatthe sensing bandwidth entirely covers signals from both WTsIEEE 802.15.1 and IEEE 802.15.4. Additionally, the TPR ofIEEE 802.15.1 STI interferences is independent of the number of interfering signals, and therefore also it is independent ofthe frequency channel of the utilized signal.In contrast, the TPR of a wide-band IEEE 802.11 b/g STIsignal is worse and drops even further with multiple signals.Such signals are only partially within the sensing bandwidth ofa snapshot. Another reason is that the inter-signal overlappingis significant. Therefore, the differentiation between interferingsignals and the utilized signal is more difficult. Additionally,the high variation of the TPR indicates the frequency channeldependency of the utilized signal.
C. Cross-Technology Interference Classification
Figure 9 shows the evaluation for CTIs. It is noticeable,that the TPR for three IEEE 802.11 b/g interfering signalsincreases. This unexpected behavior may result from thecombinatoric property that for three possible options theprobability for one or three correct ones is higher than twocorrect ones. It is also possible that the CNN is overfitting forthree IEEE 802.11 b/g interferers.For CTI with an IEEE 802.11 b/g utilized signal the pro-posed approach results in a slightly better TPR. This is becausethe bandwidth of the IEEE 802.15.1 and IEEE 802.15.4 signalsare narrow compared to the snapshot bandwidth and thereforeless overlapping interference occurs.In case of a wide-band IEEE 802.11 b/g utilized signalthe narrow-band interfering signals suffer from the spectralintersection. It results in a decrease of the TPR. Additionally,the high variation of the TPR indicates the frequency channeldependency of the utilized signal.VI. C
ONCLUSION
The steadily growing use of license-free frequency bandsrequires reliable coexistence management and therefore properwireless interference identification (WII). In this work, wepropose a WII approach based upon deep convolutionalneural network (CNN) which extends our work [6]. TheCNN naively learns its features through self-optimizationduring an extensive data-driven training process. In contrastto our previous work, we target coexistence managementfor cooperative utilized wireless comunication system (WCS)which are interfered by non-cooperative WCSs. We analyzedWCSs with the wireless technologies (WTs) IEEE 802.15.1,IEEE 802.11 b/g and IEEE 802.15.4. Hence, our approachclassifies multiple interfering signals in the presence of autilized signal. Therefore, it is multi-class and multi-labelclassification problem.For multi-label dataset generation, the single-label one formSchmidt et al. [6] has been combined. They are frequency-and time-limited with the bandwidth of 10 MHz and durationof 12.8 µs, respectively. Therefore, the 2.4- GHz-ISM bandis divided into eight spectral non-overlapping sensing sub-bands. Each sub-band contains fifteen frequency channels ofthe WTs, and therefore also fifteen classes. The multi-labeldataset contains in total 450,000 snapshots. Each snapshotcombines one utilized signal with up to six interfering signalswith a signal-to-interference ratio (SIR) of one. T P R a) IEEE 802.11 b/gIEEE 802.15.4 1 2 3 4 5 6Number of interferers b) IEEE 802.15.1IEEE 802.15.4 1 2 3 4 5 6Number of interferers c) IEEE 802.15.1IEEE 802.11 b/g
Fig. 9. TPR of the interfering signals with increasing number of signals for CTI. Interference and utilized signal are different technologies. The technologyof the utilized signal is a) IEEE 802.15.1, b) IEEE 802.11 b/g and c) IEEE 802.15.4 with varying channels. The transparent area shows how the TPR variesas the channel of the signal changes.
The approach shows promising results for same- as wellas for cross-technology interference (STI, CTI). The same-technology interference (STI) classification accuracies ofspectral non-overlapping narrow-band IEEE 802.15.1 andIEEE 802.15.4 are approximately 100 %. In contrast, spectraloverlapping wide-band IEEE 802.11 b/g suffers from lowaccuracy of at least 78 %.For cross-technology interference (CTI) IEEE 802.15.1 andIEEE 802.15.4, the accuracy slightly drops down to at least95 %, because the utilized and interfering signals are partlyspectral overlapping. However, for IEEE 802.11 b/g the accu-racy even increases with at least 90 %. It takes advantage ofthe narrow-band utilized signal.For future work, the evaluation has to be experimentallyvalidated within industrial environments.VII. A
CKNOWLEDGEMENT
Part of this research was founded by KoMe (IGF 18350BG/3 over DFAM, Germany) and HiFlecs (16KIS0266 overBMBF, Germany). R
EFERENCES[1] D. C. Cires¸an, U. Meier, and J. Schmidhuber, “Multi-column deep neuralnetworks for image classification,”
CoRR , vol. abs/1202.2745, 2012.[2] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classificationwith deep convolutional neural networks,” in
Advances in NeuralInformation Processing Systems 25
CoRR , pp. 180–185, 2017.[7] D. Block, D. Toews, and U. Meier, “Implementation of efficient real-timeindustrial wireless interference identification algorithms with fuzzifiedneural networks,” in , Aug 2016, pp. 1738–1742.[8] T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutionalradio modulation recognition networks.” [Online]. Available: http://arxiv.org/pdf/1602.04105v3 [9] B. Danev and S. Capkun, “Transient-based identification of wirelesssensor nodes,” in
Proceedings of the 2009 International Conferenceon Information Processing in Sensor Networks , ser. IPSN ’09.Washington, DC, USA: IEEE Computer Society, 2009, pp. 25–36.[Online]. Available: http://dl.acm.org/citation.cfm?id=1602165.1602170[10] D. P. Kingma and J. Ba, “Adam: A method for stochasticoptimization,”
CoRR , vol. abs/1412.6980, 2014. [Online]. Available:http://arxiv.org/abs/1412.6980[11] F. Chollet et al.