PaperlayerView exact revision on arXiv

Source-linked AI summary

Wireless Interference Identification with Convolutional Neural Networks

Malte Schmidt, Dimitri Block, Uwe Meier

arXiv:1703.00737v1cs.LGcs.CV

TL;DR

Shared license-free bands need reliable coexistence management, motivating wireless interference identification. The paper applies a deep CNN to short, bandwidth-limited snapshots; it reports improved performance over the NFSC and accuracy above 95% for SNRs greater than -5 dB.

  • Problem

    Reliable coexistence management in growing license-free bands requires wireless interference identification for interference mitigation.

    Source: §I · p. 1, §VI · p. 6

  • Method

    A deep CNN self-optimizes its features to classify 15 wireless technology and channel classes from 12.8 µs snapshots with 10 MHz acquisition bandwidth.

    Source: §VI · p. 6

  • Results

    The CNN’s average performance gain and classification accuracy improvement over NFSC under similar constraints are at least 5.32 dB and 8.19%, respectively; accuracy exceeds 95% for SNRs greater than -5 dB.

    Source: §VI · p. 6

  • Takeaways & Limitations

    The reported results make the CNN a promising classifier for WII under the evaluated conditions.

    Source: §VI · p. 6

    The prototype requires enhancement and field validation before it can become a viable coexistence-management classifier; training data also needs greater diversity for channel and hardware impairments.

    Source: §VI · p. 6

Abstract

from arXiv · show

The steadily growing use of license-free frequency bands requires reliable coexistence management for deterministic medium utilization. For interference mitigation, proper wireless interference identification (WII) is essential. In this work we propose the first WII approach based upon deep convolutional neural networks (CNNs). The CNN naively learns its features through self-optimization during an extensive data-driven GPU-based training process. We propose a CNN example which is based upon sensing snapshots with a limited duration of 12.8 {\mu}s and an acquisition bandwidth of 10 MHz. The CNN differs between 15 classes. They represent packet transmissions of IEEE 802.11 b/g, IEEE 802.15.4 and IEEE 802.15.1 with overlapping frequency channels within the 2.4 GHz ISM band. We show that the CNN outperforms state-of-the-art WII approaches and has a classification accuracy greater than 95% for signal-to-noise ratio of at least -5 dB.

I. INTRODUCTION

Wireless interference identification supports coexistence management in shared license-free bands, where interference can cause packet loss and latency. The paper proposes a deep-CNN approach, while contrasting it with compressed sensing methods aimed at finding underused spectrum rather than identifying specific radio signals.

I. INTRODUCTION

  • Interference can cause packet loss and transmission latency, making wireless interference identification relevant to coexistence management in shared license-free bands.The introduction describes heterogeneous wireless systems sharing the 2.4 GHz ISM band.Source: §I · p. 1
  • The proposed deep-CNN WII approach distinguishes 15 classes using 10 MHz sensing bandwidth and 128 IQ samples spanning 12.8 µs.The evaluation uses IEEE 802.11 b/g, IEEE 802.15.4, and IEEE 802.15.1 systems.Source: §I · p. 1

A. Compressed Sensing

  • Compressed sensing classifies frequency bands as white, gray, or black according to spectral power density, indicating where interference-free transmission is likely.Its stated goal is to find sparsely used spectrum, not to identify which radio signals are present.Source: A. Compressed Sensing · pp. 1–2
  • Compressed sensing promises faster decisions and lower hardware requirements through below-Nyquist sampling, but this entails information loss that may aid decision making.The method computes the power density spectrum from a compressed input vector in a trade-off with accuracy.Source: A. Compressed Sensing · p. 2

B. Neuro-Fuzzy Signal Classifier

The neuro-fuzzy signal classifier (NFSC) classifies frequency bands according to known wireless technologies using predefined features and decision rules. Its layered process filters a fuzzificated signal, compares it with reference shapes, and aggregates results for temporal evaluation.

B. Neuro-Fuzzy Signal Classifier

  • The NFSC is an expert system that classifies frequency bands with respect to known wireless technologies.Its purpose is similar to the proposed CNN, and the paper later compares their performance.Source: B. Neuro-Fuzzy Signal Classifier · p. 2
  • The NFSC computes the input signal’s power density spectrum, normalizes it, and filters it with a predefined shape that can exceed the target channel bandwidth.The simplest filter shape is a rectangle wider than the radio signal’s channel bandwidth.Source: B. Neuro-Fuzzy Signal Classifier · p. 2
  • It compares the filtered signal with a predefined reference shape, thresholds the similarity to classify the radio signal, and permits multiple filter and reference shapes.A statistics layer aggregates consecutive results before the interference layer decides whether a signal belongs to a frequency-hopping system.Source: B. Neuro-Fuzzy Signal Classifier · p. 2
  • Unlike the CNN’s learned feature extraction and decision process, the NFSC uses predefined features and a fixed decision process, trading flexibility for easier analysis.The paper states that the CNN’s learned decision process is more difficult to analyze.Source: B. Neuro-Fuzzy Signal Classifier · p. 2

C. Convolutional Modulation Recognition

Prior work found CNNs performed best at low SNR for modulation classification. This paper adapts that CNN as a starting point for spectral-domain WII, using time-domain IQ snapshots in the referenced study.

C. Convolutional Modulation Recognition

  • For classifying 11 modulations, CNNs achieved the best results at low SNR, while most classifier differences were comparable at high SNR.The comparison covered eight digital and three analog modulations using CNNs, deep neural networks, and several conventional classifiers.Source: C. Convolutional Modulation Recognition · p. 2
  • O’Shea et al. [9] studied complex-valued temporal radio signals, whereas this paper investigates the spectral domain and adapts their best-performing CNN as a starting point.The referenced CNN used a structure inspired by networks for visual tasks such as MNIST.Source: C. Convolutional Modulation Recognition · p. 2
  • The referenced CNN was trained on approximately 96,000 time-domain snapshots, each containing 128 IQ samples supplied as a 128 × 2 matrix.One matrix column contained I samples and the other Q samples, without information about their link being given to the network.Source: C. Convolutional Modulation Recognition · p. 2

III. GENERATION OF DATA SET

The dataset combines generated and recorded wireless packet transmissions, with restrictions on signal composition and channel conditions to keep the first prototype’s problem simple. It covers several modulation and packet variants across the studied standards.

III. GENERATION OF DATA SET

  • Recording was triggered only during transmissions, excluding interframe gaps, and VSG packets used maximum-length payloads chosen at random.The passage does not state that packet payloads were otherwise controlled.Source: §III · p. 2
  • The IEEE 802.11 b/g data varied CCK, PBCC, and OFDM modes and allowed bit rates; IEEE 802.15.1 varied transport modes and packet types, while IEEE 802.15.4 used ACK frames..Source: §III · p. 2
  • Each snapshot contained exactly one signal class, with no concurrent signals, and data generation used a single emitter.The dataset also assumed a flat-fading channel because the generator and analyzer were connected by coaxial cable.Source: §III · p. 3
  • Thermal-noise reception distortion was modeled as additive white Gaussian noise over SNRs from -20 dB to 20 dB in 2 dB steps.The noise was added during post-processing with a SIMULINK model.Source: §III · p. 3
  • Training and validation used 151,200 and 74,025 sensing snapshots, respectively.The signals were generated with a vector signal generator and recorded using a real-time spectrum analyzer.Source: §III · pp. 2–3

IV. CONVOLUTIONAL NEURAL NETWORK DESIGN

The CNN classifies allocated frequency channels for IEEE 802.11 b/g, IEEE 802.15.4, and IEEE 802.15.1 transmissions using a 10 MHz sensing bandwidth. Eight parallel classifiers are needed to cover the 2.4 GHz ISM band, with bandwidth choices also affecting class count and computation.

IV. CONVOLUTIONAL NEURAL NETWORK DESIGN

  • The CNN classifies allocated frequency channels and corresponding wireless technologies for IEEE 802.11 b/g, IEEE 802.15.4, and IEEE 802.15.1 packet transmissions.The channel classes are illustrated for an example CNN in the 2.4 GHz ISM band.Source: §IV · p. 3, Figure 1 · p. 3
  • A 10 MHz sensing bandwidth requires eight parallel CNNs to observe the full 2.4 GHz ISM band.The CNN maps technology-specific relative channel numbers to absolute channel numbers using its index and technology-specific offsets.Source: §IV · p. 3
  • The design distinguishes 15 classes: ten IEEE 802.15.1, two IEEE 802.15.4, and three IEEE 802.11 b/g channels.The illustrated CNN is centered at 2426.5 MHz; IEEE 802.11 b/g channels are only partly within its sensing bandwidth.Source: §IV · p. 3
  • Increasing sensing bandwidth increases the number of observable channels and classes, while also increasing input size, computation complexity, and required IQ samples.With snapshot duration fixed, input data size grows proportionally to bandwidth.Source: §IV · p. 3

A. Frequency-Domain Sensing Snapshots

The sensing snapshot contains 128 IQ samples over 12.8 µs, which are transformed into frequency-domain features for classification.

A. Frequency-Domain Sensing Snapshots

B. Network Structure

The CNN takes frequency-domain snapshots as input and returns scores over 15 classes; training used Adam with a fixed learning rate and 50 epochs.

B. Network Structure

  • The input is a 128 × 2 matrix representing 128 complex-valued frequency bins through separate real and imaginary floating-point values.The CNN structure was derived from O’Shea et al..Source: B. Network Structure · p. 3
  • The output is a 15-entry vector whose values indicate each class's likelihood relative to the other classes.Source: B. Network Structure · p. 3

C. Network Training

  • Training used Adam at learning rate 0.0001 for 50 epochs, with input normalization and batch size 1024; no hyperparameter optimization was performed.The setup was selected from small parameter variations, and the batch size was near the graphics-card memory limit.Source: C. Network Training · p. 3
  • Table I is the wireless-technologies channel-mapping table, while Table II presents the CNN structure.Source: Table I · p. 4, Table II · p. 4

D. Network Size Reduction

The authors reduce CNN size in relation to the number of training snapshots, aiming to improve generalization while shortening GPU training time.

D. Network Size Reduction

  • The reduced CNN was set to approximately the same number of degrees of freedom as training snapshots, following a rule intended to discourage learning noise and improve generalization.The authors note that the degrees-of-freedom rule is difficult to apply when the training-snapshot count itself must be determined.Source: D. Network Size Reduction · p. 4
  • The reduced CNN used learning rate 0.001, 200 epochs, and batch size 1024.Its structure is listed in Table III.Source: D. Network Size Reduction · p. 4, Table III · p. 4

E. Implementation

  • On the reported GPU setup, reduced-CNN training fell from about 74 to 3 seconds per epoch and from 501 to 20 ms per batch.The implementation used Keras with Theano as backend on a system with an Nvidia GTX 960 GPU.Source: E. Implementation · p. 4

V. RESULTS

On validation data, CNN accuracy varies by technology and channel, with IEEE 802.11 b/g hardest and IEEE 802.15.1 strongest. In a constrained comparison over IEEE 802.15.1 and IEEE 802.15.4, CNN outperforms NFSC across SNRs.

V. RESULTS

  • IEEE 802.11 b/g channels have the lowest CNN accuracy, and their greater channel width, modulation diversity, and bit-rate diversity make them the most complex signals here.The passage reports validation-data accuracy and points to Fig. 2.Source: §V · p. 4, Figure 2 · p. 4
  • IEEE 802.15.1 channels achieve the best accuracy, while border-channel distortion, overlap, and shared center frequencies slightly worsen recognition for four classes.Classes RCHIEEE802.15.1 0 and 9 are affected by the anti-aliasing filter; classes 3 and 8 share center frequencies with IEEE 802.15.4 classes.Source: §V · p. 4
  • CNN accuracy exceeds that of O’Shea et al.’s modulation-recognition CNN across all SNRs; the authors suggest frequency-selective classification is likely easier.They also point to significant CNN reduction without a large accuracy loss as supporting this interpretation.Source: §V · p. 5
  • The authors infer that classification may worsen when signals share a center frequency, consistent with slightly poorer recognition of overlapping IEEE 802.15.1 and IEEE 802.15.4 classes.The shared-center-frequency classes are IEEE 802.15.1 classes 3 and 8.Source: §V · pp. 4–5

A. Comparison of CNN and NFSC

  • 5.32 dB performance gain and 8.19 % classification-accuracy improvement: CNN beats NFSC across SNRs on averaged IEEE 802.15.1/802.15.4 classes; IEEE 802.11 b/g was excluded.The comparison excluded IEEE 802.11 b/g because NFSC had suboptimal out-of-band signal identification.Source: Figure 3 · p. 5, A. Comparison of CNN and NFSC · p. 5

B. Frequency- and Time-Domain Sensing Snapshots

Frequency-domain sensing snapshots outperform time-equivalent inputs, while a reduced CNN retains validation accuracy comparable to the original despite having over 99% fewer degrees of freedom. The reduced model is much faster and generalizes better.

B. Frequency- and Time-Domain Sensing Snapshots

C. Reduced CNN

  • The reduced CNN is much faster than the original, with over 99% fewer degrees of freedom.Fig. 5 presents training and validation accuracy for both CNNs.Source: C. Reduced CNN · p. 5, Figure 5 · p. 6
  • Validation accuracy remains comparable after reducing the CNN’s degrees of freedom by over 99%, while the reduced model generalizes better.The original CNN’s training accuracy is higher, especially at low SNR, but the authors interpret that gap as memorization of sensing patterns including noise.Source: C. Reduced CNN · p. 5

VI. CONCLUSION

The paper proposes a deep CNN for wireless interference identification using short, bandwidth-limited snapshots from three wireless standards. It reports high accuracy at low SNR and better performance than an NFSC, while noting that field validation and broader training data remain necessary.

VI. CONCLUSION

  • The paper proposes the first deep-CNN approach to WII, with features learned through data-driven self-optimization and a network reduced by 99% of total degrees of freedom.The authors report better generalization during training after reducing the network size.Source: §VI · p. 6
  • The CNN classifies 15 classes using snapshots lasting 12.8 µs with 10 MHz acquisition bandwidth, covering IEEE 802.11 b/g, IEEE 802.15.4, and IEEE 802.15.1 transmissions.The evaluation used 151,200 training snapshots and 74,025 validation snapshots spanning 19 modulation-type and symbol-rate variants.Source: §VI · p. 6
  • Average classification accuracy exceeds 95% for SNRs greater than -5 dB, and the CNN outperforms NFSC by at least 5.32 dB average performance gain and 8.19% classification accuracy improvement.The reported results also note performance drops for wideband signals clipped by the limited acquisition bandwidth.Source: §VI · p. 6
  • The prototype must be enhanced and validated in the field before it can become a viable coexistence-management classifier.The authors also call for more diverse training data to better represent channel and hardware impairments, and for a CNN suitable for multi-label identification of concurrent transmissions.Source: §VI · p. 6
Loading 1703.00737v1…