PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative AI Enabled Robust Data Augmentation for Wireless Sensing in ISAC Networks

Jiacheng Wang, Changyuan Zhao, Hongyang Du, Geng Sun, Jiawen Kang, Shiwen Mao, Dusit Niyato, Dong In Kim

arXiv:2502.12622v1eess.SP

TL;DR

ISAC sensing needs substantial, high-quality CSI data, but collection is costly, labor-intensive, and privacy-sensitive. The paper proposes DiRA, combining diffusion-based sample expansion and quality enhancement with CSI-derived acceleration-jerk features, and reports up to 70% improved target-detection performance.

  • Problem

    Reliable sensing-model training requires extensive CSI samples, while data collection is costly, labor-intensive, and privacy-sensitive.

  • Method

    The paper proposes DiRA, which expands sensing samples, enhances their quality, and augments CSI-derived acceleration and jerk parameters for target detection.

  • Results

    DiRA increases average target-detection accuracy by 0.29, representing a 70% improvement compared with using no augmentation.

  • Takeaways & Limitations

    The results support DiRA as a robust data-augmentation approach for sensing tasks when original ISAC samples are scarce or unevenly distributed.

Abstract

from arXiv · show

Integrated sensing and communication (ISAC) uses the same software and hardware resources to achieve both communication and sensing functionalities. Thus, it stands as one of the core technologies of 6G and has garnered significant attention in recent years. In ISAC systems, a variety of machine learning models are trained to analyze and identify signal patterns, thereby ensuring reliable sensing and communications. However, considering factors such as communication rates, costs, and privacy, collecting sufficient training data from various ISAC scenarios for these models is impractical. Hence, this paper introduces a generative AI (GenAI) enabled robust data augmentation scheme. The scheme first employs a conditioned diffusion model trained on a limited amount of collected CSI data to generate new samples, thereby expanding the sample quantity. Building on this, the scheme further utilizes another diffusion model to enhance the sample quality, thereby facilitating the data augmentation in scenarios where the original sensing data is insufficient and unevenly distributed. Moreover, we propose a novel algorithm to estimate the acceleration and jerk of signal propagation path length changes from CSI. We then use the proposed scheme to enhance the estimated parameters and detect the number of targets based on the enhanced data. The evaluation reveals that our scheme improves the detection performance by up to 70%, demonstrating reliability and robustness, which supports the deployment and practical use of the ISAC network.

I. INTRODUCTION

ISAC sensing relies on machine-learning models but faces costly, time-consuming, privacy-sensitive CSI collection and insufficient, unevenly distributed training data. This paper proposes DiRA, which expands and enhances sensing data, extracts acceleration and jerk parameters from CSI, and improves target-count detection.

  • ISAC combines communication and sensing using shared network and hardware resources, making it a fundamental 6G technology.
  • Training reliable sensing models requires large, high-quality CSI datasets, but collecting them demands costly equipment, substantial labor, and user participation.User participation also raises privacy concerns because CSI can reveal physiological information such as gait, breathing, and heart rate.
  • DiRA combines diffusion-based sample quantity expansion with quality enhancement to augment sensing data when original samples are scarce or unevenly distributed.Its quantity module generates new samples, while its quality module further improves generated-sample quality.
  • The paper proposes estimating acceleration and jerk of signal propagation path-length changes from CSI for wireless sensing.These parameters capture characteristics associated with user activities, behaviors, and location changes.
  • DiRA augments estimated signal parameters and trains machine-learning models for target detection, increasing average detection accuracy by 0.29, or 70% compared with no augmentation.
  • The proposed pipeline extracts CSI spectra, augments them with DiRA, and uses the augmented spectra to train a sensing model that detects the number of targets.During online operation, input sensing data passes through quality enhancement before entering the sensing model.

B. GenAI-based Signal Augmentation

The paper combines CSI-based estimation of propagation-path acceleration and jerk with generative models that produce and refine signal data for ISAC sensing.

  • Diffusion, VAE, and GAN models have been used to generate or enhance CSI, RF, received-signal-strength, and time-series signal data.
  • Signal Parameter Estimation: The proposed method estimates acceleration and jerk from CSI by analyzing phase accumulation caused by dynamic propagation-path changes.Static paths are suppressed through frequency-domain processing before correlation and spectral analysis.
  • Signal Parameter Estimation: A keystone transformation decouples time and lag variables before two-dimensional FFT processing forms an acceleration-jerk spectrum.The transformation resolves coupling between Δt and τ, allowing signal energy to accumulate for spectrum formation.
  • Signal Parameter Estimation: Each dynamic propagation path produces a peak in the acceleration-jerk spectrum, while cross-terms can also appear under specific path conditions.The cross-term interference condition is associated with equal path-length-change jerk values, r1 = r2.
  • Signal Parameter Estimation: The number of acceleration-jerk spectrum peaks reflects the number of dynamic targets, supporting target-count detection and applications such as fall detection.

B. Spectrum Augmentation

The proposed DiRA scheme augments acceleration-jerk spectra in two stages: it first expands sample quantity and then purifies generated spectra to improve their quality.

  • DiRA trains a conditional diffusion model on collected acceleration-jerk spectra to expand the available training data.A second purification stage addresses noise in spectra generated from insufficient original samples.

1) Spectrum Quantity Expansion:

The quantity-expansion module uses diffusion’s forward noising and learned reverse denoising processes to generate conditional acceleration-jerk spectra from limited datasets.

  • The diffusion process gradually adds noise to acceleration-jerk spectra, while the reverse process iteratively removes noise to generate new spectra.
  • A parameterized model pθ(Pt−1|Pt) estimates the difficult reverse transition used to denoise sampled Gaussian noise.
  • Starting from PT sampled from N(0, I), the reverse process iteratively samples Pt−1 for t = T, …, 1 to generate a new spectrum.
  • Conditional generation trains on paired spectrum-condition samples and incorporates the condition into the reverse-process distribution.
  • For limited datasets, the trained conditional diffusion model generates additional spectra to increase the overall sample quantity.

2) Spectrum Quality Enhancement:

DiRA enhances generated AJ spectra by injecting and then removing noise with a second diffusion process guided toward the conditioned spectrum. The guidance strength and injected-noise threshold balance denoising against fidelity to the underlying spectrum.

  • Spectrum Quality Enhancement: DiRA uses a second diffusion model to remove inherent noise from generated AJ spectra while preserving their useful characteristics.The forward process adds Gaussian noise, and the reverse process removes both injected and inherent noise.
  • Spectrum Quality Enhancement: The threshold T* controls the injected noise needed to perturb noise κ while preserving the noise-free spectrum P_zr.If T* is too large, the output may deviate from P_zr; if too small, κ may remain.
  • Spectrum Quality Enhancement: Conditional guidance uses the noisy spectrum P′c to encourage denoised outputs to approach the condition while reducing noise.The guidance term is based on an MSE distance evaluator.
  • Spectrum Quality Enhancement: The guidance scale s′ trades off denoising strength against retaining noise from the condition P′c.Larger s′ is used after more diffusion noise has disrupted the condition; smaller s′ is used when P′c remains noisy.
  • Spectrum Quality Enhancement: The enhancement algorithm performs W rounds of enhancement for each spectrum.The paper summarizes this procedure in Algorithm 1.

C. Target Detection

The target-detection pipeline expands and enhances AJ spectra, then trains a ResNet-based classifier to detect the number of dynamic targets. Conditional generation equalizes category sizes before quality enhancement and classification.

  • C. Target Detection: The classifier detects dynamic-target counts from augmented AJ spectra because AJ-spectrum peak counts equal the number of dynamic targets.The classifier is ResNet-based.
  • C. Target Detection: Conditional diffusion expands AJ-spectrum quantities to equalize uneven category sizes, ensuring every category contains generated and collected spectra.An example expands categories with 1,000 and 100 original spectra to 2,000 each.
  • C. Target Detection: DiRA uses conditional embeddings for sample quantity expansion and time-only embeddings for sample quality enhancement.The diffusion architecture includes these different conditioning choices for the two stages.
  • C. Target Detection: The quality-enhancement procedure is summarized as Spectrum Quality Enhancement and returns the enhanced spectrum P′0.Its input includes a distance gradient, gradient scale s′, and P′c.
  • C. Target Detection: The target classifier uses ResNet-18, comprising one convolutional layer, two pooling layers, eight residual units, and one fully connected layer.Each residual unit contains two 3x3 convolutional layers and uses ReLU activation.
  • C. Target Detection: After training, the ResNet classifies newly collected AJ spectra to improve target-detection accuracy.The model is trained on enhanced spectra before recognizing AJ spectra.

IV. IMPLEMENTATION AND EVALUATION

The evaluation validates AJ-spectrum estimation, examines DiRA’s quantity and quality enhancement, and tests augmented-data robustness through cross-scenario target-count detection.

  • IV. IMPLEMENTATION AND EVALUATION: The evaluation covers AJ estimation, DiRA’s spectrum quantity and quality enhancement, and ResNet-based detection of target counts.These experiments assess augmented data for ISAC sensing support.
  • IV. IMPLEMENTATION AND EVALUATION: AJ spectra are extracted from CSI collected in scenarios with different numbers of targets.Figure 4 presents the resulting spectra.
  • IV. IMPLEMENTATION AND EVALUATION: The evaluation uses cross-validation across various scenarios to test DiRA’s support for ISAC network sensing capabilities.A ResNet-18 is trained with different datasets for target-count detection.

A. Experimental Configurations

The experiments collect CSI with commercial IEEE 802.11 access points using 4x4 MIMO, two frequency bands, directional antennas, and 256 subcarriers at 5.805 GHz.

  • A. Experimental Configurations: CSI is collected using multiple commercial IEEE 802.11 access points supporting four radio-frequency channels and up to 4x4 MIMO transmission.The platform covers both 2 GHz and 5 GHz frequency bands.
  • A. Experimental Configurations: The experiment uses 5.805 GHz signals with 80 MHz bandwidth across 256 subcarriers and directional antennas to reduce noise.The collected data is processed after measurement.

B. Experimental Method

The experiments evaluate DiRA’s AJ-spectrum generation and quality enhancement using imbalanced or limited datasets, with FID measuring distribution similarity and cross-validation assessing target-detection performance.

  • The study collects CSI from scenarios with varying numbers of dynamic targets and computes AJ spectra for augmentation evaluation.The AJ spectra support subsequent generation and target-detection experiments.
  • A conditional diffusion model is trained on an imbalanced dataset to generate 200 new AJ spectra for sample-generation assessment.The experiment uses 150,000 training steps.
  • DiRA’s generative performance is tested with conditional diffusion models trained for 100,000 steps on datasets containing different total sample numbers.
  • FID measures distribution similarity between generated and original samples, with lower values indicating closer agreement with collected spectra.
  • ResNet-18 cross-validation evaluates target-number detection across different scenarios using datasets produced by the augmentation process.

1) Spectrum Analysis:

The proposed AJ-spectrum analysis extracts acceleration and jerk from CSI across scenarios with different target counts, then compares DiRA’s generated spectra with CGAN using FID and generation visualizations.

  • Spectrum Analysis: The proposed algorithm computes AJ spectra from CSI collected in scenarios containing 3 to 7 targets.
  • Spectrum Analysis: Acceleration and jerk differ across targets because their locations, movement directions, and speeds differ, enabling target-related patterns in the AJ spectrum.
  • Spectrum Analysis: More targets introduce additional interference and cross-terms, increasing noise while still allowing target counts and related parameters to be identified.
  • Spectrum Analysis: DiRA and CGAN generate AJ spectra for five scenarios with 3 to 7 dynamic targets using an uneven dataset of 1,800, 1,600, 1,400, 1,200, and 1,000 samples.
  • Spectrum Analysis: The generation process includes outputs at 1,000, 800, 600, and 400 steps alongside CGAN outputs for three- and seven-target scenarios.
  • Spectrum Analysis: Figure 5 presents AJ-spectrum generation and CGAN comparisons across scenarios, with acceleration on the x-axis and jerk on the y-axis.
  • Spectrum Analysis: Figure 6 compares DiRA and CGAN using FID for AJ-spectrum generation.

3) Quality Enhancement:

DiRA enhances AJ-spectrum quality after quantity expansion by reducing noise while retaining target-related peaks, and the resulting augmented data improves target-number detection across scenarios.

  • Quality Enhancement: After quantity expansion, DiRA evaluates AJ-spectrum quality enhancement through a process that adds and removes noise.
  • Quality Enhancement: The enhanced spectra retain peaks associated with dynamic targets and their locations while exhibiting less noise than the originals.
  • Quality Enhancement: Quality enhancement brings original and generated AJ spectra closer to noise-free reference spectra, with average FIDs of 181.172, 172.94, and 178.094 for the original, generated, and overall samples.
  • Quality Enhancement: Using T ∗ = 50 and W = 2 yields higher enhancement FIDs than 50-step and 100-step enhancement while retaining original-sample features and reducing noise.
  • Target Detection: ResNet-18 target-number detection reaches accuracies of 0.67, 0.71, and 0.71 across three scenarios with DiRA-augmented data.
  • Target Detection: DiRA-augmented data improves detection performance by up to 70% compared with training without DiRA and shows less fluctuation across scenarios than 50- and 100-step enhancements.
  • Conclusion: The paper concludes that DiRA combines sample quantity expansion and quality enhancement, supporting robust AJ-spectrum augmentation and target-number detection in ISAC networks.
Loading 2502.12622v1…