Pattern Recognition Algorithm in DAS Using Physics-Informed Neural Networks

Wang, Haotian

doi:10.3788/AOS252086

Pattern Recognition Algorithm in DAS Using Physics-Informed Neural Networks

2026-04-01

0

PRE

AI

DOI：10.3788/AOS252086

Original

Save

Abstract

En 中文

Objective Distributed acoustic sensing (DAS) provides long-range, continuous vibration monitoring by interrogating phase and intensity perturbations in Rayleigh backscattering along optical fibers. In conveyor-belt health monitoring and other industrial deployments, DAS signals are often corrupted by strong environmental noise and exhibit pronounced nonstationarity, with time-varying spectral components and intermittent transient impulses. At the same time, labeled fault samples are typically scarce, and the most discriminative cues can shift across domains: some faults are more separable in the raw waveform through periodicity or amplitude modulation, while others are more evident in localized time-frequency energy patterns. Conventional recognition pipelines that rely on fixed preprocessing, handcrafted features, or a single representation domain therefore struggle to maintain accuracy and generalization under small-sample training and distribution shift. This study aims to develop an end-to-end, cross-domain learning framework that can adaptively exploit both time-domain and time-frequency-domain information for robust DAS fault pattern recognition, while remaining configurable for real-world deployment constraints. Methods We propose a short-time Fourier neural operator (STFNO) framework that combines differentiable time-frequency transforms with operator-style feature learning and cross-domain fusion. Given a raw DAS time-series segment, STFNO first applies a short-time Fourier transform (STFT) to obtain a localized time-frequency representation, enabling the model to capture transient events and time-dependent spectral changes that are difficult to represent with global spectral operations alone. Next, a learnable operator-style module is applied on the time-frequency plane to extract multi-scale structures and to enhance informative regions, yielding refined time-frequency features. To align these enhanced cues with waveform-level decision making and to preserve signal fidelity, STFNO performs an inverse STFT (ISTFT) to project the refined representation back into the time domain, producing a residual signal that encodes complementary information discovered in the time-frequency branch. This residual is then fused with the original waveform using a residual connection, after which a lightweight time-domain classification head with global pooling outputs the final fault label. To quantify the value of cross-domain collaboration, we construct representative baselines: an FNO-based fusion network operating on time-series inputs, a purely time-domain network, a time-frequency classification network that treats spectrograms as grayscale inputs, and two standard vision backbones (ResNet-18 and Inception v3) trained on grayscale time-frequency images. Experiments are conducted on a conveyor-belt fault dataset containing normal operation and multiple fault types, including roller eccentricity, roller rupture, and opening rupture. Typical waveform patterns and their corresponding time-frequency distributions are provided. To emphasize practical small-sample learning, we adopt a challenging protocol with 10 labeled training samples per class and 190 test samples per class for both three-class and four-class recognition tasks. Results and Discussions On the simulated three-class task, STFNO achieves the highest overall accuracy of 97.02%, outperforming all baselines. The gain is most pronounced for roller eccentricity, whose signature appears as a highly concentrated, narrowband energy ridge in the time-frequency representation. By explicitly enhancing localized structures on the time-frequency plane and then projecting the refined information back to the waveform through ISTFT-based residual fusion, STFNO effectively converts such spectral cues into waveform-level evidence for classification. Meanwhile, normal operation is more strongly characterized by regular temporal patterns, which are naturally captured in the time domain; the cross-domain design therefore mitigates the domain preference discrepancy across classes and yields more balanced performance under small-sample training. In the more challenging four-class setting that additionally includes opening rupture, STFNO still delivers the best overall performance, reaching 85.18% accuracy. The dominant failure mode for all methods is the mutual confusion between roller rupture and opening rupture, which exhibits substantial overlap in both waveform morphology and time-frequency structure. Even under this low separability regime, STFNO improves overall class balance relative to single-domain models and vision-backbone baselines, indicating that cross-domain fusion helps aggregate weak but complementary discriminative cues that are insufficient in isolation. Ablation results confirm the functional necessity of each major component. Removing the time-frequency enhancement module yields the largest degradation, demonstrating that learnable refinement in the time-frequency domain is central for nonstationary DAS signals. Excluding the ISTFT projection or residual fusion also reduces accuracy, showing that returning enhanced information to the time domain and preserving original waveform fidelity are both critical for stable, effective inference. Eliminating the STFT stage collapses the framework into a single-domain model and produces the weakest results, underscoring the importance of time-frequency localization. Complementary evidence from the multi-domain validation further indicates that domain utility is fault dependent: spectrally distinctive faults benefit more from frequency-related cues, whereas normal operation relies more on temporal regularity, supporting the motivation for adaptive cross-domain integration. Field experiments highlight additional distribution shift in noise level, amplitude statistics, impulse density, and energy dispersion; representative differences between simulated and site data are shown. Under these conditions, the configuration study provides a practical deployment guideline: with a tight parameter budget, allocating more capacity to the time-frequency branch tends to be more stable; with larger budgets, shifting capacity toward the time-domain branch improves performance by enabling richer sequential modeling and stronger robustness to operational variability. Conclusions This work presents STFNO, a cross-domain operator learning framework for DAS-based fault pattern recognition that addresses the combined challenges of nonstationarity, heavy noise, limited labeled samples, and distribution shift. By integrating differentiable STFT/ISTFT transforms, learnable time-frequency operator enhancement, and residual fusion into an end-to-end classifier, STFNO achieves 97.02% accuracy on a small-sample three-class conveyor-belt fault task and the best overall performance of 85.18% on a four-class task. Ablation studies confirm the indispensability of time-frequency enhancement, ISTFT projection, and residual fusion. Field evaluations further demonstrate that STFNO remains effective under distribution shift and provides actionable configuration guidance across model sizes. Overall, STFNO offers a robust and engineering-oriented solution for industrial DAS fault diagnosis where discriminative information may vary across domains and labeled data are limited.

Keywords:

optical fiber sensing

physics-informed neural network

pattern recognition

Pattern Recognition Algorithm in DAS Using Physics-Informed Neural Networks

Abstract

Journal

Organization

Cited Papers