MTFNet: A Lightweight Multi-Scale Time–Frequency Neural Network for Multi-Class Event Recognition in DAS-Based Perimeter Security

Abstract

Distributed Acoustic Sensing (DAS) has emerged as a pivotal technology for perimeter security. However, existing DAS event recognition methods have inherent limitations. Single-domain modeling has limited feature representation, while conventional multi-domain fusion methods require converting one-dimensional signals into two-dimensional images. This conversion adds computational overhead and may discard part of the original phase and amplitude information. In addition, the robustness of lightweight recognizers under low signal-to-noise ratio (SNR) conditions remains insufficiently examined. This paper proposes a lightweight multi-scale time–frequency network (MTFNet) for direct processing of one-dimensional DAS signals. The model adopts a parallel time–frequency dual-branch architecture. The time-domain branch utilizes a lightweight 1D Inception-ResNet variant to capture short-term impulse features and long-range temporal dependencies via multi-scale parallel convolutions and residual connections. The frequency-domain branch performs multi-scale segmented spectral analysis to extract complementary discriminative information at diverse time and frequency resolutions, followed by a lightweight attention-free fusion module for effective time–frequency feature fusion. Experimental results on a self-built dataset covering Wind, Facility Noise, Hammering, Sawing, and Walking events show that MTFNet achieves a recognition accuracy of 99.55% with 0.12 M parameters, outperforming SVM, 1D-CNN, TCN, MS 1D-CNN, CNN-BiLSTM, and attention-based multimodal fusion methods. It maintains an accuracy of 94.06% at an SNR of 0 dB, demonstrating its robustness against additive white Gaussian noise (AWGN) and practical engineering application potential.