Small leak detection of gas pipelines with Φ-OTDR using an improved MobileNetV3-BiLSTM network

This paper presents a lightweight deep learning approach for detecting small leaks in buried gas pipelines using a phase-sensitive optical time domain reflectometry (Φ-OTDR) system. Vibration signals acquired from a distributed optical fiber vibration sensor (DOFVS) are converted into time–frequency images via continuous wavelet transform (CWT) to facilitate joint spatiotemporal feature extraction. An improved model that integrates a lightweight MobileNetV3 backbone with a bidirectional LSTM (BiLSTM) module is proposed to capture both spatial patterns from individual CWT images and temporal dependencies across consecutive frames. Experimental results demonstrate an overall accuracy of 95.96% in classifying four leak pressure levels, including the reliable detection of leaks as small as 1/16 inch under pressures as low as 0.1 MPa. Compared to conventional CNN models, an improved recognition accuracy of 8%–12% is demonstrated for the identification of low-pressure and small-aperture leakage conditions.