Hybrid deep learning framework for acoustic emission-based gas pipeline leak detection

Abstract Natural gas pipeline leak detection faces significant challenges, including high-intensity environmental noise, complex leakage signal characteristics, and difficulties in time-frequency domain feature fusion. To address these challenges, this paper proposes a pipeline leakage detection model based on dual-path feature extraction and long sequence modeling (MFCC-MSRNet-BiLSTM), aiming to improve the accuracy and robustness of pipeline leakage fault detection. The model innovatively integrates time-domain features from raw Acoustic emission (AE) signals with frequency-domain Mel-frequency cepstral coefficient (MFCC) features through a dual-path feature extraction architecture, enabling collaborative time-frequency information processing. The time-domain branch employs a lightweight Multi-scale residual network (MSRNet) that incorporates multi-scale depthwise separable convolutional modules, Multi-head self-attention mechanisms (MSA), and contraction modules. This architecture achieves multi-granularity feature extraction, long-range dependency modeling, and noise suppression. The frequency-domain branch extracts MFCC features through a dedicated processing network to capture signal cepstral characteristics in the frequency domain. After coupling the dual-branch features through a fusion layer, a Bidirectional long short-term memory (BiLSTM) network captures temporal dynamic characteristics for fault classification. validated on the public GPLA-12 dataset, achieving 96.52% fault detection accuracy and outperforming traditional classification methods. This research provides a novel approach for AE signal-based natural gas pipeline leak detection.