Natural Gas Pipeline Leak Detection Based on Dual Feature Drift in Acoustic Signals

Detecting leaks in natural gas pipelines using acoustic signals typically requires extensive prior knowledge and complex parameter designs, making it challenging to handle background noise and data distribution disparities simultaneously. This article proposes a dual-feature drift framework utilizing a nonparametric design approach for acoustic signal-based leak detection. This framework consists of two core technologies: first, feature backward normalization. Low-dimensional drift factors are designed based on transformed acoustic signals to exponentially normalize the time-periodic features of the signal feature matrix, thereby eliminating strong background noise. Second, constructing the feature drift layer within a one-dimensional convolutional neural network. Weighted parameters constrain a high-dimensional feature matrix, developing drift factors that perform exponential drift on each feature, thus enhancing gradient constraints and eliminating data distribution differences during model training. This framework achieves a fault identification accuracy of 95.46% for natural gas pipeline leaks, outperforming competing methods and representing a novel approach to intelligent pipeline leak detection.