Research on Infrasound-based Pipeline Leakage Signal Identification Using Random Forest

To enhance the identification of infrasound leakage signals, particularly improving the detection and recognition rate of persistent small-scale infrasound leakage signals, this study constructs a multi-feature collaborative leakage identification system based on infrasound signals. In the research process, characteristic parameters of leaked infrasound signals—including time-domain, frequency-domain, and waveform parameters—are first extracted. The random forest algorithm is then employed to rank the importance of these features, enabling preliminary screening of optimal features. Subsequently, based on the eigenvalue distribution law of non-leakage signals, the 3σ criterion is utilized to establish an adaptive dynamic threshold range for non-leakage signal features, with secondary feature selection conducted in conjunction with multi-feature collaboration. Finally, a binary classification framework is constructed to achieve leakage discrimination. Experimental results demonstrate that the optimal feature combination for sudden leakage is peak-to-peak value and frequency mean, yielding a discriminant accuracy of 99.5%. For persistent leakage, the optimal feature combination is variance and frequency root mean square, with a discriminant accuracy of 96%. Additionally, leakage aperture exerts a significant influence on model performance: as the leakage aperture increases, the discriminant accuracy exhibits a stepwise upward trend.