Joint Denoising and Feature Selection for Leakage Identification in Mine Air-Supply Pipelines

This study investigates the identification of leakage in mineral air-supply pipelines using vibration signals. To address the issue of severe environmental noise and high signal dimensionality, a unified framework integrating joint denoising, feature selection, and leakage-type identification is proposed. First, the dung beetle optimization (DBO) algorithm adaptively tunes the parameters of variational mode decomposition (VMD) to achieve effective signal decomposition. The resulting modes are grouped by energy and kurtosis into transient leakage modes and major/minor sustained-leakage modes. An improved wavelet packet decomposition (WPD) is then applied to denoise the minor modes, which are reconstructed with the remaining modes to obtain a clean signal. Next, a wrapper feature selection method based on an improved binary DBO (IBDBO) is introduced. The Bernoulli map and extreme gradient boosting (XGBoost) are used to initialize an elite population, and a grey wolf optimization mechanism is incorporated to avoid premature convergence. Finally, the selected features are fed into a random forest to identify the types of leakage. Experiments show that the proposed method achieves an average accuracy of 95.68% with only 4.35 selected features on average, effectively improving leakage identification while reducing the number of features.