<![CDATA[Pipeline networks are critical infrastructure for oil and gas transport because the occurrence of leaks can rapidly escalate into safety, economic, and environmental crises. Operators are practically required to identify the presence and type of leaks; however, applying multiclass recognition is challenging when labeled data and computing power are limited. Therefore, this study proposes a three-stage pipeline which consists of: (1) adopting the GPLA-12 dataset of acoustic or vibration signals spanning 12 leak types; (2) extracting multi-domain features by combining time-domain descriptors with Power Spectral Density (PSD)-based spectral features; and (3) applying a genetic algorithm (GA) as a wrapper for feature selection to enhance discriminability and reduce dimensionality, which was followed by benchmarking seven conventional classifiers and GA-based refinement of the top model with a focus on the feature subset and hyperparameters. A maximum accuracy of 96.35% was achieved on the GPLA-12 dataset with low computation time and a simple model architecture. The proposed pipeline also attained similar or better accuracy at substantially lower complexity and data requirements compared with prior deep CNN approaches. These results support timely multiclass decision-making in resource-constrained industrial settings. A key observation was that the focus was on supervised leak-type classification from acoustic or vibration signals, while localization, severity estimation, and multi-sensor fusion were beyond the scope of this study.]]>
Copyrights © 2026
