Intelligent diagnosis of gas pipeline condition through multivariate analysis of acoustic emission signal-based imaging

Intelligent gas-pipeline condition monitoring is vital for managing risk and reducing costs, but traditional methods using acoustic emissions (AE) encounter challenges like fluid-pressure changes, flange vibrations, and inconsistent leak length, resulting in unreliable outcomes. This research proposes an intelligent solution that overcomes these obstacles, enabling real-time-accurate monitoring. The initial solution involves resampling and segmenting AE signals, followed by two-parallel processes to enhance the density of AE features. The first process involves stacking the segments to create a 2D-acoustic-time representation (ATR), while the second process uses the Hilbert transform to derive the relative magnitudes of frequency components from the envelope-power-spectrum from each segment. This results in a 2D-acoustic-frequency representation (AFR). A data-level-fusion is then proposed by combining the ATR and AFR to create the multivariate-acoustic-imaging-representation (MAIR), which serves as input for the multivariate-convolutional neural network (MCNN). Unlike previous research in this area, MAIR enables image-based input to MCNN, allowing the utilisation of image augmentation for the first time. This mitigates data limitations and enhances the generalisation performance of the model before training. Testing on the GPLA-12 dataset demonstrates the robustness of the proposed approach, achieving a pipeline-leak detection accuracy of 97.88% for 12 classes, surpassing state-of-the-art methods by at least 8.47%.