An Explainable Transfer Learning Framework Based on Acoustic Spectral Portraits for Pipeline Leak Diagnosis Across...

This study proposes an explainable and lightweight pipeline leak-diagnosis framework based on a leak-oriented Acoustic Spectral Portrait (ASP) processing pipeline and transfer learning. The ASP processing pipeline converts one-dimensional acoustic emission signals into compact two-dimensional spectral representations through AR pre-whitening, overlapping Hann-window segmentation, zero-padded fast Fourier transform analysis, logarithmic scaling, percentile-based normalization, and temporal spectral aggregation. These spectral portraits are processed by a lightweight convolutional neural network, in which source-domain knowledge is adapted using limited labeled target-domain data. Gradient-weighted class activation mapping is further employed to provide qualitative post hoc interpretation of the spectral regions contributing to model predictions, rather than direct physical validation, supported by supplementary perturbation, model-randomization, and attribution-stability analyses. The framework was evaluated using three gas- and water-pipeline datasets through six directed source-to-target transfer tasks, with each experiment repeated 10 times using mutually exclusive data partitions. Scenario-averaged classification accuracies of 98.36%, 97.09%, and 98.58% were achieved across the three transfer scenarios. For the representative DS-1 to DS-2 transfer task, transfer learning improved target-domain accuracy from 91.80% to 96.85% while reducing target-model training time from 17.3 min for training from scratch to 9.2 min for fine-tuning. The proposed spectral portrait representation also achieved a mean scenario-averaged classification accuracy of 98.01%, outperforming conventional time-frequency representations. These results demonstrate effective source-to-target adaptation for pipeline leak diagnosis across the three investigated composite operating and acquisition domains.