MTLD-CS: Multi-Task Deep Learning-Based Fault Diagnosis Integrating Interpretable Classification and Sparse Waveform Segmentation for Mechanical Systems
Main article
Abstract
Intelligent fault diagnosis of rotating machinery has achieved remarkable success under laboratory conditions, yet real-world industrial deployments continue to face critical performance limitations: existing diagnosis approaches degrade substantially under low signal-to-noise ratio (SNR) conditions, fail to generalize across varying operating conditions due to data distribution shifts, and provide limited interpretability that restricts engineering trust and deployment acceptance. Existing ante-hoc and post-hoc interpretable diagnosis approaches have largely been studied in isolation, preventing synergistic benefits from their combination. To address these interconnected challenges, this paper proposes MTLD-CS, a Multi-Task Deep Learning-based Fault Diagnosis framework that effectively integrates the advantages of interpretable classification and sparse fault feature segmentation within a unified multi-task learning objective. MTLD-CS comprises three key modules: (1) a feature extraction module incorporating a wavelet pooling layer alongside residual blocks and dense atrous convolution blocks that facilitate multi-level feature representation learning robust to noise; (2) a sparse segmentation module that enhances fault feature identification through feature decoding blocks and a class-weighted Lovász-softmax loss function optimized for sparse fault signal patterns; and (3) an interpretable fault diagnosis module that combines ante-hoc interpretability with class activation mapping to provide both accurate fault classification and physically meaningful fault localization. The multi-task joint training objective synergistically optimizes classification accuracy and segmentation precision through a weighted combination of cross-entropy classification loss and class-weighted Lovász-softmax segmentation loss. Experimental evaluation on the Case Western Reserve University (CWRU) bearing dataset and an additional industrial planetary gearbox dataset demonstrates that MTLD-CS achieves 97.6% fault classification accuracy at 0 dB SNR, 89.8% at -4 dB SNR, and 82.4% segmentation IoU, outperforming five state-of-the-art deep learning baselines. Cross-condition evaluation confirms 88.9% accuracy on unseen operating conditions, demonstrating substantially better domain generalization than non-adaptive alternatives.
