期刊:2021 IEEE 10th Data Driven Control and Learning Systems Conference (DDCLS)日期:2026-05-01卷期号:: 347-352
标识
DOI:10.1109/ddcls71227.2026.11610170
摘要
To address the core challenges of data distribution discrepancy and target domain label scarcity in high-speed train axle box bearing fault diagnosis, this study proposes a multimodal feature-driven adversarial domain adaptation framework. Based on national bearing standards, target domain parameters are assumed to calculate fault characteristic frequencies, and fault-sensitive features are systematically extracted from time, frequency, and time-frequency domains with wavelet transform and short-time Fourier transform. High-precision source domain models are built using LightGBM, XGBoost, BiLSTM-Attention, and ResNet18, achieving nearly 100% accuracy. Domain adaptation technologies such as DANN and MMD enable knowledge transfer from test bench to real operational data, completing fault classification for 16 unlabeled target domain samples. Model interpretability is enhanced through MMD loss monitoring, t-SNE visualization, attention mechanism, and Grad-CAM, aligning decisions with bearing fault physical mechanisms. This framework provides an effective solution for intelligent operation and maintenance of industrial equipment under data-scarce scenarios.