Cross-attention multi-source fusion and domain adaptation strategy for cross-domain mechanical seals fault diagnosis

Acoustic emission (AE), seal face temperature, and torque signals are key non-destructive testing (NDT) data sources for assessing mechanical seal health. Under varying operating conditions (e.g., speed and load), NDT signal characteristics change substantially, causing diagnostic models trained in a single domain to perform poorly when transferred across domains. Unimodal analysis is also insufficient to capture holistic features under complex faults. To address these issues, MS-MLDA is developed as a framework that combines multi-source fusion with multi-level domain adaptation for robust cross-domain fault identification. BispectralNet extracts spatial-frequency features from the bispectral amplitude matrix of AE signals, while a hybrid 1D-CNN-Transformer encoder captures dynamic patterns in multi-channel time-series signals (AE RMS, friction torque, and temperature). Cross-attention then fuses the two modalities into a comprehensive state representation. For enhanced transferability, adaptive bandwidth maximum mean discrepancy (ABMMD) aligns source and target feature distributions at multiple levels before and after fusion. A multifunctional test bench for contact mechanical seals is established, and a multi-source NDT dataset covering diverse operating conditions and fault types is collected using AE, torque, and thermocouple sensors. Experimental results show that MS-MLDA achieves 93.93% classification accuracy under the largest domain shift, outperforming multiple baselines and existing domain adaptation methods.