The complex and variable engineering environment in underground mines often leads to a mapping distribution bias between the training source domain and the application target domain. Moreover, models trained on a single modality struggle to fully exploit the potential collaboration across different feature spaces, resulting in difficulties in constructing clear decision boundaries between classes, which in turn affects the reliability and robustness of fault diagnosis. To address these issues, this paper proposes a fault diagnosis method for electromechanical equipment in underground mines based on time-frequency domain collaborative adaptation. This method optimizes the decision boundaries between different fault states by fitting the domain-invariant feature distributions in both the time and frequency domains, thereby achieving high-performance unsupervised fault diagnosis in the target domain. Specifically, we employ the Local Maximum Mean Discrepancy (LMMD) algorithm to measure the feature distribution distance between the source and target domains, and utilize a domain adversarial network to extract shared domain-invariant features for distribution alignment. Additionally, to fully leverage the advantages of time-frequency collaboration, a dual-classifier is established to accurately distinguish between different fault categories. Finally, experimental results based on multiple public datasets demonstrate that the proposed method significantly enhances the reliability and generalization capability of fault diagnosis.

Domain Adaptation,Fault Diagnosis,Rotating Machinery,Deep Learning,Local Maximum Mean Discrepancy