Bearing fault diagnosis method based on unsupervised pseudo-labels adapted for Locality Preserving Projections

Currently, the detection of anomalous machine operations and the identification of faulty machine components are essential for maintaining the stability of manufacturing processes and ensuring product quality. In engineering applications, variations in environmental and operating conditions significantly influence the performance and efficiency of fault monitoring models. A bearing fault diagnosis method based on unsupervised domain adaptation with popular embeddings is proposed to address the dual challenges of unsupervised data environments and diverse engineering requirements while meeting demands for model costs, speed, and accuracy. Initially, an optimal convolutional neural network architecture is selected to extract significant feature data. Subsequently, an unsupervised manifold learning approach is utilized for dimensionality reduction, and a heuristic optimization algorithm is employed for hyperparameter optimization, thereby enhancing the model’s performance and generalization capability. Furthermore, a robust classifier developed from source domain data and an accuracy calculation method based on decision fusion are designed to significantly improve the model’s robustness. Finally, experiments on datasets with varying noise levels demonstrate that the proposed model achieves up to 20% higher accuracy in bearing fault diagnosis compared to other monitoring methods, showcasing its excellent practical utility in bearing anomaly detection and fault diagnosis.