Domain-adaptive gearbox fault diagnosis using deep residual sparse autoencoder with symmetric dot pattern encoding

Abstract

The intelligent operation and maintenance of gear systems face challenges like diverse and dynamic working conditions, which may cause distribution discrepancies in the feature space, thereby significantly reducing the effectiveness of traditional diagnostic models. In most multi-source domain adaptation methods, the need for consistent state spaces and the loss of valuable information with one-dimensional data limit the models’ generalization ability. To address these limitations, the present study proposes a deep residual sparse autoencoder (DRSAE) for cross-domain fault diagnosis, which converts vibration data into symmetric dot pattern images to retain more effective information, and trains the model using a two-stage sequential training strategy. In the unsupervised pre-training stage, residual modules and sparsity constraints are added to remove redundant information, and the Hilbert–Schmidt independence criterion is employed to minimize the dependency between features and domain-specific distributions, extracting domain-invariant features with maximum independence. In the supervised fine-tuning stage, the model is fine-tuned using partial source domain labeled information. Validation on the wind turbine (WT) and Huazhong University of Science and Technology (HUST) gearbox datasets shows that the DRSAE achieves average cross-domain diagnostic accuracy of 93.74% and 95.67%, respectively, outperforming other methods. The framework requires fewer labeled samples, enhancing its applicability for real-world gearbox monitoring under varying working conditions.