Hybrid feature fusion and CNN-GRU-based intelligent fault diagnosis method for rolling bearings

Abstract

Rolling-element bearings are critical components of rotating machinery, and their operating condition directly affects equipment safety and reliability. Existing rolling-bearing fault classification methods often rely on a single type of feature or manually specified model parameters, which may limit fault-feature representation and classification performance. To address these limitations, this study proposes an intelligent rolling-bearing fault diagnosis method based on hybrid feature fusion and CNN-GRU. First, the CWRU Bearing Dataset was used to construct a multiclass fault classification task, and vibration signals were segmented using non-overlapping windows with a length of 1,024. A random forest was then employed to rank the importance of the raw vibration-signal sampling points and handcrafted statistical features. The 512 most informative raw sampling points, denoted as RawTop512, were fused with the 20 most informative time- and frequency-domain statistical features, denoted as EngineeredTop20, to form a 532-dimensional feature vector. Next, a CNN-GRU classification model was developed, in which the CNN extracted local feature patterns and the GRU captured sequential dependencies among the extracted features. Particle swarm optimization (PSO) was further used to tune the key hyperparameters of CNN-GRU. Comparative experiments with several machine-learning and deep-learning models showed that CNN-GRU achieved the best performance among the baseline models, with an Accuracy of 98.63% and a Macro-F1 of 98.62%. After hyperparameter optimization, PSO-CNN-GRU achieved an Accuracy of 99.05% and a Macro-F1 of 99.06%. These results demonstrate that the proposed hybrid feature representation enhances the discriminability of bearing fault information and that PSO-assisted hyperparameter tuning can further improve CNN-GRU classification performance under the evaluated experimental conditions.