A bearing fault diagnosis method across operating conditions based on wavelet packet decomposition and impulse spatiotemporal modeling
Abstract
To address distribution shift and degraded generalization in rolling bearing fault diagnosis under varying operating conditions, such as fluctuating speeds and loads, this paper proposes a hybrid diagnostic framework termed wavelet packet decomposition-one-dimensional convolutional neural network-spiking neural network (WPD-1DCNN-SNN). The proposed framework aims to achieve robust fault identification in scenarios where target-domain samples are unavailable during training. Specifically, wavelet packet decomposition (WPD) is first used to transform raw vibration signals into a multi-scale frequency-band representation, thereby capturing fault-related local impulse characteristics from non-stationary sequences. A one-dimensional convolutional neural network (1DCNN) is then employed to extract discriminative local features, followed by a spiking neural network (SNN) with Leaky Integrate-and-Fire neurons to perform event-driven spatiotemporal modeling within discrete time steps. The proposed method is validated on 12 cross-condition diagnostic tasks constructed from the Case Western Reserve University (CWRU) dataset and a self-constructed bearing fault dataset. Experimental results show that the proposed framework achieves average diagnostic accuracies of 98.35% and 96.53% on the CWRU and self-constructed datasets, respectively, with an overall average accuracy of 97.44% across all tasks. Compared with representative transfer learning and domain adaptation baselines, the proposed method achieves superior accuracy as well as consistently high Precision, Recall, and Macro − F1 values. Further analysis based on confusion matrices and t-distributed stochastic neighbor embedding visualization confirms that the proposed framework effectively leverages structural inductive biases to suppress the influence of operating-condition-induced amplitude distribution drift. This study provides an effective solution for cross-condition bearing fault diagnosis without requiring explicit target-domain alignment during training, demonstrating strong stability and generalization capability under complex operating environments.