Learnable Wavelet Transform and Domain Adversarial Learning for Enhanced Bearing Fault Diagnosis

Unsupervised domain adaptation techniques have been widely used to detect the health conditions of rolling bearings. Despite the importance of cross-domain fault diagnosis, it has not received much attention for applications in noisy environments. To address this issue, we propose a novel architecture that combines learnable wavelet packet transform with domain adversarial neural networks (DANN-LWPT). The proposed method involves utilizing the learnable wavelet packet transform (LWPT) and wavelet packet transform (WPT) to decompose and reconstruct signals from the source and target domains. These reconstructed signals are then fed into a domain adversarial neural network (DANN). We introduce a guidance loss that dynamically enforces similarity between the source and target domain signals in the time-frequency domain during the process of decomposition and reconstruction, promoting the learning of domain-invariant and discriminative features. We compare our proposed method with other representative domain adaptation approaches, and the results of the evaluation show its superiority.

Keywords: Rolling bearing, Fault diagnosis, Unsupervised domain adaptation, Learnable wavelet packet transform.

Download PDF