DWAE-LIP with joint-TF loss: a new fault extraction method for rolling bearings under strong noise

Abstract

This paper proposes a new fault extraction method to address the challenge in rolling bearing fault diagnosis, where traditional denoising methods inadequately balance noise suppression and impulsive feature preservation. The method utilizes a discrete wavelet autoencoder (DWAE) model that incorporates local importance pooling (LIP) and a joint time–frequency (TF) optimization loss. First, a DWAE network with a TF decoupling mechanism constructs a mapping between noisy and clean signals in feature space, thereby enhancing its blind denoising capability. Second, an attention-based dynamic importance strategy for the LIP module enhances fault sparsity, suppresses noise, and preserves time-domain impulses. Finally, a multi-domain composite loss joint optimization that combines Pearson correlation coefficient loss for time-domain fidelity and Hilbert spectrum-based loss for frequency-domain feature preservation is used to improve diagnosis interpretability via TF joint supervision. Simulation and experimental results demonstrate that the method maintains strong denoising performance while significantly improving time-domain sparsity and enhancing fault features in the frequency domain and demodulation spectra, offering an effective solution for rotating machinery condition monitoring.