Learnable wavelet-driven physically interpretable networks for bearing fault diagnosis under variable speed

Abstract

Variable speed conditions pose great challenges for intelligent bearing fault diagnosis. The popular intelligent fault diagnosis models neglect the effect of rotational speed when extracting features, meantime the extracted features lack the physical meaning. To this end, a learnable wavelet-driven physically interpretable (LWPI) network is proposed for diagnosing bearing faults under variable speed. Firstly, an entropy-based local peak search (LPS) algorithm with an adaptive instantaneous frequency (IF) search range is designed to extract the rotational frequency ridge from vibration signals, and it can effectively suppress the noise and adjacent interference components. Based on the extracted rotational speed information, a learnable multi-wavelet filter layer is constructed to guide the model in adaptively mining features with physical meaning. Next, a convolution block is constructed to mine the high-dimensional features, followed by a linear dense layer designed for fault classification. Experiments on bearing fault diagnosis under variable speed conditions demonstrate that the proposed LWPI network consistently outperforms five advanced methods. Meanwhile, the efficacy of the entropy-based LPS algorithm and learnable multi-wavelet filter layer are respectively verified by the ablation experiments.