Multi-Frequency Wavelet-Attention Network for Robust Fault Diagnosis via Multi-Source Heterogeneous Time Series Fusion Under Class Imbalance

Owing to the structural complexity of industrial systems, equipment fault diagnosis often requires the integration of multi-source time series data. However, traditional diagnostic methods struggle to effectively process and fuse heterogeneous signals. In this paper, a novel Multi-Frequency Wavelet-Attention Network (MWANet) is proposed to effectively extract and fuse features from both high-frequency and low-frequency signals for enhanced fault classification. The high-frequency branch applies Morlet wavelet transform to generate time-frequency representations, followed by wavelet convolution layers to enlarge the receptive field and capture transient fault characteristics. The low-frequency branch combines 1D convolution with self-attention mechanisms to extract stable temporal features and capture global dependencies. Feature fusion is achieved through an Efficient Channel Attention (ECA) module, which adaptively emphasizes complementary information. To address class imbalance issues commonly found in real-world datasets, a cross-entropy loss function regularized by Kullback-Leibler (KL) divergence towards prior distributions is designed, improving minority class recognition. Experimental results on a real machine tool dataset demonstrate that MWANet outperforms several benchmark models in accuracy and robustness.