DBMSCNN: A Dual‐Branch MultiScale CNN Integrating LMSWT and CBAM for Robust Bearing Fault Diagnosis

ABSTRACT Rolling bearing fault diagnosis involves capturing key fault information from vibration signals, including transient temporal features in raw signals and spectral‐spatial patterns in time‐frequency representations. However, it remains challenging to jointly preserve transient temporal information from raw signals and extract concentrated spectral‐spatial patterns from time‐frequency representations. To address this, we propose the dual‐branch multiscale convolutional neural network (DBMSCNN), a dual‐branch framework guided by two synchronized signal representations, namely raw vibration signals and LMSWT‐based time‐frequency representations. Specifically, the 1‐D branch employs multiscale convolutional kernels and gated recurrent units (GRUs) to extract local fault patterns and temporal dependencies from raw sequences. The 2‐D branch uses the local maximum synchrosqueezing wavelet transform (LMSWT) to generate high‐resolution time‐frequency images, followed by an enhanced ConvNeXt V2 backbone and a convolutional block attention module (CBAM) to extract discriminative spectral‐spatial features. The extracted features are fused via feature‐level concatenation before classification to retain information from both branches. Experiments on three benchmark datasets show that DBMSCNN achieves improved diagnostic performance over the evaluated single‐branch baselines and alternative fusion methods.