Application of Multisensor Data Fusion and Inception–SE Network in Low-Speed Bearing Fault Diagnosis

The diagnosis of low-speed bearing faults in wind turbines remains challenging due to weak fault signatures and significant noise interference under low-speed operating conditions. To address these limitations, this study proposes MSISFNet, a novel multisensor fusion framework integrating enhanced inception modules, squeeze-and-excitation (SE) networks, self-attention mechanisms, and a joint training strategy for robust fault detection. Specifically, the enhanced inception module employs parallel multiscale convolutional kernels to simultaneously process acoustic emission (AE), effectively capturing sparsely distributed high-frequency transient impacts and low-frequency resonance modulation features inherent to low-speed faults. The SE module dynamically adjusts channel-wise feature responses through adaptive weight learning, suppressing noise-dominated frequency bands while enhancing the sensitivity of key channels associated with fault harmonics. To enhance multisensor collaborative optimization, we implement a phased joint training paradigm consisting of two stages. Initially, sensor-specific subnetworks undergo pretraining to extract distinctive and discriminative features from individual modalities. Subsequently, a self-attention mechanism is used to fuse features from different sensors at the feature level through adaptive weighting, so as to improve the sensitivity to fuzzy subtle fault features in individual modes. Comparative experiments with single-sensor CNN models and other advanced models highlight its superior diagnostic performance in bearing fault detection under low-speed conditions. This article provides insights and reference for future research on bearing fault diagnosis under low-speed conditions by leveraging the integration of information from multiple AE sensors.