Multi-scale and multi-branch 1D-AM-CNN network for fault diagnosis based on Bayesian data fusion

To overcome information loss and difficulty in fusion of heterogeneous signals due to forced temporal transimaging in existing convolutional neural network CNN bearing fault diagnosis methods, this paper proposes a multi-scale one-dimensional attention convolutional network (1D-AM-CNN) based on Bayesian optimization. The method synchronously extracts fault cross-frequency band features through a multi-scale parallel architecture, uses Bayesian optimization to achieve adaptive optimal weighted fusion of current and vibration signals, and employs a stylized recalibration module with a coordinate attention mechanism to perform channel-space bi-dimensional feature recalibration. Experiments on the PU 6203 bearing dataset show that the proposed method achieves an accuracy of 99.09% for six types of fault identification with an AUC of 0.991 and an F1 score of over 99.11%, and a recall rate of 98.85% for early weak faults (damage area ≤ 2%), which demonstrates the effectiveness and practical applicability of the proposed framework for intelligent fault diagnosis.