A Multi-Sensor Time-Frequency Fusion Fault Diagnosis Framework Based on Diffusion Models for Gearboxes under Imbalanced Conditions

With the increasing demand for intelligent diagnosis in modern industry, multi-sensor collaborative fault diagnosis technology has been widely developed. However, these studies focus on balanced datasets, while industrial practice often encounters data environments with far fewer fault samples than normal samples. To address this problem, this paper proposes a multi-sensor collaborative fault diagnosis framework under imbalanced conditions. Specifically, the improved DiffWave is applied in the denoising diffusion probabilistic model to obtain a high-quality multi-source augmented hybrid dataset. Then, a multi-source multi-scale time-frequency fusion convolutional neural network (MMT-FCNN) is proposed for multi-sensor collaborative fault diagnosis. Multiscale gated convolution in MMTFCNN employs a gating mechanism and several parallel large-scale depthwise convolutions to extract diverse features in the time and frequency domain branches. The efficient adaptive fusion of MMTFCNN introduces cross-space learning to improve the global-local information interaction in attention feature fusion and introduces efficient grouping to achieve efficient multi-branch data fusion. The experimental results based on the publicly available and constructed gear datasets show that the proposed method can achieve high diagnostic accuracy and stability under extreme imbalance conditions.