Intelligent Fault Diagnosis with Both Health Condition Label Scarsity and Rotating Speed Variations

Intelligent fault diagnosis (IFD) can significantly reduce the workload of operators, and has important research value and engineering application significance. Existing intelligent diagnostic algorithms still require a large amount of labeled data for training. Self-supervised learning utilizes unlabeled data to automatically learn its intrinsic structure, which can significantly reduce the demand for labeled data in IFD. Therefore, this article conducts research on IFD algorithms that effectively utilize unlabeled monitoring data based on self-supervised learning. Firstly, a self-supervised learning method based on vibration signal time-frequency graph is constructed, and two new data augmentation methods are introduced to assist in self-supervised network training, learning valuable representations for downstream diagnosis tasks. Afterwards, based on the self-supervised learning network, global fine-tuning is performed with a small amount of labeled data to achieve the construction of a diagnostic network. Finally, validations were conducted based on the bearing fault dataset of the multi-stage transmission experimental platform. Extensive experiments have shown that the self-supervised learning based network can hold performance with only a small amount of labeled fault data. The research content of this article will provide some reference for the effective utilization of unlabeled monitoring data in real-life applications.