Large Model for Rotating Machine Fault Diagnosis Based on a Dense Connection Network With Depthwise Separable Convolution

Most of the existing intelligent fault diagnosis models are suitable for only a type of rotating machine or equipment. To achieve the intelligent fault diagnosis for various rotating machines, it is significant to construct a diagnostic model with a powerful generalization ability. Thereupon, this work explores a large fault diagnosis model for a variety of rotary machines. To process the big data from a number of rotating machines and mine their fault characteristics effectively, a dense connection network with depthwise separable convolution (DCNDSC) is proposed as the large model. In this network, a dense connection with depthwise separable convolution block (DCDSCB) is designed for representing the complex vibration data and suppressing the over-fitting, and then a series of DCDSCBs are stacked so that DCNDSC can well extract various complicated characteristics caused by different faults and working conditions. A large rotating machine dataset including almost all public rotating machine data and our private data are built to train the large model. For enhancing the diagnostic ability of large model on the new monitoring data, a diminutive network fine-tuning strategy is proposed, while the main feature extraction capability of the pretrained DCNDSC is preserved. Ten fault datasets are applied to verify the high accuracy and strong generalization ability of the developed large model. This model is not only effectively applied to the fault diagnosis of actual rotating machinery but also first provides a pretraining large model for the field of mechanical fault diagnosis. Codes of our work are released at: https://qinyi-team.github.io/2024/04/Dense-connection-network-with-depthwise-separable-convolution/.