Fault diagnosis method based on recursive federated transfer learning under multi rate sampling

The number of samples and the use of sample information affect the accuracy of fault diagnosis model based on deep learning. In the actual industrial equipment health monitoring, there are often multi rate sampling sensors, which makes the number of samples with complete structure based on deep learning training very few. Only using the samples with complete structure for fault diagnosis will lead to the waste of incomplete structure sample information. In this paper, a recursive federation strategy is proposed for fault diagnosis. The obtained sensor observations at different rates are regarded as sample sets of different institutions. At a single time, the local model recursively federates according to the variable relationship between institutions. The federal central model is jointly optimized with the local models of each institution to obtain a global fault diagnosis model with high accuracy at that time, then the multiple time models are recursively federated according to the previous sample time and the next sample time to obtain a more optimized fault diagnosis model. The effectiveness of this method is proved by the case analysis of gearbox fault diagnosis.