Fault diagnosis method for harmonic reducer based on personalized federated aggregation strategy with skip cycle weight

Abstract

In order to solve the problem of low fault diagnosis accuracy caused by the difference in data distribution between users of different harmonic reducers under data islands, a privacy-preserving fault diagnosis method for harmonic reducers based on personalized federated learning (PFL-HR) is proposed. First, a model construction method based on second aggregation is proposed to deploy personalized local models among users, reducing differences in data distribution. Second, a federated aggregation strategy based on cycle weight is proposed to update the global model parameters, accelerating the convergence of the global model. Finally, in the global model parameters distribution stage, a model parameters’ skip aggregation strategy is proposed to extend the training paradigm, further improving diagnosis accuracy. Through multiple groups of experiments on the harmonic reducer data collected from the self-built experimental platform, the results show that PFL-HR improves accuracy by an average of 6.08%. compared to other personalized federated learning methods.