Federated knowledge amalgamation with unbiased semantic attributes under cloud–edge collaboration for heterogeneous fault diagnosis

Abstract

Federated fault diagnosis attracts increasing attention in industrial cloud–edge collaboration scenarios due to the consideration of privacy and security issues. However, statistical heterogeneity caused by non-independent and identically distributed (non-IID) fault data is a ubiquitous problem. Previous methods designed for this problem mainly apply weighted averaging or knowledge distillation to integrate biased local knowledge from clients, ignoring the discrepancy between knowledge extracted from non-IID data solely. This is prone to result in slow and unstable model convergence and global model drift. In this article, a Federated Knowledge Amalgamation (FedKA) framework is proposed to alleviate the knowledge discrepancy caused by statistical heterogeneity. To mitigate the knowledge discrepancy, we first reveal and mine unbiased public knowledge for industrial fault diagnosis, i.e., fault attributes, which describe the fault characterization and mechanism without changing with data distribution. Then, a generative mapping between unbiased fault attributes and fault data is established for clients to incorporate unbiased knowledge into the diagnostic model. Moreover, to mitigate knowledge discrepancy and overcome the global model drift, a knowledge amalgamation algorithm is designed. Through the proposed layer-wise alignment, the multi-level fault semantic knowledge that stems from fault attributes is finely exploited and aggregated. Additionally, unlike the common fixed weighting strategy, we explore an adaptive weighting mechanism through the sample-wise diagnostic confidence to further alleviate the global model drift. Extensive experiments on a real thermal power plant process demonstrate the superiority of the proposed method for federated fault diagnosis.