DeceFL to detect bearing faults from CWRU benchmark dataset. (A) Data were divided into IID samples for all 4 clients. (B) Data were divided into Non-IID unbalanced samples. Each client locally specified its data size and sample distribution. (C) Illustration of communication topology for FedAvg, SL and DeceFL. (D), (E) Performance of DeceFL on IID data using logistic regression and DNN, respectively, with reference performance of FedAvg and SL. (F), (G) Performance of DeceFL on Non-IID data using logistic regression and DNN, respectively, with reference performance of FedAvg and SL. (D)–(G) The boxplots at bottom illustrate the performance comparison between DeceFL and each client that was trained independently (that is, each client trained its own model only using its associated local data without communicating with any other clients). The experiments on CWRU benchmark dataset confirms the similar performance of DeceFL, as a fully decentralized framework, in comparison to FedAvg and SL, for industrial datasets, while DeceFL offers more freedom on the choice of communication topology to handle practical issues in organizing clients for large-scale federated learning applications.