Collaborative Drift Compensation

Federated learning (FL) enables multiple clients with limited data to participate in collaborative learning to build a global shared learning model. It captures the knowledge from all the client-level models while keeping their data at local premises and private. While the data distribution homogeneity and stationarity across the clients are taken for granted in FL, this is rarely the case in industrial automation applications. The client-level data distribution shifts lead to a divergence in the learning of the global model and a drift in the model performance. Such a drift is common in industrial settings, where data patterns change over time based on operating and environmental conditions. Existing approaches have proposed ways to adapt drift in a central setting or during FL by assuming the same type of drift across the clients, where the clients may observe different drifts over time and space. Hence, we propose a novel collaborative model adaptation algorithm (\(\texttt{CLMA}\)), which uses only model parameters and does not need any additional information or data from the client to be communicated to the server. Our approach works for various types of drift found in the industries and addresses the shortcomings of FL for model adaptation in both scenarios: while doing FL and also after model deployment for inference. Through numerical experiments, we evaluate our algorithm on a real industrial dataset, demonstrating its superior efficacy in handling various types of drift and non-IID data in an FL setting.