Knowledge-based Clustering Federated Learning for fault diagnosis in robotic assembly

Fault diagnosis in industrial robots is a critical aspect of intelligent manufacturing. However, the accuracy of fault diagnosis models can be significantly affected by the few-shot problem, which refers to the limited availability of labeled data for training. Traditional methods often rely on combining data from different sources, which can raise privacy concerns due to the sensitive nature of the data involved. Federated Learning has emerged as a privacy-preserving approach, but it faces challenges in dealing with the Non-IID (Non-Independent and Identically Distributed) data distribution across different robot systems. In this study, we propose a novel approach called Knowledge-Based Clustering Federated Learning (KCFed) to address both the few-shot problem in robot fault diagnosis and the Non-IID problem in Federated Learning. KCFed incorporates Federated Learning principles to ensure privacy protection during the model training process. Additionally, by leveraging the Knowledge-based Clustering Mechanism (KCM) and Knowledge Accumulation Mechanism (KAM), KCFed aims to improve the performance of fault diagnosis models by clustering similar tasks and accumulating useful prior information. To evaluate the effectiveness of KCFed, we conducted experiments using data collected from industrial robots performing 12 different tasks, resulting in a diverse set of 48 states. The experimental results demonstrate the promising performance of KCFed in improving fault diagnosis accuracy while preserving privacy in a federated learning setting.