Intelligent Diagnosis of Substation Equipment Based on Federated Learning and Inspection Robot Image Data

Accurate assessment of substation equipment conditions is crucial for ensuring the reliable operation of power systems. Traditional methods rely on centralised training, which requires extensive labelled data often owned by different electricity companies. Driven by the need for privacy protection and competitive advantage, such data cannot be shared, creating significant barriers to effective model development. To address these challenges, this paper proposes a federated learning‐based Transformer model (FL‐TM). The Transformer network is utilised to extract global features from equipment images captured by inspection robots, while a multi‐layer perceptron (MLP) classifier performs classification tasks. To minimise communication overhead, only the MLP is updated locally, while the Transformer network remains frozen during training. The proposed FL‐TM framework incorporates a knowledge distillation‐based mechanism, where soft labels generated by edge nodes are aggregated at the cloud server to optimise the global model while preserving data privacy. Experimental results show that, in distributed and privacy‐sensitive environments, the proposed method enhances recognition accuracy and achieves lower communication costs.