PPFGED: Federated learning for graphic element detection with privacy preservation in multi-source substation drawings

Abstract

In digital recognition of substation drawings, a significant challenge is how to accurately detect complex graphic elements in each drawing. Existing detection schemes often operate in isolation, with data dispersed across multiple sources, which impedes the development of high-accuracy models. A straightforward solution of consolidating drawings from different substations for training does not satisfy the stringent security requirements of substations. To address these challenges, we propose a novel graphic element detection framework based on federated learning for multi-source substation drawings. Our framework involves a server provider, multiple power companies, and distributed substations, and it employs a universal model generation scheme that protects data privacy. Local substations utilize the designed universal graphic element extraction scheme to categorize drawings and extract smaller elements, achieving enhanced accuracy. Subsequently, we design a privacy-enhanced diffusion model for substations to protect elements before uploading them. Power companies then train local model and apply a designed dimension-compressed local differential privacy scheme to the model parameters to protect the training process. To deter server provider from maliciously exploiting model, a game theory-based optimization solution is designed between server provider and power companies. Finally, experiments conducted using substation drawings provided by multiple cooperating units demonstrate that our framework can accurately perform drawing classification and graphic element extraction, and while obtaining the strongest privacy guarantees and low time cost, it achieves an intelligent system for graphic element detection with an accuracy of 88.14% across different regional substations.