A Lightweight-Model Based Federated Learning Approach for Fault Diagnosis in Intelligent Power Distribution Systems.

The federated learning framework plays a pivotal role in diagnosing faults in distribution equipment. Alleviating the computational complexity and communication burden of its micro-terminals is essential for enhancing the efficiency of the federated learning framework. Therefore, this paper focuses on effectively lightening the fault diagnosis model by employing a combination of model pruning and quantization techniques, aiming to compress the fault diagnosis model. To achieve this, we employ a two-step approach. Firstly, neural networks are pruned selectively based on the significance of the model's structure and weights, resulting in model compression. Secondly, a significant portion of the model parameters is mapped to a smaller set of representative vectors, leading to data compression, reduced storage requirements, and lowered data transmission costs. Compared with traditional methods, this approach reduces the computational load of the fault diagnosis model from 901.38M to 442.43M, with only a minor 2.5% decrease in diagnostic accuracy. This methodology not only ensures the security of the data but also adeptly addresses the constraints faced by micro-terminal devices in computing and communication within the distributed federated learning environment.