An Efficient Federated Learning Framework for Enhancing Data Diversity and Communication in Industrial IoT Fault Diagnosis

Abstract

In the field of smart manufacturing, the widespread application of Industrial Internet of Things (IIoT) has prompted various data silos to generate and accumulate huge amounts of industrial data. To ensure the security of multiparty data, the industry has introduced federated learning (FL) technology to realize the circulation of data value by training machine learning models across silos. However, the nonindependent and identically distributed (Non-IID) nature of data distribution among different silos in real industrial environments, which results in degradation of model performance and increase in communication overheads, has become a key challenge that needs to be addressed. In this article, we propose a sparse quantization compression solution based on FL multiagent generative adversarial network (MAD-GAN) named SQC-FLMGAN, aiming to alleviate the problem of Non-IID data across silos in industrial environments and to improve communication efficiency. The scheme generates data samples with diversity and high quality for each client by training MAD-GAN models in an FL environment. This process effectively mitigates the modal collapse problem and improves the generalization ability and diagnostic accuracy of the model. In order to further reduce communication overhead, we have introduced an advanced model compression technique that combines Top-k sparsification and quantization methods. The Top-k algorithm selects crucial model parameters for transmission, while quantization further reduces their precision, eliminating redundancy in communication reducing the amount of data transmitted. Through experiments on the bearing dataset in industrial scenarios, we have demonstrated that the SQC-FLMGAN scheme significantly reduces communication overhead while maintaining model performance. The code framework is available at https://github.com/sqcflmgan/SQC-FLMGAN.