Secure Aggregation-Based Federated Learning for Fault Detection in Distributed Manufacturing Systems

The growing use of the Industrial Internet of Things (IIoT) technologies in current manufacturing has provided an opportunity to monitor the work of the machines and the processes in real time, however, it also creates significant concerns about data privacy, communication overhead, and compliance with regulations. Conventional centralized fault detection models need that massive amounts of sensitive process information be collected, which is not always feasible in distributed industry scenarios. In order to cope with them, the given paper suggests the use of a Secure Aggregation-Based Federated Learning (SecAgg-FL) architecture to detect faults in distributed manufacturing systems. The suggested solution enables every manufacturing location to be trained only with its data and only exchange masked parameter changes with a central server so that no raw data or specific gradient information is visible. Strong privacy guarantees are offered by using a cryptographically secure aggregation protocol and the impact of non-IID data distributions and client dropouts are addressed through an adaptive federated training strategy. The experimental findings on three benchmark datasets (TEP, SECOM, and MIMII) indicate that SecAgg-FL with centralized models has similar fault detection performance, as indicated by the AUC values of over 0.96 and low overhead. The analysis of communication further indicates that the use of compression technique has greatly cut on the bandwidth used thus making the framework scalable and efficient. This piece of work illustrates the possibility of implementing privacy-sensitive, real-time fault detection in Industry 4.0 settings without negatively affecting the accuracy and reliability of such systems.