Enhancing the Aggregation of the Federated Learning for the Industrial Cyber Physical Systems

Industrial Cyber-Physical Systems (ICPS) have gained considerable interest in the last decade from both industry and academia. Such systems have proven particularly complex and provide considerable challenges to master their design and ensure their functionalities. In this paper, we intend to tackle some of these challenges related to the performance and decision supports of ICPS by proposing a federated learning based framework, called FedGA-ICPS. First, we initiate a ICPS modeling formalism to specify such systems structure and behaviors. Then, based on the ICPS generated data from the industrial sensors, FedGA-ICPS analyzes their performance by proposing locally embedded learning models. Then, federated learning is powered by genetic algorithm to accelerate and improve the aggregation. Finally, transfer learning is applied to broadcast the performed parameters of the leaning models over different constrained entities. FedGA-ICPS has been applied on MNIST and showed prominent results.