A Novel Federated Learning Framework for System Identification

In this paper, we integrate the federated learning approach into the System Identification (SYSID) problem and propose the Incremental Clustering-based federated learning method for SYSID (IC-SYSID). IC-SYSID provides a framework capable of managing the SYSID problem of multiple systems while does not depend on prior knowledge of the SYSID dataset. In IC-SYSID, the necessity of prior knowledge is removed by ClusterCraft which is an incremental clustering method. ClusterCraft initializes one cluster model at the beginning and conditionally generates new clusters during the training. It determines the similarities between local workers and groups of similar ones in the newly generated clusters. As ClusterCraft can generate similar clusters, we develop ClusterMerge to merge similar clusters thus diminishing the number of clusters. Furthermore, we suggested a scaled Glorot initialization to initialize stable cluster models and incorporated a regularizer into the loss function of each local worker to prevent instability during local training. To handle large-scale SYSID data, we also train local models with a mini-batch deep learning structure. We conduct experiments with a synthetic SYSID dataset and demonstrate that IC-SYSID gives sufficient SYSID performance while handling SYSID challenges that emerge from multiple data sources.