Enhancing Edge-Based Federated Learning With Privacy-Preserving Gradient Transmission for Tool Wear Detection

This article focuses on the development of a personalized federated learning (PFL) edge training platform that ensures privacy during gradient transmission (both uploading and downloading). We set up a client-side training environment for federated learning on a Raspberry Pi 4 and use the file transfer protocol (FTP) for uploading training gradients to a server-side PC for aggregation. The personalized gradients are then sent back to the local clients via FTP for model updates. To validate the feasibility of the PFL algorithm, we employ a 1-D convolutional neural network (CNN) for detecting tool wear. We present the comparison results between our proposed PFL and the federated averaging (FedAvg) algorithms to demonstrate performance enhancements. The PFL model is more closely aligned with local client needs, leading to better predictive performance, especially in scenarios of data scarcity and heterogeneous data. Experimental findings reveal that PFL can effectively manage tool wear across various notches or machines. Through the sharing of local model gradients via federated learning, we facilitate training and achieve personalization using a personalized detector. Both the personalized and global models show a 10%–15% increase in accuracy, supporting preventive maintenance efforts.