Real-time PEM fuel cell fault classification in smart manufacturing plant by using uncertainty aware federated deep learning

The Proton Exchange Membrane Fuel Cells (PEMFCs) is not the least significant component of the modern industrial, yet the reliable operation of the smart manufacturing environment cannot be guaranteed since the breakdowns of PEMFCs are multiple and highly complex in character. The article describes the UFDL-Edge, an uncertainty-based federated deep learning model that was created to detect the defects of PEMFC in real-time through an IIoT-based manufacturing plant. The objective of the proposed architecture is to offer the adaptive feature detector and an open-ended model evolution without compromising the privacy of the data in the distributed manufacturing locales through integrating uncertainty-aware feature detectors (UFD) with temporal Hadoop Distributed File System (HDFS)-based streaming. It employs a hierarchical and federated learning approach to obtain a combination of the advantages of Byzantine-resistant aggregation and uncertainty quantification to enhance the resilience of fault-detection in a variety of operating conditions. The 10.3% improvement in F1-score on the initial anomaly detection, 23.7% reduction in false positive ratio, and less than 100ms latency of real-time diagnostics obtained with huge-scale experimentation in three testbeds of smart manufacturing are better results. Foot printing framework eliminates drastic challenges such as cold start issue, water flooding, thermal controller malfunction and catalyst degradation as well as data sovereignty across manufacturing plants used.