Conceptual Design and Preliminary Embedded Applicability Evaluation of an AI-Based Predictive Maintenance Residual Current Device Using Time-Series Analysis of Leakage Current

This study proposes a conceptual AI-based predictive maintenance architecture for residual current devices (RCDs) using time-series analysis of leakage current and evaluates the preliminary embedded applicability of deep learning models. Leakage current data were collected at 1 s intervals for 120 h using two ACS712 current sensors. LSTM, GRU, and Transformer models were trained and evaluated using the same dataset configuration and evaluation framework. Their performance was compared using MAE, RMSE, trainable parameter count, estimated FP32 weight memory, and CPU inference time. LSTM and GRU showed lower prediction errors than the Transformer. GRU achieved prediction accuracy comparable to LSTM while requiring the fewest parameters and the shortest CPU inference time, suggesting its potential as a lightweight candidate for future MCU-based implementation. Based on these results, the proposed architecture is designed to support early warning and predictive maintenance while retaining the conventional RCD protection function.