A robust lifetime prediction method for photocouplers considering bias-deviation point for accelerated lifetime testing data

Abstract Photocouplers provide electrical isolation in critical systems, and accurate lifetime prediction is essential for maintaining operational safety. Conventional approaches rely on accelerated lifetime testing (ALT) and predefined degradation models to estimate lifespan under normal conditions. However, these methods often overlook key sources of error: (1) manufacturing uncertainties that cause variability in degradation bias and rate, and (2) shifts in failure mechanisms during testing that distort degradation patterns. This study proposes a novel lifetime prediction framework that systematically addresses both issues. By applying a differential analysis of degradation curves, the proposed method identifies the bias point—where degradation begins due to process or material variation—and the deviation point—where alternative failure mechanisms begin to dominate. Only data between these points are used for prediction, and a Monte Carlo dropout-based deep learning model estimates the lifetime with associated confidence intervals. The predicted lifespan under ALT conditions is subsequently translated into an expected lifetime under standard operating conditions. Experimental validation confirms that the proposed method improves prediction accuracy, facilitating reliable maintenance and risk mitigation for photocoupler-equipped systems. The proposed method estimates a 23 years lifetime, matching the 25-year technical datasheet, while the conventional method overestimates at 58 years, demonstrating superior accuracy and practical relevance.