In data-driven control systems, machine learning-based controllers are widely adopted, but their performance and safety may degrade when inputs deviate from the training data range. While conventional fail-safe designs rely on anomaly detection followed by shutdown, modern applications demand continuous operation even under unexpected conditions. This study proposes a safety design approach that detects early signs of input deviation based on changes in model outputs and performs model correction or retraining as needed. The goal is to achieve fail-soft operation that ensures safety without interrupting control.
Saito et al. (Wed,) studied this question.