Robustness and Explainability of Machine-Learning-Based Medium Voltage Circuit Breaker Condition Assessment Method
This article explores the robustness and explainability of a convolutional neural network-based fault detection method for medium-voltage circuit breakers. The robustness is analysed by evaluating the method's performance under the presence of stationary and non-stationary disturbances in the vibration signature. Additionally, the impact of sensor ageing on performance indices is investigated to assess long-term reliability. Since the condition assessment method is focused on binary classification, the detection outcome interpretation aspect is addressed by providing recommendations for operator or autonomous system actions. Both aspects are demonstrated using datasets collected from real-world medium-voltage circuit breakers.