Online monitoring and state-of-health estimation of circuit breakers using data analysis and AI

To analyse these high-dimensional, complex datasets, a tiered AI framework is designed and planned to address sensor multimodality, initial data scarcity, and the distinct temporal requirements of circuit breaker diagnostic. A CNN-YOLO (Convolution Neural Network – You

Only Look Once) architecture performs posterior fault classification using 2D spectrograms, while an Encoder-only Transformer enables real-time, intra-cycle monitoring. Using a selfsupervised masked token strategy, the Transformer learns the electromechanical "distance to completion" from nominal switching cycles, allowing for the detection of millisecond-scale deviations without requiring extensive fault-labeled datasets. The LOESS (Locally Estimated

Scatterplot Smoothing) algorithm decomposes long-term switching data into interpretable mechanical drift trends for state-of-health (SoH) estimation. Early modality-specific "seed" models transit to a unified, early-fusion multimodal Transformer as the field database expands.

The framework ensures immediate diagnostic utility while providing a scalable roadmap for long-term predictive maintenance and switching prognosis.