Modern engineering systems increasingly face failure modes driven by stochastic physical phenomena that are difficult to detect using conventional signal processing techniques. DC arc faults are a representative example, as their highly variable behaviour, dependence on operating conditions, and transient nature make reliable detection extremely challenging. To address this challenge, we developed a data-driven fault detection framework based on an extensive experimental dataset and machine learning-assisted parameter optimisation, enabling robust identification of arc-fault signatures under realistic operating conditions.

A key challenge was translating algorithms originally developed in MATLAB into a deployable embedded solution. The final implementation runs in real time on a standard STM32 microcontroller without the use of dedicated AI accelerators, requiring extensive model optimisation, memory reduction, and adaptation to hardware constraints while maintaining detection performance. Using real-world DC arc fault applications from both the residential and aerospace sectors, this presentation demonstrates how complex physical phenomena can be transformed into practical, robust, and cost-effective embedded protection systems.