Feasibility study of machine learning capabilities integration into the microcontroller of the new generation of ABB circuit breakers

With the growth of computational power capabilities of electronic embedded devices, an increasing number of tasks are moving from centralized processing units to edge devices, implementing edge computing paradigms. This thesis, developed as a collaboration between Politecnico di Torino and ABB, is about the adoption of this kind of distributed architecture in ABB circuit breaker systems, aiming to evaluate the potential integration of Machine Learning capabilities directly into the microcontroller of the circuit breaker mainboard. Specifically, the treatment starts with a description of the development of a Load Forecasting Machine Learning algorithm exploiting the TensorFlow framework, discussing in detail the dataset collection, training and testing processes. After that, a preliminary benchmarking evaluation has proven that the resulting ML model is too complex to be executed on resource-constrained embedded devices. Therefore, the model is converted through the TensorFlow Lite tool, which produces a simplified lightweight version to be wrapped into a TinyML application. TinyML techniques, together with the TensorFlow Lite for Microcontroller tools, made it possible to run ML inference on tiny devices obtaining interesting results in terms of required CPU time, memory consumption and extra software. The thesis continues reporting this complete performance evaluation that is carried on an ST Microelectronics demo board that emulates the circuit breaker mainboard since it is equipped with the same microcontroller. The outcomes of this evaluation are promising and the company is planning to continue the project development. Lastly, the final section deals with a discussion about the limitations and potential improvements of the project along with an illustration of the future of edge computing systems.