TinyML-based approach for Remaining Useful Life Prediction of Turbofan Engines
Georgios Athanasakis, Gabriel Filios, Ioannis Katsidimas, Sotiris Nikoletseas, Stefanos H. Panagiotou
Abstract
In the recent years, artificial intelligence, machine learning and IoT technologies have enabled a great number of industrial applications with profitable results. Predicting the remaining useful life (RUL) of turbofan engines constitutes a successful example of industrial AI, and it has received thorough attention from the researchers worldwide, with numerous novel and effective methods being proposed in the literature. Meanwhile, TinyML is a recent trend that has emerged in the AI field and demonstrates, amongst others, promising potential to break through the existing barriers of trusting and deploying real-time critical industrial AI solutions. In this context, this paper aims to further contribute to the literature and demonstrate the realization of RUL predictions in the extreme edge via TinyML, using the popular C-MAPSS dataset from NASA Ames Research Center, X-CUBE-AI tool and STMF767ZI microcontroller for the deployment of ML models. We benchmark different ML algorithms, with a special focus on deep learning algorithms (LSTMs and CNNs). The results indicate that there is potential for deploying machine learning models for RUL prediction in resource-scarce IoT devices, with acceptable accuracy loss, while taking advantage of the benefits TinyML has to offer over cloud-based AI inference.