Automating the production of Machine Learning models for imbalanced datasets - application to Industry 4.0

This paper presents a tool automating the production of deep learning models for industrial pervasive applications characterized by imbalanced datasets. Pervasive computing has led to innovative services in Industry 4.0 but remains extremely complex to handle. Indeed, data from IoT devices possess highly specific characteristics that make them challenging to utilize in a learning context. Precisely, these data can be noisy, scarce, unlabeled, unbalanced. In this paper, we focus on the class imbalance problem which is very acute in many industrial applications. The problem statement outlines class imbalance as a hindrance to accurate predictions, especially in scenarios where anomaly detection is vital. The approach described in this paper is the delevery of a tool including best-of-breed techniques and metrics to deal with the class imbalanced problem and providing the best model depending on a set of constraints expressed by the domain experts. It is illustrated on a use of case on motor malfunction detection in Industry 4.0, utilizing a dataset provided by Schneider Electric.