Trustworthy Machine Learning Operations for Predictive Maintenance Solutions

Abstract

With the ever-growing capabilities of data acquisition and computational units in industry, development, and deployment of data-driven models (e.g., predictive maintenance solutions) have become more abundant. However, when not trained and maintained properly, these models can be counterproductive as their predictions are not correct, reliable, or interpretable. In addition, unlike conventional software, the issues with such models manifest themselves in reduced productivity and not in forms of traceable software error. Therefore, in this proposal we aim to use model evaluation measures introduced in trustworthy AI operations (TrustAIOps) to trigger re-evaluation of different parts of the data pipeline and the deployed data-driven model given machine learning operations (MLOps) requirements. We argue that by creating an ecosystem capable of monitoring different aspects of a data-driven solution by integrating and managing the implementation concepts in TrustAIOps and MLOps, it is possible to boost the performance of models given the constant changes induced by the specifications of Industry 4.0.