Knowledge Incorporation for Machine Learning in Condition Monitoring: A Survey

Abstract

Model-based condition monitoring (MBCM) solves the inverse problem of inferring a systems state, including possible faults, from sensor observations. Constructing these models in a knowledge-based manner following the laws of physics is hard due to the inverse nature of the problem and unknown fault types. As a result, it has become more attractive to build a model solely from past observations via machine learning (ML). Although highly promising, shortcomings of ML in the scientific domain, including physically inconsistent results and lack of interpretability, became apparent. This led to recent efforts to enhance machine learning with scientific knowledge including a combination of knowledge-based and data-driven modelling, often referred to as hybrid models. The main contributions of this work are: (1) a link of shortcomings of machine learning in CM to a lack of knowledge; (2) a categorization of unique approaches with respect to required knowledge and mechanism of incorporation that have either been applied in condition monitoring or show potential from their application to scientific problems; (3) derivation of promising research directions uncovered as vacant spaces in the categorization.