Assessment of the Trustworthiness of Grey-Box Models for Condition Monitoring of Industrial Components and Systems

In this paper we propose a method to assess the trustworthiness of Grey-Box (GB) models for Condition Monitoring (CM) of industrial components and systems. We consider a GB architecture that leverages the strengths of physics-based White Box (WB) and data-driven Black Box (BB) models by using the BB to correct the WB prediction. Specifically, the BB receives in input the measured signals and the output of the WB. We define trustworthiness in relation to model accuracy and consistency with the laws of physics. The latter is evaluated using Shapley Additive exPlanations (SHAP) to quantify the extent of reliance of the BB on the WB output. The rationale is that if the BB is sensitive to the WB output, the GB is trustworthy because it indirectly embeds the laws of physics. The effectiveness of the proposed method is demonstrated on a synthetic case study which mimics the condition monitoring of an industrial system.

Keywords: Condition monitoring, Neural network, Grey-box model, Explainable artificial intelligence, Shapley additive explanations.