A Data-Driven Interpretable Method Based on Reversely Built FIS and Its Applications in Measurement

With the rapid development of machine learning, model interpretability has become a major topic in recent studies, especially in the field of measurement. However, there has always been a dilemma in interpreting machine learning models: those intrinsically interpretable methods cannot achieve desirable results in complex applications, whereas the methods with good performances are not interpretable. To resolve this dilemma, this article proposes an interpretable method based on fuzzy inference system (FIS), an intrinsically interpretable method capable of handling complicated applications. First, a method of reversely building FIS is proposed, and each part of the FIS is obtained by driving data, so that the reversely built FIS is enabled to interpret machine learning models. Second, three optimization methods are presented, and the complexities of the antecedents, the consequent, and the rules in the built FIS are greatly reduced. Then, based on fuzzy logic, a multilevel knowledge extracting method is put forward to acquire both the global knowledge and the local knowledge, so that the target machine learning model can be comprehensively interpreted. Finally, the proposed method is evaluated by real applications in industrial fields. The evaluation results show that the proposed method is effective in interpreting machine learning models.