Desirable Properties Based Neural Network Explanations Evaluation Method for Fault Diagnosis

Neural networks have been widely used in fault diagnosis due to their excellent performance, while the ‘black box’ nature of neural networks remains a question whether the results of fault diagnosis are trustworthy. Many explanation methods have been used to solve this problem. However, it is hard to decide which explanation method to trust for fault diagnosis if explanations generated by different methods are contradictory. Thus, in this paper, a desirable properties based evaluation method is proposed to evaluate neural network explanations for fault diagnosis. Firstly, the desirable properties that explanations for fault diagnosis ought to satisfy are summarized, which could be used as evaluation criterion. Then the input features that have relationships between input and output of the neural network are generated, which could be used to quantify desirable properties. Finally, desirable properties are quantified using these input features to evaluate explanations. Experiments are conducted to verify the proposed method and the prior knowledge is used to verify the effective of the proposed method, and the results show that our method can be used to evaluate neural network explanations for fault diagnosis.