An embedded physical information network for blade crack detection considering dynamic multi-level credibility

With the widespread application of deep learning across various industrial fields, there has been a growing demand for explainable artificial intelligence (XAI) and reliable decision-making processes. In addressing the issues of lacking inherent interpretability and credibility support for decision-making results within the blade crack detection models of industrial centrifugal fan, an embedded physical information network for blade crack detection considering dynamic multi-level credibility method is proposed in this paper, aiming to enhance the model’s interpretability and improve the balance between the accuracy of blade crack identification and the credibility of the decision-making process. Firstly, the spectral density offset (SDO) indicator is designed to implant the physical loss function space with the monotonicity constraint. Secondly, combined with the simulation a priori information, the gradient-weighted class activation mapping (Grad-CAM) and diversity-pick Shapley (DP-Shapley) methods are adopted to quantitatively assess its credibility from the data level and feature level. Subsequently, the composite credibility is embedded in the training process of the network at the decision-making level, and the accuracy and composite credibility relationship is balanced by adjusting the physical loss term weights. Finally, a dynamic updating method for the weights of the physical loss terms based on the gradient descent is designed, which can maintain a high crack detection accuracy during the model training while gradually improving the credibility of the overall network. Through the data verification of the blade damage test bed for the centrifugal fan, the model has high detection accuracy, stability, and interpretability, and can be applied to the credible detection of the blade crack damage.