An Interpretability Method for Broken Wire Detection

As an indispensable piece of production equipment in the industrial field, wire rope is directly related to personnel safety and the normal operation of equipment. Therefore, it is necessary to perform broken wire detection. Deep learning has powerful feature-learning capabilities and is characterized by high accuracy and efficiency, and the YOLOv8 object detection model has been adopted to detect wire breaks in electromagnetic signal images of wire rope, achieving better results. Nevertheless, the black box problem of the model brings a new trust challenge, and it is difficult to determine the correctness of the model's decision and whether it has any potential problems, so an interpretability study needed to be carried out. In this work, a perturbation-based interpretability method-ESTC (Eliminating Splicing and Truncation Compensation)-is proposed, which distinguishes itself from other methods of the same type by targeting the signaling object instead of the ordinary object. ESTC is compared with other model-agnostic interpretable methods, LIME, RISE, and D-RISE, using the same model on the same test set. The results indicate that our proposed method is objectively superior to the others, and the interpretability analysis shows that the model predicts in a way that is consistent with the priori knowledge of the manual rope inspection. This not only increases the credibility of using the object detection model for broken wire detection but also has important implications for the practical application of using object detection model to detect wire breaks.