An Explainable Deep Learning Algorithm for Damage Detection and Classification of Airship Envelopes

Damage data from a specially designed airship envelope test bench were utilized, and a novel explainable deep learning method called Neuron-Focused Class Activation Mapping (NF-CAM) was developed to classify and interpret damage types. Post-training interpretation algorithms were applied to existing classification neural networks to enhance understanding without compromising performance. Feature weights were refined by calculating logit values between similar features, thereby improving focus on target regions. Feature maps were sorted and accumulated to reduce the impact of edges and irrelevant areas, effectively highlighting key feature regions. NF-CAM was employed to re-weight feature maps, while occlusion saliency maps were used to validate the reliability of the emphasized regions. The proposed method demonstrated high accuracy in classifying various damage types. It showed a more significant decrease in accuracy during NF-CAM occlusion experiments compared to mainstream methods, highlighting its superior explanatory capabilities.