Implementation of knowledge distillation for onboard defect detection on an Unmanned Aircraft System for light aircraft general visual inspections

Visual inspections of aircraft are a vital part of routine procedures for maintenance personnel in the aviation industry. However, these inspections take up a considerable amount of time to perform and are susceptible to human error. To mitigate this, utilising image classification for detecting defects is proposed, leveraging transfer learning and knowledge distillation within MATLAB to develop an efficient and deployable model. Transfer learning is applied to a ResNet-50 model, adapting it to classify aircraft defects using a curated dataset. This fine-tuned model is then utilised as a teacher in the knowledge distillation process, where a compact SqueezeNet model (the student) learns from both hard and soft labels to replicate its performance while significantly reducing computational demands. This allows for optimising deep-learning models for deployment on smaller hardware, making the student model suitable for use on an Unmanned Aircraft System (UAS) to filter out images that do not contain a defect, reducing workload for ground personnel. The proposed method offers a solution for improving the efficiency and accuracy of defect detection during a general visual inspection in the aviation industry. Targeted defects here are damaged_skin , missing_or_damaged_rivets , and panel_missing alongside a class denoting no_defect . The knowledge-distilled SqueezeNet model achieves 95.37% validation accuracy and 90.72% inference accuracy, with a 96.9% reduction in model size compared to ResNet-50. The teacher model has a size of 85.77 MB, while the student model is significantly smaller at 2.66 MB, making it ideal for deployment on embedded systems with limited resources.