A Lightweight and Explainable Data-Driven Scheme for Fault Detection of Aerospace Sensors

Compared with traditional model-based fault detection and classification (FDC) methods, deep neural networks (DNNs) prove to be more accurate for aerospace sensors. An emerging approach, called image-based intelligent FDC, converts sensor data into images, treating FDC as abnormal region detection problem on the image. Although promising advances have been claimed, due to the small size of the stacked image, diminutive convolutional kernels and shallow DNN layers were used, which hinders the FDC performances. In this paper, we propose a data augmentation technique to enlarge the image size (corresponding to VGG16 net in the machine vision realm), followed by fine-tuning the FDC neural network using VGG16 as a baseline. To compress the network, a layered channel pruning algorithm evaluates the importance of the convolutional filters enclosed in the net. Channel pruning with a knowledge distillation algorithm is then implemented, yielding a more lightweight net. The fusion of Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) methods are adopted for explainability analysis of the FDC net, which clarifies the enclosed mathematical operations both locally (e.g., convolutions) and globally (e.g., deep Shapley values). Through data augmentation, fine-tuning, and pruning, the FDC net achieves 98.99% accuracy across 4 aircraft at 5 diverse flight conditions. We also deploy the FDC net on real data from the flight of a model airplane, which claims an accuracy at 98.44%. The high accuracy and explainability analysis presented in this paper justifies the efficacy of the proposed scheme.