Real-time fault diagnosis of subsonic aircraft based on lightweight convolutional neural network

This paper presents a real-time fault diagnosis framework for subsonic aircraft on resource-constrained platforms. Initially, a baseline convolutional neural network (CNN) is designed and validated, demonstrating high diagnostic accuracy. To mitigate the computational demands of this baseline, we propose a lightweight CNN that integrates depthwise separable convolutions and a Squeeze-and-Excitation attention mechanism to significantly reduce computational complexity. Experimental results indicate that the lightweight model achieves a substantial reduction in computational load and memory usage while preserving diagnostic accuracy comparable to the baseline. Validation on an embedded hardware platform confirms the practical feasibility of the proposed approach for real-time applications.