Attention-Enhanced CNN–LSTM Integration for Rapid Detection of Aviation DC Series Arc Faults

Abstract

More-electric aircraft increasingly use high-voltage direct-current (HVDC) distribution, where low pressure, vibration, and variable loads complicate DC series arc-fault detection. This article proposes a mechanism-guided convolutional neural network–LSTM attention model using four current descriptors: standard deviation, current signal power, mean frequency, and power bandwidth. Under an eventwise train-test split, the model achieves 99.62% accuracy and a 99.55% F1-score. Held-out temperature-humidity and resistive-load tests evaluate performance outside the training conditions. Shapley additive explanations and attention visualization indicate that the model uses spectral and current-power descriptors within the tested distribution. STM32F767 deployment requires 5.69 KiB RAM and achieves 4.007 ms measured inference latency. The results support real-time edge inference for aviation HVDC branch protection under the tested conditions.