Arc-FaultNet: A Lightweight Dual-Branch CNN with Channel and Cross-Attention for Generalizable Series Arc Fault Detection

Abstract

Series arc faults in low-voltage electrical installations pose severe fire hazards due to their intermittent and load-dependent nature, making reliable detection across diverse conditions a persistent challenge. This paper presents Arc-FaultNet, a lightweight dual-branch convolutional neural network that jointly exploits temporal and spectral representations of the line current through complementary attention mechanisms. The temporal branch extracts four physically derived channels via a 1D convolutional stack enhanced with Squeeze-and-Excitation (SE) attention, while the spectral branch processes log-power Short-Time Fourier Transform (STFT) spectrograms through a learnable frequency gate. Both representations are fused via a cross-conditioned channel attention mechanism, enabling mutual temporal–spectral guidance. To assess generalization capacity and architectural stability, several training protocols were conducted—including strict GroupKFold cross-validation and single-model training—consistently yielding strong results: up to 98.77% accuracy, 98.68% F1-score, and 99% under single-model training, and 90.16% accuracy and 94.57% specificity under cross-validation, where cross-attention fusion outperforms naive concatenation by +5.62 percentage points (pp) in F1-score. An enhanced variant equipped with SE blocks and a deep classifier head further reduces performance variance by 28–51%. The consistently strong performance across all protocols confirms the generalization capacity and architectural robustness of Arc-FaultNet. With fewer than 365K parameters, Arc-FaultNet offers a practical pathway toward IEC 62606-compliant embedded deployment.