Arc-Fault Detection Using Stage-Wise Alignment and Feature Fusion of Dual Learnable Time–Frequency Representations

Arc faults in low-voltage smart-farm distribution systems are difficult to detect reliably because normal current waveforms vary with load composition and operating conditions. This paper presents a progressive three-stage time–frequency learning framework that identifies series arc faults directly from normalized current waveforms. In Stage I, fast Fourier transform (FFT)-based and wavelet-based learnable front-ends independently learn complementary global spectral and local transient representations. In Stage II, their representation spaces are structured using supervised contrastive learning with view-level and cross-spectrum alignment. In Stage III, the aligned feature extractors are frozen, and only the feature fusion block and final classifier are trained for 10-class classification. Current signals are acquired from a practical smart-farm distribution system and evaluated using a measurement-session-level group-wise split. The framework achieves 99.92% accuracy and 99.93% macro recall while maintaining 1.06 M parameters and a model-only inference latency of 8.70 ms. The results demonstrate that the proposed framework provides robust discrimination across unseen measurement sessions within the evaluated load categories and operating conditions.