Arc faults pose challenges to electric safety, which can cause serious fire hazards. However, the commonly used arc fault detection method is prone to nuisance tripping. This paper proposed a hybrid arc fault detection method based on the improved Mel-Frequency Ceptral Coefficients (MFCC) for preprocessing and a neural network model for arc identification called ARC_MFCC. As per the IEC 62606, twelve different loads/scenarios are considered for this research. An arc tangent-based core filter is employed to improve the MFCC to enhance the arc features within a bandwidth of 3 kHz to 7 kHz. A lightweight neural network model of fully connected cascaded with the MFCC-based preprocessing, which can distinguish the arc fault with normal operation under different test conditions. As a verification result, the ARC_ MFCC can achieve an accuracy of 99.34%. Moreover, the proposed method is implemented by Raspberry pie 4B. Test results show an average running time of about 4.2 ms per sample, which Ensures that the tripping time can meet IEC 62606.

Arc fault detection devices are mandatory worldwide for mitigating DC series arc faults in photovoltaic systems. However, they are prone to nuisance tripping. Artificial intelligence-based approaches can be a solution, but they are "black boxes" and challenging to modify. This paper proposes an explainability and attention-based method to investigate the intensive details of such an algorithm. The contributions of an arc feature to the proposed model can be visualized by the proposed interpretable methodology so that insensitive arc features can be removed to reduce the quantity of input data. Additionally, the structure of the proposed model can be optimized by cutting the redundant layers. Thus, an accuracy of 99.63% is achieved with only 48.48% of the parameters compared to the original model. Finally, the optimized model is implemented by a Cortex M7-based microprocessor with a runtime of only 7.8 ms, making it ready for industrial application.

Arc faults can cause a severe electric fire, especially series arc faults. Artificial intelligence (AI)-based arc fault detection methods can have a boasted detection accuracy. However, the complexity and large parameter quantity of the AI-based algorithm will hinder its real-time performance for detecting series arc faults. This paper proposes a lightweight arc fault detection method based on the EffNet module, which can make the algorithm less complex with the same detection accuracy level. An arc fault test platform was constructed to collect arc current data, covering eight types of loads required by the IEC 62606 standard. The raw arc current data are used directly as an input for the proposed algorithm, reducing the module’s complexity. According to features of arc current mainly represented in the time domain, the first and last convolution layers of the EffNet module can be improved. Additionally, the spatially separable convolution is well-tuned and trimmed to achieve a more lightweight and better-performance architecture for arc fault detection called Arc_EffNet. Remarkably, this model achieves an impressive arc detection accuracy of 99.75%. An arc fault detection prototype was built using the Raspberry Pi 4B to evaluate the real-time detecting performance of the proposed method. The experimental results show that the prototype takes a time of about 72 ms to respond to a series arc fault, which can fulfill the requirement of real-time detection for arc faults.