Abstract:

To address the insufficient generalization performance of deep learning-based arc fault detection methods in unknown multi-load circuits，a Filtering Long Short-Term Memory （Filtering LSTM） neural network driven by high-frequency coupled analog signals is proposed. By combining this network with a Lightweight convolutional neural network （Lightweight CNN），a Filtering LSTM-Lightweight CNN arc fault detection model is constructed. The high-frequency coupled signals of multi-load circuits can be simulated through the linear superposition of high-frequency coupled signals from single-load circuits. These analog signals are then used to drive the Filtering LSTM，which filters out unknown features in the multi-load circuit signals and reconstructs the signals. Finally，a Lightweight CNN optimized by the tree-structured Parzen estimator is employed to perform arc fault detection on the reconstructed signals. Experiments demonstrate that the Filtering LSTM-Lightweight CNN achieves an arc fault detection accuracy of 99.45% among 136 000 unknown multi-load circuit samples. Compared with detection algorithms that do not adopt Filtering LSTM，the proposed method improves the detection accuracy by up to 14.05%，significantly enhancing the generalization ability of the arc fault detection model.

Key words: series arc fault, feature filtering, Lightweight convolutional neural network （Lightweight CNN）, fault detection