An arc fault detection method based on a one-dimensional dilated convolutional neural network

[Objective] The arc fault of the low-voltage distribution system is one of the primary causes of residential fires. Due to the diverse load types and complex connection methods in residential areas, arcs fault exhibit many similar and concealed characteristics, making them difficult to detect. This frequently leads to issues with arc fault protection devices, such as false alarms and missed detections. The conventional detection method, which is based on manually extracting arc fault feature vectors, is incomplete and heavily relies on expert-designed features. Consequently, this impedes the development of highly generalizable models. In addressing these challenges of multiload systems, this paper proposes a method for serial arc fault diagnosis and load recognition based on a one-dimensional dilated convolutional neural network (1D-DCNN). [Methods] First, a series of experiments on multiload arcs fault are conducted using a custom-designed experimental platform. This platform supports both single-load and dual-branch load conditions during testing. Normal and faulty current data on the main bus are sampled under various operating conditions at a rate of 500 kHz. During the data processing phase, the continuous time series data are discretized and normalized based on the half-cycle length. Subsequently, a 1D-DCNN is used to extract features from the high-sampling-rate arc fault current data. Furthermore, the scaled exponential linear unit activation functions and residual connections are introduced to address the challenges of gradient vanishing and network degradation. Moreover, the cyclic padding method is adopted to alleviate boundary effects and enhance the model s robustness to dataset shift. The arc fault detection model is developed by integrating average ensemble learning with a Softmax multiclassifier. Precision, recall, and specificity are used to assess the efficiency of the model. Finally, the accuracy of the proposed model in load classification, load state recognition, and overall accuracy is compared with that of other classical models, providing a comprehensive assessment of its efficacy. [Results] The findings of this method were as follows: (1) The accuracy of arc detection using a recurrent neural network was considerably low, primarily due to gradient vanishing and exploding gradients, making it difficult to effectively train the model. (2) Under the condition of equal parameter count between a 1D-DCNN and 1D-CNN, the dilated convolutional operation expanded the receptive field, resulting in greater accuracy than the 1D-CNN model. (3) The proposed method achieved a remarkable accuracy of 99.67% in detecting arc faults for both single-load and mixed-load scenarios, with an accuracy of 99.95% and an overall accuracy of 99.62%. [Conclusions] This research presents a unique model capable of autonomously learning features from high-sampling-rate current data without requiring manual feature extraction. It efficiently detects arcs fault while identifying the type of faulty load simultaneously. The model outperforms typical convolutional neural networks on validation of the test set, thereby meeting the requirements for arc fault identification. This advancement has major implications for serial arc fault detection and load recognition applications. © 2024 Tsinghua University. All rights reserved.