Abstract:The series arc faults of the electric lines in the low-voltage distribution network can lead to devastating fires, posing a significant threat to the lives and safety of residents. Aiming at the problems that arc faults are challenging to identify due to the variety of load types in the lines, a new diagnosis method of two-level classification for arc faults based on data random fusion and MC-MGCNN network is proposed. Firstly, the first level of load classification is carried out. Extreme weighted fuzzy entropy is calculated to separate the dimmer load that could easily lead to misjudgment from the total load pool, and the concepts of simple load pool and complex load pool are established. Secondly, the corresponding classifiers are designed for the two different load pools for the second level of diagnosis classification. Specifically, the time–frequency domain features are extracted for the load waveforms in a simple load pool, and extreme gradient boosting (XGBoost) is used for rapid diagnosis. On the contrary, for the nonlinear load waveforms and multi-load combination waveforms in the complex load pool, the random fusion mechanism of data is applied for feature enhancement, and a new one-dimensional data set is constructed. Finally, the latest data is input into the multi-channel convolution neural network combining multi-head attention and gated recurrent unit (MC-MGCNN) network for arc fault diagnosis. The experimental results show that the first level of load classification can effectively reduce the number of misjudgment samples in the complex load pool and improve the overall binary classification accuracy by 7.54 %. The new data set gains different degrees of improvement in the identification accuracy of other comparative networks. In addition, the new network has good feature learning ability and diagnostic accuracy of arc faults. The accuracy of the multi-classification is up to 99.5 %, and the binary classification accuracy is up to 99.94 %. The test is carried out in the Raspberry PI 4b environment, and the detection time is 52.3 ms, which verified the feasibility of hardware deployment.

FauIt Arc Identification Method Based on K－means CIassification and BP NeuraI Network

Fault Diagnosis Method of High Power Charging Equipment Based on Neural Network

Research of Arc Fault Detctor Based on Neural Networks

Arc-Fault Recognition Based on BP Neural Network

Identification Method for Low-Voltage Arc Fault Based on the Loose Combination of Wavelet Transformation and Neural Network

An Constraint-based Clustering Integrated BP Networks Approach for Analog Circuit Fault Diagnose

Series Arc Fault Identification Method Based on Multi-Feature Fusion

DC Arc Fault Risk Degree Evaluation Based on Back Propagation Neural Network

A New Methodology for Identifying Arc Fault by Sparse Representation and Neural Network

Series AC Arc Fault Detection Method Based on Hybrid Time and Frequency Analysis and Fully Connected Neural Network

A two-level classification diagnosis method for AC arc faults based on data random fusion and MC-MGCNN network

Voltage‐based fault arc detection based on PCA‐RF

A Novel Series Arc Fault Detection Method Based on Mel-Frequency Cepstral Coefficients and Fully Connected Neural Network

An Arc Contacts Life Assessment Method Based on Dynamic Resistance Measurement and Bp Neural Network

Series Arc Fault Identification Method Based on Energy Produced by Wavelet Transformation and Neural Network

Fault Identification in Double-Circuit on Same Pole Using Neural Network

A Novel Arc Fault Detection Method Integrated Random Forest, Improved Multi-scale Permutation Entropy and Wavelet Packet Transform

Research and Application on Intelligent Fault Diagnosis Based on BP Neural Net

Series arc fault identification method based on wavelet transform and feature values decomposition fusion DNN

Fault Arc Detection Based on Channel Attention Mechanism and Lightweight Residual Network

Series Arc Fault Detection Based on Wavelet Compression Reconstruction Data Enhancement and Deep Residual Network