Research On Low-voltage DC Arc Fault Diagnosis Method Based On Deep Learning

In recent years,distributed photovoltaic power generation has developed rapidly,and battery energy storage systems and DC power supply systems are in the ascendant.The subsequent low-voltage DC arc fault detection problem has become the focus of research by scholars at home and abroad.How to accurately identify the arc fault of the low-voltage DC system and make protective actions in time is of great significance to ensure the safe operation of the DC power supply system.Thesis conducts research on low-voltage DC series arc fault diagnosis methods.Through multi-dimensional feature analysis of the current waveform of low-voltage DC series arc faults,the effective feature quantity is selected to form a feature vector group to express the fault arc information,and the low-voltage DC based on deep learning is used.The fault arc diagnosis strategy realizes the fault identification.The main work of thesis is as follows:(1)DC fault arc experiment and characteristic analysisAccording to the relevant test standards for arc faults at home and abroad,a low-voltage DC arc fault test platform was established.Conduct experiments on resistors,LED lights,fans and mixed loads to obtain current waveforms and analyze the characteristics of current waveforms of various loads under arc fault and normal operating conditions.(2)Extraction of characteristic quantities of fault current waveformIn time domain analysis,statistical characteristic parameters(pulse index,coefficient of variation)and singular entropy are used to characterize the distortion characteristics of current signal;In the aspect of time-frequency analysis,thesis proposes a time-frequency analysis method based on the combination of wavelet transform and morphological fractal dimension from the aspect of multi-scale feature analysis.This method uses the wavelet morphological fractal dimension to describe the arc fault,and realizes the description of the change characteristics of the phase frequency domain before and after the arc fault.Comprehensive analysis and construction of low-voltage DC series arc feature vector provide data support for arc fault identification.(3)Low voltage DC arc fault diagnosis model based on deep learningIn this dissertation,a probabilistic neural network model for low voltage DC series arc fault identification is designed.According to the influence of human behavior experience on the classification accuracy of the probabilistic neural network,the linear exhaustive method is used to select the appropriate spread factor,so as to improve the recognition accuracy of the probabilistic neural network.The multi-dimensional feature sample data set is used to train the probabilistic neural network and verify its effectiveness.In addition,the current waveform distortion caused by short-time arc and load start-up will also interfere with the identification of fault arc.Therefore,thesis optimizes the fault diagnosis model to reduce the influence of short-time current waveform distortion on the identification accuracy.