低压交流串联故障电弧辨识方法

Traditional protection switching devices can not open the line that occurs series arc fault in time because the current amplitude when the line is in a series arc fault state is lower than that in a normal state. Recently, the development of machine learning has provided a new idea for detecting an arc fault. However, a large amount of data is needed to train an arc fault detection model, and it is expensive to make a circuit module that uses a machine learning algorithm to recognize arc faults, which causes troubles in promotion and application of arc fault detection device (AFDD). Therefore, this paper introduces an arc fault recognition method based on multi-feature fusion. Firstly, a test device is constructed, and series arc fault experiments are carried out under different load types, accumulating a large number of test data. Secondly, the arcing processes of arc fault are observed within a smaller time scale by a high-speed camera to establish the relationship between the arcing process’s physical phenomenon and arc current waveform. Besides, the feature of circuit current waveform is analyzed when the line occurs arc fault. Thirdly, based on current zero and current ripple features of circuit current, three characteristics, current average, harmonic amplitude sum, and wavelet energy entropy, are extracted from the time domain, frequency domain, and signal disorder degree, respectively. To avoid the influences of different load types on characteristic threshold selection, the characteristic ratio between arcing state and normal state serves as a reference for characteristic threshold selection. Finally, the prototype of AFDD is made by a universal microcontroller (STM32), and the accuracy of the arc fault detection method is verified by experiments. Besides, the stability of the method is also verified by arc fault simulations (arc generated by carbonization line) and special loads starting experiments. The photos of arcing processes intuitively reflect the current zero phenomena and instability of arcing. These physical phenomena can be characterized by the current zero feature and current ripple feature of the current waveform. Meanwhile, three characteristics can distinguish arcing and normal states of the line under different load types. In addition, each characteristic value is different under different load types, causing difficulty in selecting the threshold between normal and arcing states. However, the characteristic ratio exists similarity under different load types. The characteristic ratio can serve as a reference for characteristic threshold selection. Through experimental verification, the accuracy of the arc fault detection method is 90%, and the applicability of the method is not affected by modes of arc fault simulation experiment. The following conclusions can be drawn from experimental results: (1) Both the current zero feature and the current ripple feature of arcing processes can be used to identify the arcing state of the line. (2) Three characteristics, current average, harmonic amplitude sum, and wavelet energy entropy are effective for distinguishing arcing and normal states of the line under different load types. (3) The characteristic ratio of arcing and normal states can provide references for characteristics threshold selection, avoiding the effects of different load types on threshold selection.