Application and analysis of deep learning in impulse voltage aging process

作者： GOZ Electric 时间：2024-08-26 09:36:09 阅读：11

1. Model optimization

Deep learning models 1 and 2 are trained and tested for vacuum breakdown mechanism identification using breakdown voltage and current waveform samples with a time extraction range of 0 to 400 μs and 0 to 200 μs, respectively. The effect of deep learning model 2 in vacuum breakdown mechanism identification is better than that of deep learning model 1, which shows that the training and testing of vacuum breakdown mechanism identification using breakdown voltage and current waveform samples with a time extraction range of 0 to 200 μs is better than that using breakdown voltage and current waveform samples with a time extraction range of 0 to 400 μs. The schematic diagram of the comparison and analysis of typical breakdown voltage and current waveforms is shown in the figure. As can be seen from the figure, the breakdown voltage and current waveform samples with a time extraction range of 0 to 200 μs amplify the pre-breakdown process, highlight the pre-breakdown voltage and current waveform characteristics, and increase the proportion of pre-breakdown voltage and current characteristics when the deep learning model extracts voltage and current waveform characteristics. From the analysis of the physical process, the characteristics of the voltage and current waveforms before breakdown are the key to distinguishing different vacuum breakdown mechanisms, and the pre-breakdown process occurs within 200 μs. The deep learning model 2 also uses the breakdown voltage and current waveform samples with a time extraction range of 0 to 200 μs for identification training and testing, thereby achieving better results.

    The deep learning model has achieved a high accuracy of 87.99% for vacuum breakdown mechanism identification training and testing using breakdown voltage and current waveform samples with a time extraction range of 0 to 200 μs. If further improvement is to be achieved, further improvement of the breakdown voltage and current waveform samples can be considered. A schematic diagram of an improved breakdown voltage and current waveform sample is shown in the figure. The breakdown voltage and current waveform samples with a time extraction range of 0 to 200 μs still cannot accurately locate the breakdown moment, which makes the deep learning model not maximize the proportion of the voltage and current waveform features before the breakdown when extracting the voltage and current waveform features. If the breakdown voltage and current waveform samples are image processed to accurately locate the breakdown moment, the proportion of the voltage and current waveform features before the breakdown process can be improved when the deep learning model extracts the voltage and current waveform features. This algorithm still needs further research.

2. Engineering application

    In engineering applications, impulse voltage aging technology is often used to improve the withstand voltage level of vacuum circuit breakers. The impulse voltage aging process includes a series of breakdown processes, and there is a conversion of vacuum breakdown mechanism. The breakdown mechanism dominated by different stages of aging has a regular evolution trend. By identifying the breakdown mechanism in the aging process, the aging evolution stage can be distinguished, so that a more reasonable aging scheme can be designed for different stages, thereby improving the insulation performance of vacuum circuit breakers. Therefore, it is of great significance to quickly and accurately identify the vacuum breakdown mechanism during the aging process for real-time monitoring, accurately determine the aging evolution stage, optimize the aging scheme, and improve the insulation capacity of vacuum circuit breakers.

    Related literatures have described the feasibility of deep learning technology in the aging process, and this paper aims to propose an optimization method to achieve rapid and accurate identification of the vacuum breakdown mechanism by highlighting the pre-breakdown process. The deep learning model 2 established according to the optimization method has an accuracy of 88.92%, 87.99% and 92.78% for the vacuum breakdown mechanism identification of the breakdown voltage and current waveform samples of electrodes C, D and E, respectively, and the precision, recall and F1 scores have all reached 87.44%, and the time it takes to complete a vacuum breakdown mechanism identification is only about 5s. The above results show that the deep learning model 2 can accurately, effectively and quickly complete the vacuum breakdown mechanism identification. The optimization method proposed in this paper effectively improves the accuracy of deep learning technology in the identification of vacuum breakdown mechanisms, making the application of deep learning technology in the aging process have a better prospect.

3. Conclusion

1) This paper proposes a vacuum breakdown mechanism identification optimization method based on deep learning. By comparing the effects of vacuum breakdown mechanism identification using breakdown voltage and current waveform samples with different time extraction ranges, it is found that waveform processing that highlights the pre-breakdown process can effectively optimize the vacuum breakdown mechanism identification effect.

2) According to the vacuum breakdown mechanism identification optimization method based on deep learning proposed in this paper, the vacuum breakdown mechanism can be quickly identified with an accuracy of 87.99%. The effectiveness of this method is verified by precision, recall and F1 score, which has theoretical guiding significance for engineering applications.