Degradation Research Of Capacitors Based On Time Series Analysis

With the development of technology and the improvement of social demand,the application range of capacitors has become wider and wider.As a kind of component used in voltage-stabling and noise-eliminating,capacitors directly determine whether an electrical system can operate normally or not.Therefore,it is very important to study the reliability of capacitors.There are many methods to study reliability,the degradation research according to performance characteristic values is one of the frequently-used methods.As for electronic products such as capacitors,the capacity values are are generally used as performance characteristic values.In order to understand the degradation characteristic values of capacitors and to build corresponding research methods,we established an accelerated degradation experimental platform,and conducted a stress test on a batch of aluminum electrolytic capacitors.In order to effectively extract data information from the degradation characteristic values,theexperimental data of capacitors as time series were systematically studied in the dissertation.For data characteristics of capacitors,we proposed capacitor degradation prediction methods based on time series analysis.The time series characteristics such as stability of the capacitor degradation data were analyzed by means including unit root test,autocorrelation analysis and partial autocorrelation analysis.The long-term memory of the degradation data was studied by rescaled range analysis.A general autoregressive moving average model and an autoregressive moving average model with fractional difference were established,among which the autoregressive moving average model with fractional difference is suitable for data with long-term memory.There are a few application examples for this model in manufacturing industries,and the model has not been used by other people in the study of capacitor degradation.Based on the minimum information criterion and the maximum likelihood estimation,we completed the parameter selection and parameter estimation of the two models.The two degradation prediction models based on time series analysis were validated using degradation data of the experiment.The experimental results showed that under the given rules of sample partitioning,the prediction accuracy of the ordinary autoregressive moving average model is slightly better than that of the fractionaldifference autoregressive moving average model.In addition,we developed an open source package of fractional difference autoregressive moving average model with language Python,which can be directly used by other scholars who study this model.In order to prove the validity of the degradation prediction method based on time series analysis,the prediction results of the Wiener process model which has been proved to be more effective were used as the standard values.The results showed that the prediction accuracy of the proposed method is similar to that of the traditional model,so the method proposed in this thesis is effective.And the results of analysis of residuals and the prediction accuracy proved the validity of the two models applied in capacitor degradation analysis.Then,the pseudo-life and the reliability of the capacitors under the operating temperature are calculated using the prediction results from the time series models.This provides a method for obtaining estimation of the reliability of electronic components without performing long-term experiments.In addition,by using difference equation theory we proposed an over-differential prejudgment method for degradation prediction models based on time series analysis,which is helpful for model parameters selection of electronic products subject to stochastic processes.In order to compare the effectiveness of different models and their research methods in various degradation analysis,we used a model basedon extreme gradient boosting algorithm from ensemble learning models to predict and analyze the experimental data of capacitors based on time series differential operation.This algorithm is novel and popular in ensemble learning,so we used it for capacitor degradation analysis.Research results show that extreme gradient boosting algorithm has certain prediction accuracy under the designed test set and training set,but the prediction accuracy is lower than that obtained by the two degradation prediction models based on time series analysis.