Study On Condition Monitoring Data Quality Improvement And Fault Identification For Power Transformer

As one of the most important and key equipment,a power transformer is the core of energy conversion and transmission in the power grid.In order to monitor the actual operating condition of the transformer,a large number of condition monitoring devices and smart meters have been equipped.At present,the condition monitoring data of the transformer are increasing rapidly and show the characteristics of large volume and high dimensions.Under this background,how to make full use of these different types of condition monitoring data to improve the accuracy of transformer fault identification results has become an urgent research area.To solve the bottleneck problems in fault identification,such as low data quality,scarce fault data,and poor performance of conventional fault identification methods,the related research work,including quality improvement of condition monitoring data,fault condition detection,and failure probability calculation of the transformer,were studied in this dissertation.The main innovative achievements were obtained as following.(1)According to the characteristics of normal data,valid and invalid anomalies,an anomaly detection method based on auxiliary feature vector and density-based clustering(DBSCAN)was proposed.The auxiliary feature vectors of each condition parameter were constructed for clustering to recognize the accurate identification of different normal data patterns,valid anomalies containing the fault information,and invalid anomalies in condition monitoring datasets.To improve the precise and recall of anomaly detection results in unsupervised settings,a heuristic method for parameter selection of DBSCAN model based on the "Number of clusters–Eps" curve was proposed.Compared with state-of-the-art anomaly detection techniques,the proposed method shows the ability of both pattern recognition and anomaly detection,and solves the drawback of cleaning the valid anomalies.Different application examples are implemented on data of dissolved gas content in transformer oil,and the proposed method performs accurately and robustly in different datasets,provided that the condition monitoring data satisfy the assumption of stationarity.(2)According to the characteristics and severity of missing values,three types of missing data were defined,namely isolated missing value,continuous missing variable,and continuous missing sample.Following the principle of data imputation from easy to difficult,a three-step data imputation method was proposed to impute missing values in condition monitoring datasets.These missing values were imputed using the one-dimensional interpolation methods,the regression-based methods,and the stepwise extrapolation prediction model based on long short-term memory network,respectively.Two application examples are implemented on a dissolved gas analysis(DGA)dataset and a load dataset of transformers,the results show that the proposed method is appropriate for imputing both stationary and non-stationary condition monitoring data,and shows better performance than that of a kind of conventional or novel data interpolation method.(3)A fault detection method based on unsupervised concept drift recognition and dynamic graph embedding(DGE)was presented and applied in scenarios when fault data were scarce and the time of fault occurrence was unknown.Cosine similarity and multiple linear regression models were applied to identify the concept drift of the unlabeled time series data,as well as the boundaries between normal and fault data to help select the appropriate offline modeling data for the fault detection model.When detecting the faults of transformer,only normal historical data were needed for DGE model,thus the performance was not affected by the scarcity of fault data.Compared with the conventional threshold-based detection techniques,the proposed method can accurately identify the potential and less severe faults of the equipment.(4)Although traditional fault diagnosis methods can qualitatively identify the fault modes of transformer,it is difficult to evaluate the failure probability quantitatively.A failure probability calculation method for transformer based on association rules analysis technologies was proposed.The Weibull probability distribution model and the Apriori algorithm were applied to mine the association rules and quantify the importance weights between each type of condition parameter and different fault modes.After that,the failure probability of different fault modes and the transformer was calculated.The application examples show that the proposed method is fully data-driven without using any expert experience or special knowledge.Therefore,the objectivity and accuracy of the calculation results are ensured.(5)The comparative experiments in different transformer fault identification scenarios indicate that the invalid anomalies and missing values in condition monitoring datasets are key factors affecting the performance of fault identification methods.Anomaly detection and missing value interpolation can effectively improve the validity and integrity of the original data,which is of great significance for obtaining the accurate and reliable fault identification results.