Research On Fault Feature Extraction And Real-time Location Strategy Of MMC Sub-module Based On Data Mining

Modular multilevel converter（MMC）shows extremely important engineering application prospects due to its modularity and scalability.For the MMC containing many power switches,the fault diagnosis and location of the sub-module is one of the most important issues in the MMC reliability research.The sub-modules（SM）of MMC contain a large number of power devices,so that each sub-module of MMC may become a point of failure,which may distort the voltage and current in the MMC,and even damage the MMC in severe cases.Therefore,MMC reliability research has become a research hotspot in academia and industry,and the study of MMC sub-module fault feature extraction and real-time location strategy is an important subject of MMC reliability research.This paper takes the HB-SM-based three-phase six-arm symmetrical structure MMC topology as the research object,and proposes a data mining-based MMC sub-module fault feature extraction and real-time location strategy,including the MMC sub-module sliding time window fault feature extraction strategy and the two-dimensional Convolutional Neural Networks（2D-CNNs）fault diagnosis strategy based on sliding time window feature extraction,starting from the ubiquitous relationship of data,provides new ideas for the research of MMC sub-module fault feature extraction and real-time location.The research content is as follows:（1）This article analyzes the working principle of MMC in detail and establishes a detailed MMC mathematical model.On this basis,specific MMC system control is given,including overall control,capacitor voltage balance control,and double-frequency circulating current suppression.Based on the MMC mathematical model and system control,the MMC sub-module open circuit fault characteristics are analyzed in depth,and the MMC sub-module is analyzed under the power switch T₁ open circuit fault and the power switch T₂ open circuit fault.（2）This paper proposes a sliding time window fault feature extraction strategy for MMC sub-modules.The sliding time data matrix is used to analyze the MMC’s infinite flow signal sequence such as capacitor voltage,circulating current,bridge arm current,DC side voltage,three-phase current and voltage,and time.Collective sampling,combined with sliding time window theory,proposes a fast wavelet packet transform algorithm based on sliding time window to extract the comprehensive characteristics of the fault signal in the time domain and frequency domain—wavelet packet coefficients,based on the existence of a part of the same wavelet packet between adjacent sliding time windows The feature relationship of the coefficient data group,using the data overlap relationship between adjacent sliding time windows to extract the comprehensive characteristics of the time domain and frequency domain of the sliding time data matrix signal,and pre-store a part of the previous sliding time during two adjacent sliding intervals The same wavelet packet coefficients of the signal sequence collected by the window avoid the repeated calculation process.At the cost of smaller storage space,a real-time feature extraction algorithm of MMC sub-module wavelet transform with high computational efficiency is obtained.This method is a new idea of applying big data processing technology to MMC system reliability research.Starting from the data ubiquitous relationship,the MMC sub-module switch open circuit fault research is carried out.At the same time,the best sliding time window interval is based on the Nyquist sampling theorem and the central spectrum analysis theory.（3）This paper proposes a two-dimensional Convolutional Neural Networks（2D-CNNs）fault diagnosis strategy based on feature extraction of sliding time windows,and analyzes the fault feature energy entropy feature expression and input of the sliding time window.The energy entropy feature of the fault feature is input as the input to the two-dimensional convolutional neural network.After the offline training step,the training data and results of a sufficient number of samples are accumulated,and real-time online fault detection and identification are performed to generate the probability of occurrence of possible fault types P（F）,to determine whether the P（F）value meets the fault standard,it can realize the detection and location of the open circuit fault of the sub-module at the same time.At the same time,the performance evaluation and comparison of 2D-CNNs fault feature diagnosis strategies based on sliding time window feature extraction are carried out.The proposed strategy can have the advantages of concise and low-data-volume feature samples in both the time domain and the frequency domain,and can effectively locate the fault,so that the fault can be located with high precision in a short time.In addition,it does not require the manual setting of the mathematical model of MMC and the empirical threshold.（4）Based on professional tools PSCAD/EMTDC simulation software and small-sized prototypes in the laboratory,the feasibility of the proposed 2D-CNNs fault feature diagnosis strategy based on sliding time window feature extraction is verified and the validity of the MMC sub-module switch open circuit fault diagnosis under different working conditions and working modes is verified.