Fault diagnosis of modular multilevel converter based on principal component analysis and support vector machine

Modular multilevel converter (MMC) is widely used in DC transmission, new energy grid-connected power generation, transmission, reactive power compensation, power flow control and other fields. When the sub-module fails, detecting and locating the fault quickly and accurately is the key to improving the operational reliability of the converter. Principal component analysis (PCA) obtains the feature space of reduced dimension by extracting the principal components of the fault sample set, which is conducive to the extraction of fault features. Support vector machine (SVM) has good classification performance when applied to fault diagnosis. Combining the advantages of both, this paper takes modular multi-level converter as the research object, extracts the fault characteristics of MMC, and uses PCA algorithm to reduce dimensionality. Then, the SVM classifier is constructed, the processed fault samples are used for training, and the trained SVM classifier is used to perform fault diagnosis. Finally, a three-phase eleven-level MMC simulation model is built to simulate the method used. The results show that this method can effectively improve the diagnosis speed and accuracy of MMC fault diagnosis, and provides a reference for the application of the PCA-SVM method in the actual engineering of MMC fault diagnosis.


Introduction
With the continuous increase in the scale of new energy installations and the gradual development of intelligent power grids, modular multilevel converters are increasingly used in fields such as flexible DC transmission and new energy grid connection. Each bridge arm of the MMC is composed of multiple sub-modules (SM) with the same parameters in series and divided by voltage, failure of any sub-module will cause the bridge arm to work abnormally and affect the normal operation of MMC. Therefore, when a sub-module fails, the fault can be quickly and accurately determined and the faulty sub-module can be replaced through a redundant control strategy, which can greatly improve the reliability of MMC operation [1].
Among the components of MMC, the power device of the sub-module is the component with the highest failure rate. The short-circuit fault of the power device is generally detected by the hardware protection circuit, and the scheme is relatively mature, while the open-circuit fault of the power device is not easy to find [2], but it will cause serious damage to the system if it is allowed to develop. The open-circuit fault diagnosis methods for MMC sub-modules at home and abroad mainly include hardware-based methods [3], model-based methods [4], and machine learning-based methods [5]. Literature [5] proposes a pattern recognition method based on K-means clustering, which detects and locates faulty sub-modules through the two-dimensional trajectory pattern of sub-modules, but this method is computationally intensive and complex. Literature [6] proposes a feature extraction and 2 dimensionality reduction method based on the combination of WPD and PCA, then the optimized BP neural network is used to achieve fault location. Literature [7] summarizes the fault operation characteristics when different faults occur through simulation analysis, and provides a reference for accurately locating MMC faults.
This paper selects the fault diagnosis and identification method of power electronic rectifier device based on the combination of PCA-based fault feature extraction and support vector machine, and applies this method to the open-circuit fault diagnosis of the MMC sub-module. Firstly, this paper studies the mechanism of MMC under normal operation and fault conditions, and selects the signal that best reflects the open circuit fault of the sub-module as the fault characteristic parameter; Secondly, the PCA algorithm is used to reduce the dimensionality of the fault feature, the SVM classifier is constructed, the processed fault samples are used for training, and then the trained SVM classifier is used for fault diagnosis; Finally, an MMC simulation model is built to verify the fault diagnosis method used.

Open-circuit fault analysis of MMC sub-module
At present, the most commonly used MMC topology in engineering is shown in Figure 1, and its submodule structure is a typical half-bridge structure [8]. There are four modes in the internal current path of the sub-module when the MMC is running normally, as shown in Table 1.  Table 1 shows that if 1 T fails, it will affect the circuit in mode 2; if 2 T fails, it will affect the circuit in mode 2; Figure

Fault detection method based on PCA-SVM
This paper presents an MMC fault diagnosis method based on the PCA-SVM algorithm model, and then introduces the theoretical principles of the method used.

Fault feature extraction based on PCA
According to the principle of PCA, the feature extraction algorithm of fault signal based on PCA is as follows.
Step 4. Solve the eigenvalues and eigenvectors of the covariance matrix, that is, solve the eigenvalues of the construction matrix T Cx XX  and the corresponding orthogonal normalized eigenvectors i  .


Step 5. Solve the matrix A . When the dimension of the original fault signal is much larger than the fault training sample, that is There is Through the above steps, the N-dimensional feature vector of the fault signal can be effectively extracted.

Improved SVM multi-classification algorithm
Support Vector Machine is a generalized linear classifier for binary classification of data. The algorithm is derived from the two-class classification problem, which is proposed in the case of linear separability. For general nonlinear problems, nonlinear transformation is achieved by defining an appropriate kernel function, the input space is transformed into a high-dimensional space, and then the optimal linear classification surface is found in this new space [9].
SVM was originally proposed for two classification problems, and the converter fault diagnosis not only determines whether the rectifier is faulty, but also determines the location of the fault. It is a multi-value classification problem, and multiple second-class SVM classifiers need to be combined to construct the SVM multi-classifier.
Taking into account the structural characteristics of the converter circuit itself, the original "one-tomany SVM" algorithm is improved here. Specific improvement ideas include:  When constructing the i -th classifier of the N fault type classifiers, the fault training samples belonging to the i -th SVM are regarded as one category, and the category label is 1, the other remaining training samples are regarded as one category, and the category label is -1.  In order to improve the accuracy of diagnosis, the process of classifying fault test samples is improved. The decision function used is still The improved classification process is shown in Figure 3.  The fault training sample or test sample extracts the N-dimensional feature vector by PCA, which is used as the input of the trained SVM multi-classifier, and the output of the SVM is the corresponding fault type, so as to realize fault diagnosis and fault location.

Simulation
In order to verify the correctness of the PCA-SVM-based fault diagnosis method, this paper builds a three-phase eleven-level MMC simulation model based on the MATLAB/SIMULINK platform. For an eleven-level MMC inverter system, each phase has two upper and lower bridge arms, and each bridge arm is composed of a current-limiting reactor and ten sub-modules in series. The simulation model diagram is shown in Figure 5.  Figure 6 shows the A-phase output voltage and A-phase bridge arm current waveform during normal operation. The A-phase output voltage amplitude is about 4500V, which conforms to the set DC voltage parameters, and the waveform shape is an eleven-level ladder, approaching a sine wave. The current waveform of the A-phase bridge arm is completely symmetrical and without distortion, which verifys the correctness of the model.  150 data samples are selected for each failure mode in the low-dimensional feature space, each sample contains 15 feature components, and the feature training set (50 samples for each failure mode) and test set (100 samples for each failure mode) are selected to construct the SVM classifier for training and fault diagnosis. In order to ensure high classification accuracy, this paper chooses RBF as the kernel function, and the kernel width is 6.
Taking the fault set in this article as an example, the final diagnosis accuracy distribution of training samples and test samples is shown in the Table 2. The results show that when the PCA-SVM algorithm is used in the field of MMC fault diagnosis, for the fault training samples, the recognition rate is basically 100%, and for the test samples, the recognition rate can reach close to 95%, which effectively improve the diagnosis accuracy of MMC fault diagnosis.

Conclusion
Modular multilevel converters have a large number of sub-modules, which increases the probability of failure of sub-modules. Therefore, rapid diagnosis and accurate location of sub-module faults are important guarantees for continuous and uninterrupted operation of the converter. This paper analyzes the change characteristics of the bridge arm current and the sub-module capacitor voltage in each working mode when the MMC sub-module has an open-circuit fault, and introduces a PCA-SVMbased fault detection method, which combines the advantages of the PCA and SVM algorithms respectively. First, the fault signal is compressed by principal components and the fault feature vector is extracted, then the extracted fault feature vector is sent to the SVM classifier for training and fault diagnosis. This article uses the above-mentioned method to simulate the fault diagnosis of MMC. The results verify the correctness and effectiveness of the method applied to MMC fault diagnosis. In addition, this method combines the advantages of PCA and SVM algorithms, which can also be extended to other forms of power electronic equipment fault diagnosis, and for the simultaneous failure of multiple sub-modules of the same bridge arm in MMC, further research can be carried out.