A Fault Diagnosis Method for Medium Voltage Modular Multilevel Converter

Abstract

The invention discloses a medium-voltage modular multi-level converter fault diagnosis method, which belongs to the field of power electronics. The method constructs the collected capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multi-level converter MMC into a data set, wherein the data set includes normal operation state data and fault operation state data of the submodules of the modular multi-level converter MMC; performs unique-hot encoding on the sample set according to the operation state of the modular multi-level converter to obtain an encoding vector as a class label; adopts a neural network and performs parameter training on the training set based on a BP back propagation algorithm to obtain a modular multi-level converter fault diagnosis model; inputs a test set into the MMC fault diagnosis model to perform open circuit fault diagnosis on each submodule of the modular multi-level converter, and obtains the fault type by matching with the class label; and does not need to introduce new devices such as sensors, and does not increase system cost and complexity.

Description

Technical Field

[0001] The invention belongs to the field of power electronics and relates to a medium voltage modular multi-level converter fault diagnosis method. Background Art

[0002] As a converter used in medium and high voltage fields, modular multilevel converter (MMC) has attracted more and more attention due to its good scalability, flexible switch state combination, superior voltage and current resistance, and good output waveform quality. It is especially popular in the field of high-voltage flexible DC transmission and medium-voltage converter equipment in high-speed railways. However, due to the large number of switches involved, the open circuit fault of the switch has also attracted people's attention.

[0003] At present, with the development of modular multi-level converter technology, a variety of fault detection and positioning methods have been proposed. Some scholars have proposed a sub-module fault detection method using a Kalman filter. This method detects the existence of a fault by comparing the state variables estimated by the Kalman filter with the measured state variables. Some scholars have also proposed a fault detection method based on a sliding film observer, which can also detect faults. However, it is difficult to establish an accurate mathematical model for these methods. Later, some scholars proposed to add hardware circuit detection equipment, but this method increases the cost and system complexity. In recent years, some scholars have proposed the use of machine learning for fault location, but it has only been carried out to the fault arm level or fault type level, and the specific location of the fault point cannot be determined. Therefore, it is necessary to find a sub-module open circuit fault diagnosis method that does not increase the system burden and is simple to implement. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a medium voltage modular multi-level converter fault diagnosis method, comprising the following steps:

[0005] S1: collecting the capacitor voltage value of each submodule and the current value of each bridge arm of the modular multilevel converter once every N sampling periods, and recording the switch operation status data of each submodule of the modular multilevel converter MMC in each sampling period;

[0006] S2: constructing the collected capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multilevel converter MMC into a data set, wherein the data set includes normal operating state data and fault operating state data of the submodules of the modular multilevel converter MMC;

[0007] S3: constructing a data set of capacitor voltage values ​​of each submodule and current values ​​of each bridge arm of the modular multilevel converter MMC to perform feature reconstruction and sample screening to obtain new feature values, and forming a sample set with the new feature values, dividing the sample set into a training set and a test set, and performing one-hot encoding on the sample set according to the operating state of the modular multilevel converter to obtain an encoding vector as a class label;

[0008] S4: using a neural network and performing parameter training on the training set based on a BP back propagation algorithm to obtain a modular multilevel converter fault diagnosis model;

[0009] S5: Input the test set into the MMC fault diagnosis model to perform open circuit fault diagnosis on each submodule of the modular multilevel converter, and obtain the fault type by matching with the class label.

[0010] Furthermore, the step of constructing a data set from the capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multilevel converter MMC to perform feature reconstruction and sample screening to obtain a new feature value comprises the following steps:

[0011] S3-1: construct a preliminary sample set according to the preliminary sample set formula using the collected submodule capacitor voltage, each bridge arm current and recorded switch state data;

[0012] S3-2: Filter samples according to the fault type. If an open circuit fault occurs in the switch device in series with the capacitor, only retain samples when the input current of the faulty submodule is less than zero; if an open circuit fault occurs in the switch device in parallel with the capacitor, only retain samples when the input current of the faulty submodule is greater than zero; and then obtain a new characteristic value.

[0013] Furthermore, the preliminary sample set formula is as follows:

[0014]

[0015] Among them: U pji (k) represents the capacitor voltage value of the i-th submodule of the j-th bridge arm of the p-phase at the k-th moment; I pj (k) represents the current value of the j-th bridge arm of the p-phase at the time k; S pji g represents the switch operation state of the ith submodule of the jth bridge arm of phase p at time k+1-g, where '1' means the submodule is switched on and '0' means the submodule is switched off; C SM is the capacitance value of the submodule, Ts is the sampling period, G pji The reconstructed eigenvalues ​​representing the i-th submodule of the j-th bridge arm of the p-th phase are the preliminary eigenvalues.

[0016] Furthermore, the encoding rules of the one-hot encoding are as follows:

[0017] When the modular multilevel converter is in a normal state, the encoding rule of one-hot encoding is 0 0...0 1;

[0018] When the modular multilevel converter is in a fault state, the encoding rules of one-hot encoding are as follows;

[0019] If the fault of the modular multilevel converter occurs in the i(1, 2....., M-1, M)th submodule of the j(1, 2)th bridge arm of the p(1, 2, 3)th phase, then the (M*(p-1)+M*(j-1)+i)th bit from the highest bit is '1' and the rest of the bits are '0'; where p(1, 2, 3) represents the three phases A, B, and C respectively, j(1, 2) represents the upper and lower bridge arms respectively, and i(1, 2....., M-1, M) represents the submodule number of a single bridge arm respectively.

[0020] Furthermore, the neural network activation function is a sigmoid function, whose input is a 6M-dimensional vector, representing the reconstructed features of 6M sub-modules respectively, the hidden layer has 10 units, and the output layer has 6M+1 units, representing 1 normal working state plus 6M fault locations respectively.

[0021] Furthermore, the MMC fault diagnosis model includes a Softmax classifier; the Softmax classifier outputs probability values ​​belonging to various states of the converter, selects a class label that matches the maximum probability, and obtains the fault type.

[0022] Furthermore, the Softmax classifier formula is:

[0023]

[0024] In the formula, y i is the probability value corresponding to the i-th state; m is the number of output neurons, that is, the corresponding states of the switch.

[0025] A medium voltage modular multilevel converter fault diagnosis method provided by the present invention is a modular multilevel converter fault diagnosis method based on feature reconstruction and neural network, which has the following beneficial effects:

[0026] (1) In the method of the present invention, the generalization ability of the fault diagnosis model is improved through feature reconstruction and sample screening in the input data, thereby improving the diagnostic accuracy of the model;

[0027] (2) In the method of the present invention, in the MMC submodule open circuit fault diagnosis, the neural network used has only three layers of input layer, output layer and hidden layer. The structure is simple and can be quickly located, thus avoiding secondary faults and system crashes caused by slow fault diagnosis speed.

[0028] (3) In the method of the present invention, the fault location process is different from that of conventional fault detection in that fault detection and fault location are completed at one time and the algorithm is simple;

[0029] (4) The method of the present invention is different from the traditional machine learning fault diagnosis results in that it can not only locate the position of the faulty bridge arm by decoding based on the neural network output vector, but also locate the specific submodule of the fault point, which reduces the difficulty of accident investigation and fault-tolerant control.

[0030] (5) The method of the present invention does not require the introduction of new sensors and other devices, and does not increase system cost and complexity;

[0031] (6) This method deeply extracts sample features, constructs a fault vector more clearly through the newly proposed feature reconstruction method, and uses a Softmax classifier to classify the operating status and specific fault location at one time. The present invention has good diagnostic performance for sub-module open circuit faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a three-phase MMC topological structure diagram in an embodiment of the present invention;

[0033] Figure 2 : is a topological structure diagram of an MMC submodule in an embodiment of the present invention;

[0034] Figure 3 It is a flow chart of a modular multi-level converter fault diagnosis method based on feature reconstruction and neural network of the present invention;

[0035] Figure 4 It is a sample waveform diagram after feature reconstruction after the No. 1 submodule of the upper bridge arm of phase A in the MMC of the embodiment of the present invention is opened. DETAILED DESCRIPTION

[0036] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0037] Figure 1 This is a three-phase topology diagram of the MMC in the embodiment of the present invention. As shown in the figure, the MMC three-phase six bridge arms, each bridge arm is composed of three sub-modules (SM) with completely identical structures and a bridge arm inductor for limiting circulating current fluctuations. pji represents the voltage value of the submodule No. i of the j-th bridge arm of the p-phase; I pj Characterizes the bridge arm current value, L o is the bridge arm inductance, La is the load side inductance, and Ra is the load side resistance.

[0038] Figure 2: is a topological structure diagram of the MMC submodule in an embodiment of the present invention. In this embodiment, a half-bridge submodule is used, such as Figure 2 As shown, each submodule has two IGBTs and two reverse-connected diodes, and a parallel submodule capacitor, wherein VT1 and VT2 represent IGBT devices; VD1 and VD2 represent anti-parallel diodes; C SM is the capacitance value of the submodule; U pji is the voltage across the submodule.

[0039] In this embodiment, a method for diagnosing faults of a medium voltage modular multilevel converter is provided, and the process is as follows: Figure 3 As shown, the following steps are included:

[0040] S1: extract and collect the submodule capacitor voltage value of the modular multilevel converter once every 20 sampling periods, and record the switch state and bridge arm voltage of the grid module of the modular multilevel converter MMC in each sampling period. The collected operating state includes the normal state and the faults of different submodules;

[0041] S2: constructing the collected capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multilevel converter MMC into a data set, wherein the data set includes normal operating state data and fault operating state data of the submodules of the modular multilevel converter MMC;

[0042] S3: constructing a data set of capacitor voltage values ​​of each submodule and current values ​​of each bridge arm of the modular multilevel converter MMC to perform feature reconstruction and sample screening to obtain new feature values, and forming a sample set with the new feature values. The sample set is divided into a training set and a test set. 80% of each operating status data set can be divided into a training set and 20% into a test set. The two sample sets perform unique hot encoding on the sample sets according to the operating status of the modular multilevel converter to obtain an encoding vector as a class label; Figure 4 It is a sample waveform diagram after feature reconstruction after the No. 1 submodule of the upper bridge arm of phase A in the MMC of the embodiment of the present invention is opened.

[0043] S4: Use a neural network and perform parameter training on the training set based on the BP back propagation algorithm to obtain a modular multi-level converter fault diagnosis model; use known normal data and single-point open circuit fault data sets as training sets, and perform parameter training based on the back propagation algorithm. The activation function selected in this step is the sigmoid function; the structure of the neural network to be trained is a three-layer structure, the input is a 6*M dimensional vector, M represents the number of bridge arm submodules, where M is 3, the hidden layer has 10 units, the output layer has 6M+1 units, and the final output is a 6*3+1 dimensional vector (6M fault submodule positions plus a normal operating state).

[0044] The parameters of the neural network used are as follows:

[0045] Learning rate 0.001; training times 5000; number of hidden layer units 10; activation function Sigmoid.

[0046] S5: Input the test set into the MMC fault diagnosis model to perform open circuit fault diagnosis on each submodule of the modular multilevel converter. By matching with the class label, the fault type is obtained. The final fault diagnosis accuracy reaches 99%.

[0047] Further, the collected capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multilevel converter MMC are constructed into a data set, wherein the data set includes normal operation status data and fault operation status data of the submodules of the modular multilevel converter MMC;

[0048] S3-1: construct a preliminary sample set according to the preliminary sample set formula using the collected submodule capacitor voltage, each bridge arm current and recorded switch state data;

[0049] S3-2: Filter samples according to the fault type. If an open circuit fault occurs in the switch device in series with the capacitor, only retain samples when the input current of the faulty submodule is less than zero; if an open circuit fault occurs in the switch device in parallel with the capacitor, only retain samples when the input current of the faulty submodule is greater than zero; and then obtain a new characteristic value.

[0050] Furthermore, the preliminary sample set formula is as follows:

[0051]

[0052] Among them: U pji (k) represents the capacitor voltage value of the i-th submodule of the j-th bridge arm of the p-phase at the k-th moment; I pj (k) represents the current value of the j-th bridge arm of the p-phase at the time k; S pji g represents the switch operation state of the ith submodule of the jth bridge arm of phase p at time k+1-g, where '1' means the submodule is switched on and '0' means the submodule is switched off; C SM is the capacitance value of the submodule, Ts is the sampling period, G pji The reconstructed eigenvalues ​​representing the i-th submodule of the j-th bridge arm of the p-th phase are the preliminary eigenvalues.

[0053] Furthermore, the encoding rules of the one-hot encoding are as follows:

[0054] When the modular multilevel converter is in a normal state, the encoding rule of one-hot encoding is 0 0...0 1;

[0055] When the modular multilevel converter is in a fault state, the encoding rules of one-hot encoding are as follows;

[0056] If the fault of the modular multilevel converter occurs in the i(1, 2....., M-1, M)th submodule of the j(1, 2)th bridge arm of the p(1, 2, 3)th phase, then the (M*(p-1)+M*(j-1)+i)th bit from the highest bit is '1' and the rest of the bits are '0'; where p(1, 2, 3) represents the three phases A, B, and C respectively, j(1, 2) represents the upper and lower bridge arms respectively, and i(1, 2....., M-1, M) represents the submodule number of a single bridge arm respectively.

[0057] If the input vector dimension is a 6*3 vector, M represents the number of bridge arm submodules, and the output is a 6*3+1 vector:

[0058] The encoding rules are as follows:

[0059] The normal state is encoded as 0 0...0 1, that is, the first 6*3 items are 0 and the last item is 1.

[0060] The fault status is encoded as follows: if the fault occurs in the i(1,2,3)th submodule of the j(1,2)th bridge arm of the p(1,2,3)th phase, then the (3*(p-1)+3*(j-1)+i)th bit from the highest bit is '1' and the rest of the bits are '0';

[0061] Furthermore, the test sample set is input into the trained diagnostic model, and the probability values ​​belonging to each state of the transformer are output by the Softmax classifier. The class label matching the maximum probability is selected to obtain the fault type. The Softmax classifier function is:

[0062]

[0063] In the formula, y i is the probability value corresponding to the i-th state; the number of output units is 7, corresponding to each state of the current transformer.

[0064] The present invention can effectively extract the deep features of the data and effectively solve the problem of low data feature differentiation. Compared with traditional fault diagnosis methods such as Kalman filtering, sliding mode observer, tensor machine, and neural network, the present invention can locate the specific submodule of the fault point in one step, without the need for a fault detection stage, and without adding new hardware structures. The detection accuracy is significantly improved, and the detection speed can also be significantly improved.

Claims

1. A medium voltage modular multilevel converter fault diagnosis method, characterized in that: The steps include: S1: collecting the capacitor voltage value of each submodule and the current value of each bridge arm of the modular multilevel converter once every N sampling periods, and recording the switch operation status data of each submodule of the modular multilevel converter MMC in each sampling period; S2: constructing the collected capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multilevel converter MMC into a data set, wherein the data set includes normal operating state data and fault operating state data of the submodules of the modular multilevel converter MMC; S3: constructing a data set of capacitor voltage values ​​of each submodule and current values ​​of each bridge arm of the modular multilevel converter MMC to perform feature reconstruction and sample screening to obtain new feature values, and forming a sample set with the new feature values, dividing the sample set into a training set and a test set, and performing one-hot encoding on the sample set according to the operating state of the modular multilevel converter to obtain an encoding vector as a class label; S4: using a neural network and performing parameter training on the training set based on a BP back propagation algorithm to obtain a modular multilevel converter fault diagnosis model; S5: input the test set into the MMC fault diagnosis model to perform open circuit fault diagnosis on each submodule of the modular multilevel converter, and obtain the fault type by matching with the class label; The step of constructing a data set from the capacitor voltage values ​​of each submodule and the current values ​​of each bridge arm of the modular multilevel converter MMC to perform feature reconstruction and sample screening to obtain a new feature value comprises the following steps: S3-1: construct a preliminary sample set according to the preliminary sample set formula using the collected submodule capacitor voltage, each bridge arm current and recorded switch state data; S3-2: Filter samples according to the fault type. If an open circuit fault occurs in the switch device connected in series with the capacitor, only samples when the input current of the faulty submodule is less than zero are retained; if an open circuit fault occurs in the switch device connected in parallel with the capacitor, only samples when the input current of the faulty submodule is greater than zero are retained; thereby obtaining a new feature value; The preliminary sample set formula is as follows: Among them: U pji (k) represents the capacitor voltage value of the i-th submodule of the j-th bridge arm of the p-phase at the k-th moment; I pj (k) represents the current value of the j-th bridge arm of the p-phase at the time k; S pji g represents the switch operation state of the ith submodule of the jth bridge arm of phase p at time k+1-g, where '1' means the submodule is switched on and '0' means the submodule is switched off; C SM is the capacitance value of the submodule, Ts is the sampling period, G pji The reconstructed eigenvalues ​​representing the i-th submodule of the j-th bridge arm of the p-th phase are the preliminary eigenvalues.

2. A medium voltage modular multilevel converter fault diagnosis method according to claim 1, characterized in that: The encoding rules of the one-hot encoding are as follows: When the modular multilevel converter is in a normal state, the encoding rule of one-hot encoding is 0 0...0 1; When the modular multilevel converter is in a fault state, the encoding rules of one-hot encoding are as follows; If the fault of the modular multilevel converter occurs in the i(1, 2....., M-1, M)th submodule of the j(1, 2)th bridge arm of the p(1, 2, 3)th phase, then the (M*(p-1)+M*(j-1)+i)th bit from the highest bit is '1' and the rest of the bits are '0'; where p(1, 2, 3) represents the three phases A, B, and C respectively, j(1, 2) represents the upper and lower bridge arms respectively, and i(1, 2....., M-1, M) represents the submodule number of a single bridge arm respectively.

3. A medium voltage modular multilevel converter fault diagnosis method according to claim 1, characterized in that: The neural network activation function is a sigmoid function, whose input is a 6M-dimensional vector, representing the reconstructed features of 6M sub-modules respectively. The hidden layer has 10 units, and the output layer has 6M+1 units, representing 1 normal working state plus 6M fault locations respectively.

4. A medium voltage modular multilevel converter fault diagnosis method according to claim 1, characterized in that: The MMC fault diagnosis model includes a Softmax classifier; the Softmax classifier outputs probability values ​​belonging to various states of the converter, selects a class label that matches the maximum probability, and obtains the fault type.

5. A method for fault diagnosis of a medium voltage modular multilevel converter according to claim 4, characterized in that: The Softmax classifier formula is: In the formula, y i is the probability value corresponding to the i-th state; m is the number of output neurons, that is, the corresponding states of the switch.

Images