NL2020015A - Fault diagnosis method of three-level inverter based on empirical mode decomposition and decision tree RVM - Google Patents

Abstract

Disclosed is a fault diagnosis method of a three-level inverter based on an empirical mode decomposition and a decision tree RVM, in view of fault diagnosis problems of a diode neutral-point-clamped three-level inverter in a photovoltaic power generation system, first of all, analysing operating conditions of an inverter main circuit and performing fault classification; extracting each signal component with the empirical mode decomposition method by taking the middle, upper and lower bridge leg voltages as measurement signals; calculating corresponding parameters, such as an energy and an energy entropy; thereby generating a decision tree RVM classification model with a particle swarm clustering algorithm, to finally achieve multi-mode fault diagnosis of the photovoltaic diode neutral-pointclamped three-level inverter. Advantages of the present invention are that, setting of parameters is not needed, the number of classification models is small, both a calculating efficiency and a diagnosis precision are high, and the robustness is good.

Description

Patent center

The Netherlands © 2020015 (21) Application number: 2020015 © Application submitted: 04/12/2017

A PATENT APPLICATION @ Int. CL:

G01R 31/42 (2018.01)

© Priority: 26/12/2016 CN 201611216599.8 © Applicant (s): JIANGNAN UNIVERSITY in Jiangsu Wuxi, China, CN. © Application registered: 02/07/2018 © Inventor (s): Hongfeng Tao in Wuxi (CN). © Request published: Chaochao Zhou in Wuxi (CN). 04/07/2018 Yajun Tong in Wuxi (CN). Yan Liu of Wuxi (CN). Jianqiang Shen in Wuxi (CN). © Authorized representative: ir. H.Th. van den Heuvel et al. in 's-Hertogenbosch.

(54) Fault diagnosis method or three-level inverter based on empirical mode decomposition and decision tree RVM (57) Disclosed is a fault diagnosis method or three-level inverter based on empirical mode decomposition and decision tree RVM, in view of fault diagnosis problems or a diode neutral-point-clamped three-level inverter in a photovoltaic power generation system, first or all, analyzing operating conditions or an inverter main circuit and performing fault classification; extracting each signal component with the empirical mode decomposition method by taking the middle, upper and lower bridge leg voltages as measurement signals; calculating corresponding parameters, such as an energy and an energy entropy; continuously generating a decision tree RVM classification model with a particle swarm clustering algorithm, to finally achieve multi-mode fault diagnosis of the photovoltaic diode neutral-pointclamped three-level inverter. Advantages of the present invention are that, setting of parameters is not needed, the number of classification models is small, both a calculating efficiency and a diagnosis precision are high, and robustness is good.

NLA 2020015

This publication corresponds to the documents originally submitted.

Fault diagnosis method or three-level inverter based on empirical mode decomposition and decision tree RVM

The present invention relates to the field of fault diagnosis or power electronic 5 devices, and in particular, to a fault diagnosis method or a diode neutral-pointclamped three-level inverter based on an empirical mode decomposition and a decision tree Relevance Vector Machine (RVM ).

As the photovoltaic power generation technology advances and the grid-connected operation scale of the photovoltaic power generation increases, the development of the photovoltaic power generation industry has been seriously restricted by problems, such as optimization, improvement, and operating costs of the photovoltaic power generation system. In particular, the photovoltaic inverter is low in cost, but the inverter is always a weak link which easily breaks down in the whole system, for various reasons, example, the power electronic device itself used by the inverter main circuit is fragile, the inverter main circuit is complex in control and frequent in on-off control, and the external environment is severe. The inverter is susceptible to faults, such as overvoltage, overcurrent, short and open faults or power tubes, and all these conditions are closely related to safe operations of the whole photovoltaic power generation system. In order to prevent more serious accidents caused by faults, it is necessary to timely detect the failed device and determine reasons and positions of the faults, and such doing not only facilitates to reduce economic losses, but is also useful for work development or maintenance personnel . Stable, efficient and safe operation of the photovoltaic power generation system also can be achieved, which has a significant meaning for promoting the large-scale development of the photovoltaic power generation in China.

As different types and structures or inverters are gradually applied in the photovoltaic power generation system, reliability, stability and maintainability or the inverter has become increasingly important. Data show that 38% of all the grid connected inverter faults derive from damage or power tubes in the inverter main circuit. Common inverter faults mainly include short and open faults, if a short fault occurs, a hardware circuit usually performs protective processing within a time or microsecond order; however, the short fault will not cause the system to shut down immediately in most cases, but will cause secondary faults or other devices, thus finally making the system out of operation. When a fault occurs in an inverter, physical quantities, such as the voltage, the current in the circuit will differ from those in a normal state, therefore, based on different detection signals, diagnosis methods for open faults or an inverter power tube can be classified into current and voltage fault diagnosis methods. No additional sensors are needed in a current fault diagnosis method, however, in most cases, the current is relevant to an ioad, under ano load or light load condition, the current method has a very low diagnosis precision. The voltage method is to perform fault diagnosis by investigating deviation of a phase voltage, a line voltage or a bridge lay voltage or the inverter from those in a normal state, an additional sensor is needed, and such has many advantages, such as, better robustness to noise and load, lower false positive rate and less diagnosis time.

In the fault diagnosis of power electronic devices, selecting and extracting fault feature vectors always have critical factors in the diagnosis, which directly influence the accuracy of the diagnosis result. There are many switching devices for the photovoltaic three-level inverter, the types of fault problems are complicated, and a large number of measured signals are non-static signals, as a consequence, it is necessary to adopt a feature extraction method suitable for processing non3 static signals during fault diagnosis, and an empirical mode decomposition method is just such a method.

On the other hand, it is another critical step or fault diagnosis to design a classifier with a reasonable structure to perform state recognition. At present, the pattern recognition method for fault detection and diagnosis mainly comprises statistical pattern recognition and neural network recognition, meanwhile, intelligent diagnosis algorithms such as an extreme learning machine and a support vector machine also show great application potential. However, conventional statistical pattern recognition methods have their own limitations, many important issues of the neural network technology have not yet been solved theoretically, a large number of samples need to be trained by the extreme learning machine, although the support vector machine is suitable to solve small-sample, non-linear and high-dimension pattern recognition, various parameters still need to be selected empirically, and parameters such as a penalty coefficient and a kernel function radius have a great influence on the diagnosis precision. A relevant vector machine (RVM) is a learning machine constructed based on a Bayesian framework, does not need to set penalty factors, overlearning due to improper parameter settings just like the support vector machine will not occur, and this algorithm is also capable of solving high dimension, non-linear and small-sample pattern recognition problems, and thus has a good application prospect.

One object of the present invention is to provide a fault diagnosis method or a diode neutral-point-clamped three-level inverter based on an empirical mode decomposition and a decision tree RVM.

The present invention provides hereto a fault diagnosis method or a neutral-clamped three-level inverter based on an empirical mode decomposition and a decision tree RVM, characterized by including the steps of: 1) establishing a neutral-point-clamped three-diode level inverter main circuit model and performing fault classification; 2) extracting a three-level inverter main circuit open circuit fault feature vector; 3) constructing a three-level inverter fault diagnosis decision tree; 4) constructing a relevant vector machine RVM fault classification decision tree model, to finally achieve fault diagnosis or a photovoltaic diode neutral-point-clamped three-level inverter.

Step 1): establish a photovoltaic diode neutral-point-clamped three-level inverter main circuit model and perform fault classification.

The three-level inverter main circuit comprises three-phase bridge legs and consists of two clamping capacitors, twelve main switch tubes, twelve free-wheel diodes and six neutral point clamping diodes, the three-level inverter main circuit has two significant features: a ) an output voltage waveform is synthesized by a variety of levels, the harmonic content is greatly reduced and the output voltage waveform is improved, when compared to a conventional two-level mode; and b) a rated voltage value of the switch tube is half the voltage of the direct current bus, making it possible to apply a low-voltage switch tube in a high-voltage converter.

Three-phase of the photovoltaic diode neutral-point-clamped three-level inverter main circuit are symmetric, and therefore, if A-phase is tasks as an example, other phases are similar; the mainly-discussed open circuit faults of the three-level inverter include an IGBT open circuit, a fusing or fusing in series and a trigger missing pulse fault, and also an open circuit fault or a neutral point clamping diode, faults are classified into four categories and thirteen subcategories, which comprises:

1) The system is free of faults, and there is one subcategory in total.

2) A single clamping diode is open-circuited, and there are two subcategories in total.

3) A single power device is open-circuited, and any of four power tubes is open-circuited, and there are four subcategories in total.

4) Two devices is open-circuited, there are two cases: first, the two open-circuited power tubes are not located at the same bridge lay, and the two open-circuited power tubes being not located at the same bridge falls into the fault of a single device on different bridge legs, and is referred to the third fault classification that a single power device is open-circuited; second, the two failed switch tubes are located at the same bridge lay, any two of the four switch tubes fail, and there are six subcategories in total.

Step 2): extract a three-level inverter main circuit open circuit fault feature.

In the process of signal analysis, a time scale and an energy distributed over the time scale are two most important parameters of the signal. Comparing a voltage signal or an open-circuited inverter main circuit power tube with a voltage signal or a normal system, energies or signals in the same frequency band are greatly different. The energy of each frequency component or the signal contains abundant fault information, if energy or one or more frequency components changes, then there is a fault, and therefore the fault analysis is performed according to the change in energy or each frequency band.

Model a diode neutral-point-clamped three-level inverter main circuit under space vector pulse width modulation SVPWM control and neutral point potential control, after modeling, perform EMD decomposition on the bridge lay voltage when various faults occur, select the first n IMF components and residual components, and then calculate energy or each IMF component and residual component. Let Ej be the energy of each component ί ·. · = Σ / · £ = 1 i, / c (1) ·

In the formula, q k (/ = 1.2, ..., n + 1; fc = 1, 2, ..., J) are magnitudes for J discrete points of the first n IMF components and residual components. A feature vector is constructed after receiving energy or each bridge leg voltage, in particular, the feature vector is 7j:

T, = XE (.E, ... E ,, Ί ¹ L θ ¹ «+ 1J (2).

Considering that values for energy is big in most cases and for ease of subsequent classification, modify normalization processing +1 9 ⁷

ΣΙ ^ Ι '' (3).

Meanwhile, calculate corresponding IMF energy entropy, on the basis of each IMF energy (4).

In the formula, pi = Ej / _Ez , a percentage of the / -th component within the overall is 20 signal energy + 1 (5).

Considering all the above parameters, and define the fault feature vector as:

Γ, · = [£ "/ £ £, / £ ... £ ,," / £ W,] (6).

Process the upper and lower bridge legs with the same method again to obtain feature vectors T ₂ 'and T ₃ ' respectively, and define the fault feature as:

T = \ T (T2 7V1

L ¹ 2 3j

Extract features of the bridge leg voltage under various faults following the processes described above, and finally construct a data sample.

Step 3): construct a particle swarm clustering fault diagnosis decision tree.

As described above, there are thirteen fault types in total for the three-level inverter constantly classify faults into two categories with a particle swarm clustering algorithm, until the subcategory only comprises one type of samples, under the condition that a decision tree is to be constructed, which particularly comprises: i) first, process of an initial category, all the training samples are tasks as the initial category and classified into two subcategories with the clustering algorithm; and ii) second, judge a subcategory and end the algorithm under the condition that the subcategory only comprises one type of samples, otherwise, continue to perform fault classification with the clustering algorithm until all the subcategories only comprise one type of samples.

The key of constructing the decision tree is to select the clustering algorithm, and the particle swarm clustering algorithm is used here; the particle swarm clustering algorithm comprises initializing the particle swarm clustering algorithm first, randomly initializing the particle swarm, set relevant parameters, and then randomly classify each sample, calculate parameters including an adaptability and a clustering center, and set the initial velocity of the particle zero . In this way, a particle individual optimal positioning pid and a globally optimal positioning _g p d can be obtained based on the initial particle swarm. Determine clustering partition of each sample according to clustering center coding of the particles and the nearest neighbor rule and then calculate a new clustering center according to the new clustering partition, update the adaptability, and compare the adaptability again, update the pid under the condition that the adaptability is better than the individual optimal position pta; update the p ₉ d under the condition that the adaptability is better than the globally optimal position p ₉ d, end the algorithm under the condition that the maximum number of iterations is reached, otherwise, continuous the iteration.

As such, summarize clustering results to construct the structure of a fault diagnosis decision tree and provide basis for the subsequent RVM training objects.

Step 4): construct a relevant vector machine RVM fault classification decision tree model.

Classify the data samples into training sets and testing sets in a proportion of 3: 7, and train the training sets according to the decision tree structure obtained by the previous step. After training is completed, test with the testing sets to obtain indexes including precision diagnosis, an average training time and an average testing time, to finally achieve fault diagnosis or a photovoltaic diode neutral-pointclamped three-level inverter.

The beneficial effects of the present invention are as follows:

the fault diagnosis method or three-level inverter based on an empirical mode decomposition and a decision tree RVM according to the present invention is based on a data-driven idea, to combine an empirical mode decomposition, a particle swarm clustering algorithm and a relevant vector machine algorithm, to achieve fault diagnosis of a photovoltaic inverter, especially a photovoltaic diode neutral-point-clamped three-level inverter.

The present invention adopts an empirical mode decomposition EMD algorithm to extract features. This EMD algorithm is an adaptive algorithm and is very suitable for analyzing non-stationary and non-linear signals, and meanwhile, is to characterize fault information by taking the energy of each IMF component and energy entropy or signals as fault feature vectors, without needing to select parameter values empirically just like wavelets analysis.

The present invention adopts a decision tree RVM fault diagnosis model structure, and only fewer classification models are required to be constructed to accomplish the fault diagnosis task for the decision tree structure. At the same time, the RVM algorithm has many advantages about the SVM algorithm: vectors to be used are fewer, the testing time is shorter, the sparsity is stronger, and in terms of classification with fewer training samples and fewer features, the robustness is better, and setting or parameters is not needed.

The present invention will be further elucidated on the basis of the non-limitative example shown in the following figures. Herein shows:

figure 1 a fault diagnosis flow or a diode neutral-point-clamped three-level inverter; figure 2 a topological structure or a diode neutral-point-clamped three-level inverter main circuit;

figure 3 an A-phase topology or an inverter main circuit;

figure 4 a bridge leg voltage when a single device fails;

figure 5 a bridge leg voltage when two devices are open-circuited simultaneously; figure 6 a bridge leg voltage when a single clamping diode is open-circuited; figure 7 an EMD decomposition result when the inverter is normal; figure 8 a fault feature vector histogram when the inverter is normal; and figure 9 a structural diagram of a decision tree after clustering classification.

The fault diagnosis flow chart of a three-level inverter based on an empirical mode decomposition and a decision tree RVM according to the present invention is as shown in figure 1, and the specific method of the present invention include steps as follows:

In figure 2 a topological structural diagram or a diode neutral-point-clamped threat level inverter main circuit is shown, and in order to simplify the analysis, only the working state of A-phase in the inverting state of the inverter has been studied, and the circuit topology of the inverter is shown in figure 3. In the figure, the solid line is the positive direction of the current, the dashed line is the negative direction of the current, after ignoring a conductor voltage drop of power devices, the potential at point A is always equal to the potential at point P in the P state, the potential at point A is always equal to the potential at point O in the O state, and the potential at point A is always equal to the potential at point N in the N state.

Based on the topological structure, faults can be classified into four categories and thirteen subcategories, i.e., the fault classification of the diode neutral-point20 clamped three-level inverter:

1) The inverter main circuit is free of faults, power devices operate normally, and there is one subcategory in total.

2) A single clamping diode, i.e., any of VDa5 and VDa6 is open-circuited, and there are two subcategories in total.

3) A single device, i.e., any of the power tubes Sa1, Sa2, Sa3, Sa4 is open-circuited, and there are four subcategories in total.

Two devices are open-circuited, there are two subcategories for this category: first, the two open-circuited power tubes are not located at the same bridge lay, and such can be referred to the third open-circuited case without taking into account the fault classification; second, the two open-circuited power tubes are located at the same bridge lay, that is, any group of the power tubes (Sa1, Sa2), (Sa1, Sa3), (Sa1, Sa4), (Sa2, Sa3), (Sa2, Sa4) or (Sa3, Sa4) are open-circuited, and there are six subcategories in total. In view of the above, fault classification and corresponding labels are shown in Table 1 below.

Table 1 Fault Classification

Label Fault Type Label Fault Type 0 Free of faults 12 Sai and Sa2 1 Sa1 13 Sa1 and Sa3 2 Sa2 14 Sai and Sa4 3 Sa3 23 Sa2 and Sa3 4 Sa4 24 Sa2 and Sa4 5 VDaS 34 Sa3 and Sa4 6 VDaS

Establish a diode neutral-point-clamped three-phase three-level inverter main circuit model, control working states of three phases in the inverter by cooperating SVPWM control with neutral point potential control technology, to drive the threat level inverter to complete inverting work. Take the bridge leg voltage as an object to study, and then bridge leg voltages in a case of various faults can be obtained, as shown in figure 4 and figure 5. It can be found by comparing figure 4 (c) and figure 5 ( a), Sa2 and (S _a i, S _a 2) have the same level logics, which is caused by the structure of the circuit itself, then it is necessary to introduce a new measured point, ie, an upper bridge leg voltage, as shown in figure 6. Perform an EMD decomposition on each bridge leg voltage, respectively, to decompose each bridge leg voltage into four IMF components and one residual component, and the EMD decomposition result of the bridge leg voltage is normally as shown in figure 7 Calculate energy of the signal after decomposition, calculate energy entropy after unifying dimension, and finally construct a fault feature vector or a single bridge leg voltage. Integrate the single fault feature vector, and construct overall fault feature vectors in a middle-upper-lower sequence, and construct data samples according to different fault types. The fault feature vector histogram when the inverter operates normally as shown in Figure 8.

As described above, fault samples are classified with the particle swarm clustering algorithm, for example, the first classification result is that: data samples with label 0, 1.4, 5, 6 and 14 are classified as one category; data samples with label 2, 3, 12, 13, 23, 24 and 34 are classified as another category. In this way, the first layer structure of the decision tree and training samples or corresponding classification model RVM1 are also determined, and so forth. After the classification is complete, the decision tree is constructed with a final result shown in figure 9. As can be seen in the figure, only ^ classification models need to be constructed for 13 fault classification problems when a decision tree structure is adopted, however , if a one-to-one structure has been adopted, it is necessary to construct 78 classification models. Meanwhile, in terms of test models, only 2 to 6 rounds of classification operations need to be performed when adopting a decision tree structure, while a one-to-one structure still needs to perform 78 rounds of classification operations. In conclusion, adopting the decision tree structure, undoubtedly can greatly reduce the number of models to be constructed, reduce the operation time, and improve the operational efficiency.

Data samples are classified into training sets and testing sets in a proportion of 3: 7. 12 relevant vector machine classification models, i.e., RVM1 to RVM12 are trained, respectively, according to the constructed decision tree structure. In order to verify the anti-interference capability of the algorithm, white noises with 10% and 15% signal amplitudes are added to the original data for comparison, and meanwhile, the training time, the testing time and the diagnosis precision of a BP neural network (back propagation neural network, BPNN), an extreme learning machine (extreme learning machine, ELM), a relevant vector machine with a one-to-one structure (1 vs. 1) and a decision tree support vector machine (DT- SVM) are transversely compared, and the final fault diagnosis results are summarized in Table 2 and Table 3.

Table 2 Fault Diagnosis Result (10% White Noise)

Average Training Time (s) Average Testing Time (s) Precision Diagnosis (correct samples / total samples) BPNN 13.327376 0.180402 898/910 ELM 0.032876 0.021062 895/910 (1-v-1) RVM 0.543930 0.548473 905/910 DT-SVM 0.026909 0.103461 895/910 DT-RVM 0.265664 0.039128 905/910

Table 3 Fault Diagnosis Result (15% White Noise)

Average Training Time (s) Average Testing Time (s) Precision Diagnosis (correct samples / total samples) BPNN 13,476147 0.193747 885/910 ELM 0.040628 0.023154 880/910 (1-v-1) RVM 0.525205 0.548704 884/910 DT-SVM 0.035681 0.085105 887/910 DT-RVM 0.276227 0.040719 889/910

The described above are merely examples for clearly illustrating the present invention, and in no way should be constructed to limit the present of the present invention. Persons skilled in the art should appreciate that various other variations and modifications can be made on the basis of the above description.

Claims (5)

Conclusions An error diagnosis method for a three-level diode converter with clamped star point based on an empirical mode decomposition and decision tree RVM, characterized in that the method comprises the steps of: 1) establishing a main circuit model for a three-level diode converter with clamped star point and performing an error classification; 2) extracting an open circuit error characteristic vector from the three-level converter main circuit; 3) constructing a three level inverter error diagnosis decision tree; and 4) constructing an error classification decision tree model for a relevant vector machine RVM, to finally obtain an error diagnosis of a photovoltaic diode three-level inverter with clamped star point. An error diagnosis method according to claim 1, characterized in that the three-level converter main circuit from step 1) comprises three-phase bridge legs and consists of two terminal capacitors, twelve main switch tubes, twelve freewheel diodes and six star point terminal diodes. An error diagnosis method according to claim 1 or 2, characterized in that the three-level converter main circuit from step 1) is designed to provide an output voltage waveform synthesized by a plurality of levels, wherein the harmonic content is considerably reduced and the output voltage waveform is improved, compared with a conventional two-level mode and / or to provide a nominal voltage value of the switch tube with half the voltage of the DC bus. An error diagnosis method according to any one of the preceding claims, characterized in that during step 2) of extracting an open circuit error characteristic of a three-level converter main circuit a voltage signal from a power tube of an open circuit converter main circuit is compared with a voltage signal of a normal system. An error diagnosis method according to any of the preceding claims, characterized in that the method comprises modeling a three-level diode converter main circuit with clamped star point under space vector pulse width modulation SVPWM control and star point potential control, and after modeling, performing EMD decomposition on the bridge leg voltage when various errors occur, selecting the first n IMF components and remaining components, and then calculating the energy of each IMF component and remaining component; let Ede be energy of every component, 1 I> 2 ^£ . = ΣΜ '-1 (1) in the formula are c, y (/ = 1.2, ..., n + 1; fc = 1.2, sizes for Jdiscrete points of the first n IMF components and remaining components: a feature vector is constructed after obtaining energy from each bridge leg voltage, the feature vector being Tj: ^ ί ⁼ [^ Ό ··· ^ n + l] (2) considering that energy values are high in most cases and for ease of subsequent classification, modifying normalization operations, (/ 7 + 1 π V Ηςκ ^{U = 1 7} (3) meanwhile, calculating corresponding IMF energy entropy, based on each IMF energy, «4-1 = - Σ aiga (4) in the formula is p, · = E / / E _{z is} a percentage of the / -th component in the total signal energy, M + 1 Α = Σ ^£ . (5) all the above parameters considered, and the error characteristic vector determining as: Ej / E ... Ε " _{+ ί} / Ε Η,] _(θ) reprocessing the upper and lower bridge legs with the same method to obtain feature vectors 7y and Ts, respectively, and determining the error characteristic as: t = \ t; t / tS] LI 2 3J extracting bridge leg voltage characteristics at different errors according to the methods described above, and finally constructing a data sample. An error diagnosis method according to any one of the preceding claims, characterized in that the error diagnosis decision tree from step 3) comprises thirteen error types in total for the three-level converter, errors being continuously classified into two categories with a particle swarm cluster algorithm, until the subcategory only one type of samples, provided that a decision tree is constructed. An error diagnosis method according to claim 6, characterized in that the classification of errors comprises the steps of: i) firstly, editing a starting category, where all training samples are taken as the starting category and classified into two subcategories with the cluster algorithm; and ii) second, reviewing a subcategory and terminating the algorithm provided that the subcategory includes only one type of samples, and otherwise continuing to perform the error assay with the cluster algorithm until all subcategories include only one type of samples wherein, while constructing the decision tree, the cluster algorithm and the particle swarm cluster algorithm to be used are selected, wherein the particle swarm cluster algorithm comprises first initializing the particle swarm cluster algorithm, randomizing the particle swarm, setting relevant parameters, and then randomly classifying each sample, calculating an adaptability and a cluster center, and setting the initial velocity of the particle to zero; in which a separate optimal position pavoor the particle, and a global optimal position p _g aop the basis of the initial particle swarm is obtained; determining cluster distribution of each sample according to cluster center coding of the particles and the neighbor neighbor rule and then calculating a new cluster center according to the new cluster distribution, updating the adaptability, and comparing the adaptability again, updating the p « / provided that the adaptability is better than the individual optimum position p ^; updating the p _g d provided that the adaptability is better than the globally optimal position p _g d, terminating the algorithm provided that the maximum number of repetitions is reached, and otherwise continuing the repetition; and summarizing cluster results to construct the structure of a tout diagnosis decision tree and to provide a basis for the subsequent RVM training objects. An error diagnosis method according to any one of the preceding claims, characterized in that step 4) comprises the step of constructing an RVM error classification decision tree model; classifying the data samples in training sets and test sets in a ratio of 3: 7, training the training sets according to the decision tree structure obtained by the previous step; after the training is completed, test with the test sets to obtain diagnostic accuracy, an average training time and an average test time. Fault classification Extract mnltiple groups: ofeach bridge leg voltage signal Analysis \ oltage signals with EMO decomposition method Obtain fault feature vectors Geatenatea decisioniee with a cluster aigouthm Train classification tnode ls o f ea ch reievaut vector machine T cat Guit * ~ 88 * fashion model Obtain classific: ation results VD.i U / 2 - 1 a _v , L. 1 ivü _a2 VD _a -8 * · VLN U / 2. 3.J VD _a4 A -: <A a: P state b: O state c: N state a: Free of Faults b: Sai Open Circuit c: Sa2Open Circuit d: VDas Open Circuit a: Sai and S _a 2 Open Circuit b: Sai and Sa3 Open Circuit L7V 400 ' ^r - ^{1 1} '' ¹ 200 i ⁰ i f -200 -400 __________ 0.02 0.04 0.06 f c: Sai and S _a 4 Open Circuit d: Sa2 and S ₃ 3 Open Circuit fT / V y / y 400 100 ^L 250 - 400 f If ........... F \. f w f \ (1 II λ: \ j \ / \ I Like this \! 250 - \: \; ί / s ion r / s 0.06) 0.02 0.04 i: Sa2 Open Circuit 0.06 0.02 0.04 b: Sai and S _a 2 Open Circuit -50C Sampling point 400 50C 400 -50C 100 200 300 350 100 150 200 250 Satnping point 300 350 400 200 - 1 LL -200 -1-1-Γ " X '\ ZV · ^ · - · --______ J______________________1______________________L. 100 ~ C _____L „L. 500 200 250 300 Samping pcint 350 400 _____L_____ 200 _____i_____________________J „„ 250 300 350 400 -500 100 Sa * ïïfijij pom: RVM1