A three-level inverter open-circuit fault diagnosis method and system

Abstract

The application discloses a three-level inverter open-circuit fault diagnosis method and system, realizes inverter open-circuit fault diagnosis based on the characteristics of output three-phase phase voltage and terminal voltage; an NPC three-level inverter simulation model is built to simulate various fault conditions; after collecting terminal voltage and carrying out normalization processing, frequency spectrum is obtained through fast Fourier transform (FFT); after collecting three-phase phase voltage and carrying out coordinate transformation and normalization processing, ensemble empirical mode decomposition (EEMD) is carried out on the three-phase phase voltage to obtain intrinsic mode components (IMF), and Hilbert-Huang transform is carried out on each IMF to obtain a marginal spectrum; meanwhile, a voltage positive half cycle proportion coefficient is added as a characteristic vector to carry out layer-by-layer pre-training of an extreme learning machine automatic encoder (ELM-AE), and an output weight value is calculated by using a popular regularization framework; finally, an extreme learning machine is used in a top classification layer, a Gaussian kernel function is introduced, and a grey wolf optimization (AO) algorithm is used to optimize relevant parameters to complete training of a DKELM network. The application has fast diagnosis speed and can obviously improve fault diagnosis accuracy.

Description

Technical Field

[0001] This invention belongs to the field of power electronic circuit fault diagnosis technology, specifically relating to a method and system for diagnosing open-circuit faults in a three-level inverter. Background Technology

[0002] With the depletion of traditional fossil fuels and the increasing demand for electricity globally, grid-connected power generation from renewable energy sources such as solar and wind power has become a development trend in new power systems. Inverters, widely used in these power generation systems, are power electronic devices responsible for power conversion. The failure of an inverter directly affects the normal operation of power conversion and even the safe and reliable operation of the entire power grid. With the increasing demand for high-voltage, high-power power generation, the application of multi-level inverters, especially three-level inverters, is becoming more widespread. They were developed to overcome the shortcomings of traditional two-level inverters, such as high switching stress and low DC bus voltage utilization. Three-level inverters offer advantages such as low power switching voltage stress, low output voltage harmonic content, high DC bus voltage utilization, low electromagnetic interference, and low switching losses. Therefore, in addition to photovoltaic and wind power generation, they are also widely used in various fields such as military and aerospace.

[0003] However, the increase in the number of inverter levels also leads to an increase in the number of switching transistors, making accurate control of these transistors more complex and increasing the probability of power switch failures. Therefore, it is crucial to employ effective methods for accurate and rapid diagnosis and localization of switching transistor faults in three-level inverters. Switching transistor faults in NPC three-level inverters are classified into open-circuit faults and short-circuit faults. When a short-circuit fault occurs, the designed hardware protection circuit will use its own fuse to convert it into an open-circuit fault. When an open-circuit fault occurs in a power device, it is not easily detected in its early stages, but within a very short time, the output waveform will undergo significant distortion, leading to DC-side voltage imbalance. In severe cases, this may cause the system to stop working or even significantly impact the entire system. Among open-circuit fault types, the probability of three or more switching transistors simultaneously experiencing open-circuit faults is very small; therefore, this invention only studies single-transistor open-circuit faults and double-transistor open-circuit faults.

[0004] Current methods for diagnosing open-circuit faults in inverters can be summarized as follows: ① Expert system-based methods: These methods rely on accumulated expert experience to establish a knowledge base. When a fault occurs, only the fault phenomenon needs to be observed, and the knowledge base can be consulted to determine the fault type. The difficulty lies in the inability to exhaustively cover all fault phenomena. ② Analytical model-based methods: These methods require the establishment of an accurate mathematical model. However, the entire system of an NPC three-level inverter is a very complex nonlinear system, making it difficult to establish a mathematical model, thus limiting the diagnostic capabilities. ③ Intelligent algorithm-based methods: These methods select signals that characterize fault features, process them, and then use intelligent algorithms for learning or reasoning. This invention adopts the third method. To avoid the influence of load disturbances, voltage signals are processed, and deep learning methods are combined for fault diagnosis. Summary of the Invention

[0005] In order to solve the technical problems existing in the background art, the present invention aims to provide a method and system for diagnosing open circuit faults in three-level inverters, based on DKELM's NPC, so as to improve the accuracy of fault diagnosis.

[0006] To solve the technical problem, the technical solution of the present invention is as follows:

[0007] A method for diagnosing open-circuit faults in a three-level inverter, the method comprising:

[0008] S1: Build an NPC three-level inverter simulation model to simulate various fault conditions and collect multiple sets of three-phase phase voltages and terminal voltages, i.e., multiple sets of fault signals.

[0009] S2: Preprocess the multiple sets of fault signals collected in step S1 to obtain multiple sets of preprocessed fault signals;

[0010] S3: Perform ensemble empirical mode decomposition (EEMD) on the preprocessed phase voltage fault signal and obtain the marginal spectrum through Hilbert transform. At the same time, obtain the spectrum by fast Fourier transform (FFT) of the preprocessed terminal voltage. Combine the marginal spectrum, the spectrum, and the voltage positive half-cycle scaling factor of the phase voltage as the signal feature vector.

[0011] S4: Divide the signal feature vector into labeled training samples, unlabeled training samples and test samples. The labeled training samples and unlabeled training samples form a total training sample matrix. At the same time, classify all fault types using binary labels.

[0012] S5: The deep kernel extreme learning machine is trained using training samples. The output weights of ELM-AE are calculated using a popular regularization framework instead of the least squares method. The classification of the top-level extreme learning machine introduces a Gaussian kernel function and optimizes it using the Eagle Optimization algorithm. The trained DKELM model is tested on test samples, and the fault diagnosis results are analyzed.

[0013] Furthermore, in step S1, the step of sampling the NPC three-level inverter within one sampling period includes:

[0014] S11: Open circuit faults in NPC three-level inverters are classified into 12 types of single-tube open circuit faults, 12 types of dual-tube faults in the same phase but different bridge arms, 48 ​​types of dual-tube faults in different phases, plus the fault-free case, for a total of 73 types.

[0015] S12: For each fault type, the three-phase phase voltage and terminal voltage of the NPC three-level inverter are sampled simultaneously to obtain the output three-phase phase voltage U. ih and terminal voltage U jh Where: i = a, b, c, j = ao, bo, co, h is the sample obtained in the h-th sampling, h = 1, 2, ..., N, N is the maximum number of samplings in a sampling period, and N is a positive integer.

[0016] Furthermore, in step S2, the specific preprocessing process includes:

[0017] S21: Convert the three-phase phase voltage U ih and terminal voltage U jh Obtained through Clark transformation and The formula is:

[0018]

[0019] S22: The normalized vector is obtained after normalization calculation. The formula is:

[0020]

[0021] Furthermore, step S3, which involves calculating the spectrum, marginal spectrum, and voltage positive half-cycle proportionality coefficient, includes:

[0022] S31: Obtain the spectrum of the terminal voltage by performing a Fast Fourier Transform on the normalized vector of the terminal voltage obtained in step S2. The formula is as follows:

[0023]

[0024] in The calculation is saved by utilizing the properties of the Discrete Fourier Transform;

[0025] S32: Obtain the marginal spectrum from the normalized vector of the phase voltage obtained in step S2, as follows:

[0026] S321: The normalized vector of the obtained phase voltage is used to decompose the signal into several intrinsic mode components and a residual signal using the signal extremum information. However, due to the problem of mode aliasing in Empirical Mode Decomposition (EMD), i.e., the same IMF contains signal information of different frequency components, or different IMFs contain signal information of the same component, random Gaussian white noise is added to the original signal to suppress mode aliasing through multiple experiments. The final IMF is the average of M experiments. After EMD, the original signal is divided into several empirical mode components and a residual signal.

[0027]

[0028] Each IMF is:

[0029]

[0030] In the formula d i,m This refers to the i-th eigenmode component in the m-th experiment;

[0031] S322: Applying a Hilbert transform to each of the eigenmodes obtained above yields:

[0032]

[0033] Amplitude function:

[0034]

[0035] Phase function:

[0036]

[0037] Instantaneous frequency:

[0038]

[0039] The calculated HHT spectrum:

[0040]

[0041] S323: Obtain the marginal spectrum based on the result of the Hilbert transform:

[0042]

[0043] With ω fixed, we integrate over t; the definite integral in discrete signals can be approximately decomposed into the sum of the areas of multiple rectangles; in discrete signals, H(ω,t) is the time-spectrum matrix H(ω,k), where the length of the rectangle is the H(ω,k) corresponding to the k-th data point, and the width is the time interval, i.e., Δt = 1 / f s Therefore, the integral formula can be changed to the following formula:

[0044]

[0045] S33: Add a voltage positive half-cycle proportionality coefficient to the characteristic signal, the formula is:

[0046]

[0047] Furthermore, feature vectors are constructed from the spectrum of the terminal voltage, the marginal spectrum of the phase voltage, and the positive half-cycle proportionality coefficient of the phase voltage obtained in step S3; multiple sets of fault feature vectors are divided into labeled training samples, unlabeled training samples, and test samples in a 5:3:2 ratio; the labeled training samples and unlabeled training samples form the total training sample matrix; and all 73 fault types are classified using 7-bit binary labels.

[0048] Furthermore, the deep kernel extreme learning machine in step S5 can be used for multi-class classification problems. The specific steps for training and testing this method include:

[0049] First, the semi-supervised learning method ELM-AE is used to train each hidden layer. The training sample X is used as the target output of the first ELM-AE (X1 = X), and then the output weight β1 is calculated. The output weight β1 can be obtained by the following formula:

[0050]

[0051] This process of dimensionality reduction of the feature vectors is then completed.

[0052] The top classification layer uses an Extreme Learning Machine (ELM). The input to the ELM is the node information from the second hidden layer, and the output is the desired label. A Gaussian radial basis function kernel is introduced, and the parameters are optimized using the Eagle Optimization algorithm. The hidden layer of the ELM is the desired kernel matrix, as shown in the formula:

[0053]

[0054] The output weights are obtained by regularized least squares.

[0055] Furthermore, the steps in step S5 using a deep kernel extreme learning machine are as follows:

[0056] S51: Similar to other deep learning models, DKELM initially uses the semi-supervised learning method ELM-AE to train each hidden layer, but DKELM does not require fine-tuning of the network. Compared to other deep learning algorithms, DKELM does not require a long time to train the network model.

[0057] The information in the hidden layers of DKELM is extracted using an Extreme Learning Machine Autoencoder (ELM-AE). ELM-AE combines Extreme Learning Machine and autoencoder for feature extraction. An autoencoder is itself an unsupervised neural network model that can learn the hidden features of the input data (encoding). Simultaneously, the learned new features can be used to reconstruct the original input data (decoding), thus enabling feature dimensionality reduction. The input weights and thresholds of ELM-AE are generated through random unit orthogonalization, i.e., ω... T ω=I, b T b=1; To achieve better feature extraction results, a popular graph-based regularization framework is used instead of the least squares method for calculation. The principle is that if two samples x i and x j If they are close, then their labels y i and y j They should also be close to each other; the manifold regularization framework minimizes the following cost function:

[0058]

[0059] When applied to the training of ELM-AE, the output weight β is obtained by minimizing the following cost function:

[0060]

[0061] Where C is the external regularization parameter, λ is the internal regularization parameter, and Y∈R (l+u)×m It is the label Y of the labeled sample l And the label Y of the unlabeled sample u =0, W is a (l+u)×(l+u) diagonal matrix, and the first l diagonal elements are W ii =1 / N i N i This is the number of neurons in the i-th hidden layer, and the rest are 0.

[0062] For L DKELM If the derivative equals 0, then we get:

[0063]

[0064] The output weight β can be obtained using the following formula:

[0065]

[0066] Where, n k It is the number of nodes in the k-th hidden layer of DKELM.

[0067] S52: After extracting the hidden layer information, the top classification layer uses an Extreme Learning Machine (ELM). The input to the ELM is the node information of the k-th hidden layer, and the output is a label encoded in 7 bits. A Gaussian radial basis function kernel is also introduced, with the kernel matrix as follows:

[0068]

[0069] This method, when solving nonlinear classification problems, maps linearly inseparable classification problems in low-dimensional space to linearly separable problems in high-dimensional space through kernel functions, achieving good classification results. This yields the node information of the hidden layers in the Extreme Learning Machine, and the output layer weights are obtained by minimizing the regularized cost function of least squares estimation.

[0070] S53: To improve classification accuracy, the L2 regularization parameter of the output weights of the KLEM classification layer and the penalty coefficient σ of the Gaussian kernel matrix are optimized using the Eagle optimization algorithm. The fitness function is selected as the root mean square error function between the predicted output and the expected output, with the aim of minimizing the fitness function value globally. The fitness function is:

[0071]

[0072] The Eagle Optimizer algorithm used is a novel intelligent optimization algorithm with multiple exploration and exploitation strategies. Compared with other metaheuristic algorithms, the Eagle Optimizer algorithm has significant advantages. The algorithm exhibits four group behaviors, and the specific algorithm steps are as follows:

[0073] The initial location of the Skyhawk population is:

[0074]

[0075] In the formula,

[0076] x i,j =rand(0,1)×(UB-LB)+LB,i=1,2,...,N,j=1,2,...,D (20)

[0077] In the formula: x i,j Let represent the position of the i-th eagle in the j-th dimension; D represents the dimension; N represents the number of eagles; LB and UB represent the upper and lower bounds of the given problem; rand is a random number in the interval [0, 1].

[0078] A. Expand the search scope

[0079] The Skyhawk identifies prey areas and selects the best hunting grounds by flying vertically and crouching high. At this time, the AO (Agency Optimizer) sends explorers from high altitudes to fly around and determine the search area where the prey is located.

[0080]

[0081] In the formula,

[0082]

[0083] In the formula: X(t+1) is the solution for the next iteration, X best (t) is the optimal value at the current position; X M (t) represents the mean value from the first iteration to the current position.

[0084] B. Narrowing the search scope

[0085] When the prey area is spotted from high altitude, the eagle circles above the target prey, preparing to land and attack; this method is called short-gliding attack using contour flight. Here, the AO narrowly probes a selected area of ​​the target prey, preparing for the attack.

[0086] X2(t+1)=X best (t)×Levy(D)+X R (t)+(yx)×rand(0,1) (23)

[0087] In the formula, X2(t+1) is the solution for the next iteration, and Levy(D) is the Levy flight distribution function formed by formula (24); X R (t) is the random solution in the range [1,N] during the i-th iteration; y and x are spiral curve functions as shown in formula (26).

[0088]

[0089]

[0090] In the formula, Γ is the gamma function, β is a fixed value of 1.5, s is a fixed value of 0.01, and u and v are Gaussian random numbers with values ​​of (0,1).

[0091]

[0092] In the formula, θ is the helix angle and r is the helix radius, and its expression is shown in formula (27).

[0093] r=r1+U×D1 (27)

[0094] θ=-ω×D1+θ1 (28)

[0095] In the formula, U is a fixed constant U = 0.0265; D1 is an integer matrix from 1 to the length of the search space. The initial helix angle is denoted as .

[0096] C. Expand the scope of development

[0097] When the eagle has accurately identified its prey area and is ready to land and attack, it descends vertically and launches an initial attack to probe the prey's reaction. This method is called a low-altitude slow descent attack. Here, the AO (Area of ​​Arrest) uses a selected area of ​​the target to approach the prey and launch an attack.

[0098] X3(t+1)=α(X best (t)-X M (t))-rand(0,1)+δ((UB-LB)×rand(0,1)+LB) (29)

[0099] In the formula, X3(t+1) is the solution of the next iteration; α and δ are fixed values ​​of 0.1.

[0100] D. Narrow the scope of development

[0101] When the eagle approaches its prey, it attacks the prey on land based on its random movements. This method is called walking and seizing the prey, with the AO attacking the prey at the last possible moment.

[0102] X4(t+1)=QF(t)X best (t)-(G2X(t)×rand(0,1))-G1Levy(D)+G2×rand(0,1) (30)

[0103]

[0104] G1=2(1-t / T) (32)

[0105] G2=2Rand(0,1)-1 (33)

[0106] In the formula, X4(t+1) is the solution for the next iteration; QF(t) is the quality function used for balancing the search strategy; G1 represents the flight slope of the eagle, and G2 represents the various movements of the eagle used to track prey.

[0107] S54: After training the top-level parameters of DKELM based on the above-mentioned Tianying optimization algorithm, all training of DKELM for fault diagnosis is completed, which is used for open circuit fault diagnosis of NPC three-level inverters.

[0108] A three-level inverter open-circuit fault diagnosis system, the system comprising:

[0109] Fault signal acquisition module: Builds an NPC three-level inverter simulation model to simulate various fault conditions and acquires multiple sets of three-phase phase voltages and terminal voltages, i.e., multiple sets of fault signals.

[0110] Preprocessing module: preprocesses the multiple sets of fault signals collected in step S1 to obtain multiple sets of preprocessed fault signals;

[0111] Marginal spectrum and spectrum solving module: Perform ensemble empirical mode decomposition (EEMD) on the preprocessed phase voltage fault signal and obtain the marginal spectrum through Hilbert transform. At the same time, obtain the spectrum by fast Fourier transform (FFT) of the preprocessed terminal voltage. Combine the marginal spectrum, spectrum and the voltage positive half-cycle scaling factor of the phase voltage as the signal feature vector.

[0112] Sample partitioning module: Divides the signal feature vector into labeled training samples, unlabeled training samples and test samples. The labeled training samples and unlabeled training samples form a total training sample matrix. At the same time, all fault types are classified using binary labels.

[0113] Training module: The deep kernel extreme learning machine is trained using training samples. The output weights of ELM-AE are calculated using a popular regularization framework instead of the least squares method. The classification of the top-level extreme learning machine introduces a Gaussian kernel function and is optimized using the Eagle Optimization algorithm.

[0114] Test module: Tests the test samples based on the trained DKELM model and analyzes the fault diagnosis results.

[0115] A computer storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method described in any one of the preceding descriptions.

[0116] Compared with the prior art, the advantages of the present invention are as follows:

[0117] This invention utilizes EEMD decomposition to avoid the mode aliasing problem in fault signal decomposition. By combining Hilbert transform to calculate each IMF of the normalized vector, errors caused by endpoint effects between data segments can be eliminated, thus achieving continuous and smooth spectral analysis of time series.

[0118] The feature vectors formed by calculating multiple features are more discriminative and avoid misdiagnosis caused by similar fault features.

[0119] This invention is unaffected by changes in grid-side voltage frequency and amplitude, as well as load changes; it has the advantages of short training time, fast diagnostic speed, and high diagnostic accuracy. Attached Figure Description

[0120] Figure 1This is a flowchart of the NPC three-level inverter open-circuit fault diagnosis method of the present invention;

[0121] Figure 2 This is a topology diagram of the NPC three-level inverter used.

[0122] Figure 3 This is a fault diagnosis simulation diagram of an NPC three-level inverter;

[0123] Figure 4 The voltage waveform of phase a terminal of inverter Ta1 when an open-circuit fault occurs.

[0124] Figure 5 The three-phase phase voltage waveforms for an open-circuit fault in inverter Ta1;

[0125] Figure 6 EEMD decomposition diagram of the open-circuit fault signal of the Tb4Tc2 dual transistors in the inverter;

[0126] Figure 7 The marginal spectrum of the open-circuit fault signal of the inverter Tb4Tc2 dual transistors;

[0127] Figure 8 This is a diagram of the network training structure of a deep kernel extreme learning machine;

[0128] Figure 9 It is the fitness function evolution curve of the Skyhawk optimization algorithm;

[0129] Figure 10 This is the diagnostic result of an open circuit fault in Ta1Tc2. Detailed Implementation

[0130] The specific implementation of the present invention is described below with reference to embodiments:

[0131] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0132] Furthermore, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity of description and are not intended to limit the scope of the invention. Any changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0133] Example 1:

[0134] This invention provides a method for diagnosing open-circuit faults in NPC three-level inverters based on deep kernel extreme learning machines; the fault diagnosis process is as follows: Figure 1 As shown, its NPC inverter topology is as follows: Figure 2 As shown, it includes a DC power supply U dc Two DC-side capacitors C1 and C2, an IGBT with 12 anti-parallel diodes, and a three-phase load, wherein the two DC-side capacitors are identical; a fault diagnosis model of an NPC three-level inverter is built in MATLAB / Simulink. The relevant electrical parameters for implementation of this invention are set as follows: DC power supply U dc =311V, DC side capacitor C1 = DC side capacitor C2 = 5.6 × 10 -3 F, equivalent resistance R = 1 × 10 -3 Ω; The three-phase phase voltage and terminal voltage output by the inverter are used as characteristic signals for fault analysis. The specific fault diagnosis steps are as follows:

[0135] Step 1: After building the NPC three-level inverter simulation model, simulate 73 fault conditions, one of which is a fault-free normal operation; collect the output three-phase phase voltage and terminal voltage as characteristic signals; collect 50 sets of samples for each fault type, for a total of 3650 sets of samples; (e.g., ...) Figure 4 The figure shows the voltage U at phase a of Ta1 when an open-circuit fault occurs at 0.1s. ao Waveform diagram Figure 5 The three-phase voltage signals of Ta1 when an open-circuit fault occurs at 0.1s; each sample collects the data within 0.1s after the fault occurs;

[0136] Step 2: The phase voltage and terminal voltage in the 3650 sets of fault signals were subjected to Clark transformation and normalized to obtain normalized vectors. The formula is:

[0137]

[0138]

[0139] Step 3: Obtain the spectrum of the terminal voltage by performing a Fast Fourier Transform on the normalized vector of the terminal voltage obtained in Step 2). The formula is:

[0140]

[0141] in

[0142] The marginal spectrum of the normalized vector of the phase voltage obtained in step 2) is calculated. First, the normalized vector of the phase voltage is decomposed into several empirical mode components and a residual signal by EEMD decomposition. Then, Hilbert transform is performed on each intrinsic mode component and the marginal spectrum is calculated. Figure 6 The EEMD decomposition result of the Tb4Tc2 dual-transistor open-circuit fault signal is shown below. Figure 7 For the corresponding marginal spectrum; the relevant expression is:

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149]

[0150] To avoid situations where some fault characteristics are similar, a voltage positive half-cycle proportionality coefficient is added to the characteristic signal, as shown in the formula:

[0151]

[0152] These three features are combined into a feature vector. The feature vector of each sample has a dimension of 300, and the feature vectors are combined into a sample matrix of 3650*300.

[0153] Step 4: Divide the 3650*300 sample matrix into labeled training samples, unlabeled training samples, and test samples in a 5:3:2 ratio; the labeled training samples and unlabeled training samples form the total training sample matrix; use binary labels to classify 73 fault types as expected labels, and realize fault diagnosis through fault identification methods, as shown in Table 1.

[0154] Table 1. Classification of Open Circuit Faults in NPC Three-Level Inverters

[0155]

[0156] Step 5: Train the deep kernel extreme learning machine using training samples, such as... Figure 8The diagram shows the network training structure. DKELM initially uses the semi-supervised learning method ELM-AE to train each hidden layer, taking the training sample X as the target output of the first ELM-AE (X1 = X), and then calculating the output weight β1. The input weights and thresholds of ELM-AE are generated by random unit orthogonalization, i.e., ω... T ω=I, b T b = 1; the output weight β1 can be obtained using the following formula:

[0157]

[0158] Then, the output matrix H1 of the first hidden layer of DELM is used as the input and target output of the next ELM-AE to calculate β2. After training with ELM-AE twice, a 2920*70 feature matrix is ​​extracted. After extracting the hidden layer information, the top classification layer uses an Extreme Learning Machine (ELM). The input of the ELM is the node information of the second hidden layer, and the output is the desired label. At the same time, a Gaussian radial basis function kernel function is introduced, and the parameters are optimized using the Eagle Optimization algorithm. The hidden layer of the ELM is the kernel matrix obtained, and the formula is:

[0159]

[0160] The output weights are obtained using the regularized least squares method. The Skyhawk optimization algorithm optimizes the regularization coefficients and the penalty coefficients of the kernel matrix. The fitness function is the root mean square error function between the actual label and the expected label, expressed as:

[0161]

[0162] Because this algorithm employs multiple exploration and development strategies and requires few optimization parameters, the training time is significantly reduced, while simultaneously improving the accuracy of fault diagnosis; for example... Figure 9 The figure shows the fitness function curve of the Skyhawk optimization algorithm;

[0163] The trained DKELM model was saved, and the model was used to test the test samples to analyze the fault diagnosis results. By comparing multiple sets of actual labels and expected labels of the test samples, the accuracy of the test results was all above 98%, and the fault diagnosis time was completed within 0.1 seconds. Figure 10 The figure shows the diagnostic results of an open-circuit fault in Ta1Tc2. Simulation results show that the method has fast diagnostic speed and high diagnostic accuracy, and has a good effect in the field of fault diagnosis of NPC three-level inverters.

[0164] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0165] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

[0168] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. A method for diagnosing open-circuit faults in a three-level inverter, characterized in that, The method includes: S1: Build an NPC three-level inverter simulation model to simulate various fault conditions and collect multiple sets of three-phase phase voltages and terminal voltages, i.e., multiple sets of fault signals. S2: Preprocess the multiple sets of fault signals collected in step S1 to obtain multiple sets of preprocessed fault signals; S3: Perform ensemble empirical mode decomposition (EEMD) on the preprocessed phase voltage fault signal and obtain the marginal spectrum through Hilbert transform. At the same time, obtain the spectrum by fast Fourier transform (FFT) of the preprocessed terminal voltage. Combine the marginal spectrum, the spectrum, and the voltage positive half-cycle scaling factor of the phase voltage as the signal feature vector. S4: Divide the signal feature vector into labeled training samples, unlabeled training samples and test samples. The labeled training samples and unlabeled training samples form a total training sample matrix. At the same time, classify all fault types using binary labels. S5: The deep kernel extreme learning machine is trained using training samples. The output weights of ELM-AE are calculated using a popular regularization framework instead of the least squares method. The classification of the top-level extreme learning machine introduces a Gaussian kernel function and optimizes it using the Eagle Optimization algorithm. The trained DKELM model is tested on test samples, and the fault diagnosis results are analyzed.

2. The method for diagnosing open-circuit faults in a three-level inverter according to claim 1, characterized in that, In step S1, the step of sampling the NPC three-level inverter within one sampling period includes: S11: The open circuit faults of NPC three-level inverters are divided into 12 types of single tube open circuit faults, 12 types of double tube faults in the same phase but different bridge arms, 48 ​​types of double tube faults in different phases, plus the fault-free case, for a total of 73 types. S12: For each fault type, the three-phase phase voltage and terminal voltage of the NPC three-level inverter are sampled simultaneously to obtain the output three-phase phase voltage. and terminal voltage ;in: , , For the first Samples obtained from the second sampling , The maximum number of samples within a sampling period. It is a positive integer.

3. The method for diagnosing open-circuit faults in a three-level inverter according to claim 2, characterized in that, In step S2, the specific preprocessing process includes: S21: Convert the three-phase voltage and terminal voltage Obtained through Clark transformation and The formula is: (1) S22: The normalized vector is obtained after normalization calculation. The formula is: (2)。 4. The method for diagnosing open-circuit faults in a three-level inverter according to claim 3, characterized in that, In step S3, the steps of calculating the spectrum, marginal spectrum, and voltage positive half-cycle proportionality coefficient include: S31: Obtain the spectrum of the terminal voltage by performing a Fast Fourier Transform on the normalized vector of the terminal voltage obtained in step S2. The formula is as follows: (3) in This is obtained by utilizing the properties of the discrete Fourier transform to save computation; S32: Obtain the marginal spectrum from the normalized vector of the phase voltage obtained in step S2, as follows: S321: The normalized vector of the obtained phase voltage is used to decompose the signal into several intrinsic mode components and a residual signal using the signal extremum information. However, due to the problem of mode aliasing in Empirical Mode Decomposition (EMD), i.e., the same IMF contains signal information of different frequency components, or different IMFs contain signal information of the same component, random Gaussian white noise is added to the original signal to suppress mode aliasing through multiple experiments. The final IMF is the average of M experiments. After EMD, the original signal is divided into several empirical mode components and a residual signal. (4) Each IMF is: (5) In the formula This refers to the i-th eigenmode component in the m-th experiment; S322: Applying a Hilbert transform to each of the eigenmodes obtained above yields: (6) Amplitude function: (7) Phase function: (8) Instantaneous frequency: (9) The calculated HHT spectrum: (10) S323: Obtain the marginal spectrum based on the result of the Hilbert transform: (11) fixed Unchanged, for Integral; the definite integral in discrete signals can be approximately decomposed into the sum of the areas of multiple rectangles; in discrete signals, It is the time spectrum matrix The length of the rectangle is the length of the k-th data point. The width is the time interval, i.e. Therefore, the integral formula can be changed to the following formula: (12) S33: Add a voltage positive half-cycle proportionality coefficient to the characteristic signal, the formula is: (13)。 5. The method for diagnosing open-circuit faults in a three-level inverter according to claim 4, characterized in that, The spectrum of the terminal voltage, the marginal spectrum of the phase voltage, and the positive half-cycle proportionality coefficient of the phase voltage obtained in step S3 are used to construct feature vectors; multiple sets of fault feature vectors are divided into labeled training samples, unlabeled training samples, and test samples in a ratio of 5:3:2; the labeled training samples and unlabeled training samples form the total training sample matrix; at the same time, all 73 fault types are classified using 7-bit binary labels.

6. The method for diagnosing open-circuit faults in a three-level inverter according to claim 1, characterized in that, The deep kernel extreme learning machine in step S5 can be used for multi-class classification problems. The specific steps for training and testing this method include: First, the semi-supervised learning method ELM-AE is used to train each hidden layer, and the training sample X is used as the target output of the first ELM-AE. ), and then obtain the output weights. Output weights It can be obtained through the following formula: (14) This process of dimensionality reduction of the feature vectors is then completed. The top classification layer uses an Extreme Learning Machine (ELM). The input to the ELM is the node information from the second hidden layer, and the output is the desired label. A Gaussian radial basis function kernel is introduced, and the parameters are optimized using the Eagle Optimization algorithm. The hidden layer of the ELM is the desired kernel matrix, as shown in the formula: (15) The output weights are obtained by regularized least squares.

7. A three-level inverter open-circuit fault diagnosis system, characterized in that, The system includes: Fault signal acquisition module: Builds an NPC three-level inverter simulation model to simulate various fault conditions and acquires multiple sets of three-phase phase voltages and terminal voltages, i.e., multiple sets of fault signals. Preprocessing module: preprocesses the multiple sets of fault signals collected in step S1 to obtain multiple sets of preprocessed fault signals; Marginal spectrum and spectrum solving module: Perform ensemble empirical mode decomposition (EEMD) on the preprocessed phase voltage fault signal and obtain the marginal spectrum through Hilbert transform. At the same time, obtain the spectrum by fast Fourier transform (FFT) of the preprocessed terminal voltage. Combine the marginal spectrum, spectrum and voltage positive half-cycle scaling factor of the phase voltage as the signal feature vector. Sample partitioning module: Divides the signal feature vector into labeled training samples, unlabeled training samples and test samples. The labeled training samples and unlabeled training samples form a total training sample matrix. At the same time, all fault types are classified using binary labels. Training module: The deep kernel extreme learning machine is trained using training samples. The output weights of ELM-AE are calculated using a popular regularization framework instead of the least squares method. The classification of the top-level extreme learning machine introduces a Gaussian kernel function and is optimized using the Eagle Optimization algorithm. Test module: Tests the test samples based on the trained DKELM model and analyzes the fault diagnosis results.

8. A computer storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method of any one of claims 1-6.

Images