A method and system for diagnosing open circuit fault of charging power module

Abstract

The present invention discloses a method and system for diagnosing an open-circuit fault of a charging power module, and relates to the technical field of fault diagnosis. A plurality of intrinsic mode components are generated by performing variational mode decomposition on the operating data of the charging power module. Feature extraction is performed based on the correlation coefficients corresponding to the intrinsic mode components, and a feature vector set is constructed. A feature selection algorithm is used to select features from feature vectors in the feature vector set, and a training set and a test set are generated. Based on a particle swarm optimization algorithm, a training set is used to optimize key parameters of an initial extreme learning machine, and a fault classification model is constructed. The test set is input into the fault classification model for open-circuit fault diagnosis, and fault diagnosis data is generated. Signal features are extracted through variational mode decomposition and feature selection algorithms, and are input into the fault classification model after particle swarm optimization, so as to realize rapid and accurate diagnosis of open-circuit faults of power switch tubes of the charging power module, thereby improving maintenance speed and equipment safety.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method and system for diagnosing an open circuit fault of a charging power module. Background Art

[0002] In recent years, the market demand for new energy electric vehicles has been growing year by year, driving the widespread deployment of charging piles, which has led to a steady increase in the number of charging piles. In charging piles, the charging power module is the key component with the highest value. And in the charging power module, the power switch tube is its key component. Fault identification, location and diagnosis of the power switch tube are beneficial to the online detection of the charging power module status and maintenance after damage, which has important economic significance for the charging operation service industry.

[0003] The failure modes of power switch tubes are mainly open circuit failure and short circuit failure. In practical applications, once a short circuit failure of the power switch tube is detected, the protection device will convert the short circuit failure into an open circuit failure by disconnecting or other means. Although the open circuit failure will not cause instantaneous overcurrent or undervoltage, causing the system to crash immediately, it will reduce system performance. Since the failure characteristics of the power switch tube are not obvious at the beginning of the open circuit failure, it is difficult to detect and enable the protection mechanism. However, if it is not handled in time, it will quickly affect other power devices to generate overcurrent, causing minor failures to quickly evolve into major failures, and ultimately causing the charging pile to be unable to charge. Therefore, the diagnosis of the open circuit failure of the charging power module is very important.

[0004] At present, a data-driven open circuit fault diagnosis method is used to diagnose the open circuit fault of the charging power module. However, the data-driven diagnosis method has high requirements on data, requires a large amount of computing resources, has a slow fault diagnosis speed, and is prone to errors, resulting in low accuracy of the diagnostic results. Summary of the invention

[0005] The present invention provides a method and system for diagnosing an open circuit fault of a charging power module, which solves the technical problems that a data-driven diagnosis method has high requirements on data, requires a large amount of computing resources, has a slow fault diagnosis speed, is prone to errors, and leads to low accuracy of the diagnosis result.

[0006] The present invention provides a method for diagnosing an open circuit fault of a charging power module, comprising:

[0007] Acquire operation data of the charging power module, perform variational mode decomposition on the operation data, and generate multiple eigenmode components;

[0008] Extract features based on the correlation coefficients corresponding to the eigenmode components to construct a feature vector set;

[0009] Using a feature selection algorithm to select features from the feature vectors in the feature vector set to generate a training set and a test set;

[0010] Based on the particle swarm optimization algorithm, the training set is used to optimize the key parameters of the initial extreme learning machine to build a fault classification model;

[0011] The test set is input into the fault classification model to perform open circuit fault diagnosis, and fault diagnosis data corresponding to the charging power module is generated.

[0012] Optionally, the operating data includes a three-phase input voltage signal, a three-phase input current signal and a DC output voltage signal; and the step of performing variational mode decomposition on the operating data to generate a plurality of intrinsic mode components includes:

[0013] Performing sample division on the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal according to a preset division standard to generate a plurality of samples;

[0014] Performing variational modal decomposition on the samples respectively to generate multiple modal components corresponding to the samples;

[0015] An initial constraint model is constructed with the minimum estimated bandwidth of all the modal components and the sum of all the modal components being equal to the corresponding sample as the optimization goal;

[0016] The initial constraint model is updated by using a preset Lagrangian expansion expression to generate a target constraint model;

[0017] The target constraint model is used to iteratively update the modal components using an alternating direction multiplier algorithm and a preset convergence formula to generate a plurality of eigenmodal components.

[0018] Optionally, the step of performing feature extraction based on the correlation coefficient corresponding to the eigenmode component to construct a feature vector set includes:

[0019] Substituting the sample and data point data corresponding to the eigenmode component into a preset correlation coefficient formula respectively, and calculating the correlation coefficient corresponding to the eigenmode component;

[0020] The preset correlation coefficient formula is:

[0021]

[0022] In the formula, ρ i is the correlation coefficient between the intrinsic mode component and the original signal, where the original signal refers to the sample corresponding to the intrinsic mode component; y q is the qth data point of the original signal; y0 is the mean of all data points of the original signal; xiq is the qth data point of the i-th eigenmode component; x i0 is the mean value of the i-th intrinsic mode component; Q is the number of data points contained in the original signal;

[0023] Taking the eigenmode component corresponding to the maximum value in the correlation coefficient as the effective eigenmode component;

[0024] Based on the data point data corresponding to the effective eigenmode components, time domain feature calculation is performed to construct a feature vector set.

[0025] Optionally, the data point data includes the number of data points and multiple data points; the step of performing time domain feature calculation based on the data point data corresponding to the effective intrinsic mode component to construct a feature vector set includes:

[0026] Performing mean calculation using the number of data points and the data points to generate a mean corresponding to the effective eigenmode component;

[0027] Performing variance calculation using the mean, the number of data points, and the data points to generate a variance corresponding to the effective intrinsic mode component;

[0028] Performing peak value calculation using the data points to generate a peak value corresponding to the effective eigenmode component;

[0029] Calculating kurtosis using the number of data points and the data points to generate kurtosis corresponding to the effective eigenmode component;

[0030] Performing root mean square calculation using the number of data points and the data points to generate a root mean square corresponding to the effective eigenmode component;

[0031] Calculating the ratio of the peak value to the root mean square value to generate a peak factor corresponding to the effective eigenmode component;

[0032] Calculating the ratio of the peak value to the mean value to generate a pulse factor corresponding to the effective eigenmode component;

[0033] Calculating a waveform factor using the root mean square and the mean to generate a waveform factor corresponding to the effective eigenmode component;

[0034] Calculating a margin factor using the peak value, the number of data points, and the data points to generate a margin factor corresponding to the effective intrinsic mode component;

[0035] The mean, the variance, the peak, the kurtosis, the root mean square, the peak factor, the pulse factor, the waveform factor and the margin factor are used as feature vectors to construct a feature vector set.

[0036] Optionally, the step of using a feature selection algorithm to perform feature selection on the feature vectors in the feature vector set to generate a training set and a test set includes:

[0037] Using a feature selection algorithm to sort the feature vectors in the feature vector set from large to small in importance, and generate a feature vector sequence;

[0038] Performing feature selection on the feature vector sequence according to a preset number of features to generate an initial feature set;

[0039] Normalizing the initial feature set to generate a target feature set;

[0040] The target feature set is divided into data according to a preset division ratio to generate a training set and a test set.

[0041] Optionally, the step of optimizing key parameters of an initial extreme learning machine using the training set based on a particle swarm optimization algorithm to construct a fault classification model includes:

[0042] Using the sample dimension, sample number and data classification data corresponding to the training set as the number of input layers, the number of hidden layers and the number of neurons in the output layer of the initial extreme learning machine to generate an intermediate extreme learning machine;

[0043] Initializing the connection weights from the input layer to the hidden layer and the hidden layer bias parameters of the intermediate extreme learning machine as particles to generate initial particles;

[0044] The intermediate extreme learning machine is updated using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine;

[0045] Inputting the training set into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer, and generating output connection weights corresponding to the initial particles;

[0046] The classification with the highest output probability corresponding to the output connection weight is used as the classification result predicted by the target extreme learning machine;

[0047] Compare the classification result with the corresponding sample actual category to generate a classification sample error rate;

[0048] Taking the classification sample error rate as the fitness corresponding to the initial particle and counting the number of iterations;

[0049] Particles are updated according to the fitness, the number of iterations and a preset maximum number of iterations to build a fault classification model.

[0050] Optionally, the step of inputting the training set into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer to generate the output connection weights includes:

[0051] Substituting the training set into a preset least squares solution formula through the target extreme learning machine to calculate the connection weight between the hidden layer and the output layer, and generating an output connection weight;

[0052] The preset least squares solution formula is:

[0053]

[0054]

[0055] In the formula, β is the output connection weight between the hidden layer and the output layer; H is the output matrix of the hidden layer; T is the output layer; m is the number of categories; K is the number of samples. When training the training set, m is 11 and K is 2755; t j is the output of the target extreme learning machine for each sample; g() is the hidden layer activation function; ω and b are the input connection weights and hidden layer bias parameters between the input layer and the hidden layer obtained by particle swarm optimization, respectively; x is the input sample; t mj is the probability that the jth input sample of the output layer belongs to the mth category; X j =[x 1j ,x 2j ,…,x nj ] T is the jth input sample; x n is the nth input layer node; ω i =[ω i1 ,ω i2 ,…ω in ] is the connection weight from the input layer node to the i-th hidden layer node; ω L =[ω L1 ,ω L2 ,…ω Ln ] is the connection weight from the input layer node to the Lth hidden layer node; b L is the bias parameter of the Lth hidden layer neuron; b i is the bias parameter of the i-th hidden layer neuron; L is the number of hidden layer neurons.

[0056] Optionally, the step of updating particles according to the fitness, the number of iterations and a preset maximum number of iterations to construct a fault classification model includes:

[0057] Determining whether the fitness does not meet a preset threshold;

[0058] If yes, the particle position and the particle velocity are updated using a preset iterative formula to generate a target particle;

[0059] The preset iteration formula is:

[0060] v i,d [k+1]=w k v i,d [k]+c1r1(pbest i,d -x i [k])+c2r2(gbest d -x i [k])

[0061] x i,d [k+1]=x i,d [k]+v i,d [k+1];

[0062] In the formula, x i,d [k] is the position of the dth dimension of the i-th particle at the k-th iteration; v i,d [k] is the velocity of the dth dimension of the ith particle at the kth iteration; r1 and r2 are random numbers between 0 and 1, which obey uniform distribution; c1 and c2 are constants; w k is the inertia weight, and the expression of inertia weight is:

[0063]

[0064] In the formula, w max 、w min are the maximum and minimum values ​​of the inertia weight respectively; k is the current iteration number;

[0065] If not, the initial particle is used as the target particle;

[0066] Determine whether the number of iterations is equal to a preset maximum number of iterations;

[0067] If yes, the target extreme learning machine corresponding to the target particle is used as a fault classification model;

[0068] If not, the target particle is used as the initial particle, and the step of updating the intermediate extreme learning machine using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine is executed.

[0069] The present invention also provides a diagnostic system for an open circuit fault of a charging power module, comprising:

[0070] An intrinsic mode component generating module, used for acquiring operation data of the charging power module, performing variational mode decomposition on the operation data, and generating a plurality of intrinsic mode components;

[0071] A feature vector set construction module, used for extracting features based on the correlation coefficients corresponding to the intrinsic mode components and constructing a feature vector set;

[0072] A training set and a test set generation module, used to select features from the feature vectors in the feature vector set using a feature selection algorithm to generate a training set and a test set;

[0073] A fault classification model building module is used to optimize key parameters of the initial extreme learning machine using the training set based on a particle swarm optimization algorithm to build a fault classification model;

[0074] The fault diagnosis data generation module is used to input the test set into the fault classification model to perform open circuit fault diagnosis and generate fault diagnosis data corresponding to the charging power supply module.

[0075] Optionally, the operation data includes a three-phase input voltage signal, a three-phase input current signal and a DC output voltage signal; and the intrinsic mode component generation module includes:

[0076] A sample generation module, used to perform sample division on the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal according to a preset division standard to generate a plurality of samples;

[0077] A modal component generation module, used to perform variational modal decomposition on the samples respectively to generate multiple modal components corresponding to the samples;

[0078] An initial constraint model construction module is used to construct an initial constraint model with the estimated bandwidth sum of all the modal components being minimized and the sum of all the modal components being equal to the corresponding sample as the optimization target;

[0079] A target constraint model generation module is used to update the initial constraint model using a preset Lagrangian expansion expression to generate a target constraint model;

[0080] The intrinsic mode component generation submodule is used to iteratively update the modal components by using the alternating direction multiplier algorithm and a preset convergence formula through the target constraint model to generate multiple intrinsic mode components.

[0081] It can be seen from the above technical solutions that the present invention has the following advantages:

[0082] The present invention obtains the operation data of the charging power module, performs variational mode decomposition on the operation data, and generates multiple eigenmode components. Feature extraction is performed based on the correlation coefficients corresponding to the eigenmode components to construct a feature vector set. A feature selection algorithm is used to select the feature vectors in the feature vector set to generate a training set and a test set. The training set is used to optimize the key parameters of the initial extreme learning machine based on the particle swarm optimization algorithm to construct a fault classification model. The test set is input into the fault classification model for open circuit fault diagnosis to generate fault diagnosis data corresponding to the charging power module. The technical problems that the data-driven diagnosis method has high requirements on data, requires a large amount of computing resources, has a slow fault diagnosis speed, is prone to errors, and leads to low accuracy of the diagnosis result are solved. By analyzing the open circuit fault of the charging power module, analyzing the signal change characteristics under the open circuit fault of different power switch tubes, extracting the signal features through variational mode decomposition and feature selection algorithms, and inputting them into the fault classification model after particle swarm optimization, the open circuit fault of the power switch tube of the charging power module can be quickly and accurately diagnosed, thereby providing a reference for on-site maintenance personnel and improving the maintenance speed and equipment safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0084] Figure 1 A flowchart of a method for diagnosing an open circuit fault of a charging power module provided in Embodiment 1 of the present invention;

[0085] Figure 2 A schematic diagram of the bipolar structure of a charging power module provided in Embodiment 1 of the present invention;

[0086] Figure 3 A schematic diagram of the topological structure of a charging power supply module provided in the first embodiment of the present invention;

[0087] Figure 4 A schematic diagram of the extreme learning machine network structure provided in the first embodiment of the present invention;

[0088] Figure 5 A flowchart of a method for diagnosing an open circuit fault of a charging power module provided in Embodiment 2 of the present invention;

[0089] Figure 6 A flowchart of a particle swarm optimization extreme learning machine provided in the second embodiment of the present invention;

[0090] Figure 7 This is a structural block diagram of a system for diagnosing an open circuit fault of a charging power module provided in Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0091] The embodiments of the present invention provide a method and system for diagnosing an open circuit fault of a charging power module, which are used to solve the technical problems that a data-driven diagnosis method has high requirements on data, requires a large amount of computing resources, has a slow fault diagnosis speed, and is prone to errors, resulting in low accuracy of the diagnosis result.

[0092] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0093] See also Figure 1 , Figure 1 This is a flowchart of the steps of a method for diagnosing an open circuit fault of a charging power module provided in Embodiment 1 of the present invention.

[0094] A method for diagnosing an open circuit fault of a charging power module provided in Example 1 of the present invention includes:

[0095] Step 101: Acquire operating data of a charging power module, perform variational modal decomposition on the operating data, and generate multiple intrinsic modal components.

[0096] The circuit structure of the charging power module is generally divided into a two-pole structure such as Figure 2 As shown. When the charging power module is working, the front-stage AC-DC rectifier circuit converts the three-phase AC power of the power grid into stable DC power and provides it to the rear-stage circuit. The rear-stage DC-DC circuit inverts the DC power output by the front-stage into AC power, and then converts it into the required AC power through a transformer. After secondary rectification and filtering, it finally outputs appropriate DC power to supply the load for charging.

[0097] The topology of the charging power module is as follows: Figure 3 As shown, the front stage AC-DC is a three-phase VIENNA rectifier structure, and U a , U b , U c is the three-phase grid voltage, L a , L b , L cThe same filter inductor, C1 and C2 are filter capacitors. The post-stage DC-DC is a full-bridge LLC resonant converter structure. In the figure, L4 and C3 are resonant inductor and resonant capacitor respectively, C4 is the post-stage output filter capacitor, and R2 is the load resistor. 14 The 14 MOSFET power switch tubes in the charging power module circuit are also the research target of the present invention.

[0098] In an embodiment of the present invention, the three-phase input current, three-phase input voltage and DC output voltage of the charging power module are used as original signals. A section of waveform data of the three-phase input voltage, current and DC output voltage when each switch tube is operating under normal conditions is collected in the charging power module to obtain the operation data of the charging power module. The waveform data is divided into multiple samples according to a certain length and interval. That is, the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal are sampled according to the preset division standard to generate multiple samples. The samples are respectively subjected to variational modal decomposition to generate multiple modal components corresponding to the samples. The estimated bandwidth and sum of all modal components are minimized, and the sum of all modal components is equal to the corresponding sample as the optimization target, and an initial constraint model is constructed. The initial constraint model is updated by using a preset Lagrangian expansion expression to generate a target constraint model. The modal components are iteratively updated by using an alternating direction multiplier algorithm and a preset convergence formula through the target constraint model to generate multiple intrinsic modal components.

[0099] Step 102: extract features based on the correlation coefficients corresponding to the intrinsic mode components and construct a feature vector set.

[0100] In an embodiment of the present invention, the samples and data point data corresponding to the eigenmode components are respectively substituted into the preset correlation coefficient formula to calculate the correlation coefficient corresponding to the eigenmode components. The eigenmode component corresponding to the maximum value in the correlation coefficient is taken as the effective eigenmode component. The time domain feature calculation is performed based on the data point data corresponding to the effective eigenmode component to construct a feature vector set.

[0101] Step 103: Use a feature selection algorithm to select features from the feature vectors in the feature vector set to generate a training set and a test set.

[0102] In the embodiment of the present invention, a feature selection algorithm is used to sort the feature vectors in the feature vector set from large to small in importance to generate a feature vector sequence. Feature selection is performed on the feature vector sequence according to a preset number of features to generate an initial feature set. The initial feature set is normalized to generate a target feature set. The target feature set is divided into data according to a preset division ratio to generate a training set and a test set.

[0103] Step 104: Based on the particle swarm optimization algorithm, the training set is used to optimize the key parameters of the initial extreme learning machine to build a fault classification model.

[0104] In an embodiment of the present invention, Figure 4 The network structure diagram of the initial extreme learning machine is shown in Figure 1. The initial extreme learning machine includes an input layer, a hidden layer, and an output layer. The sample dimension, sample number, and data classification data corresponding to the training set are used as the number of input layers, the number of hidden layers, and the number of neurons in the output layer of the initial extreme learning machine to generate an intermediate extreme learning machine. The connection weights from the input layer to the hidden layer and the hidden layer bias parameters of the intermediate extreme learning machine are used as particles to initialize and generate initial particles. The intermediate extreme learning machine is updated using the particle position and particle velocity corresponding to the initial particles to generate a target extreme learning machine. The training set is input into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer to generate the output connection weights corresponding to the initial particles. The classification with the highest output probability corresponding to the output connection weights is used as the classification result predicted by the target extreme learning machine. The classification result is compared with the actual category of the corresponding sample to generate the classification sample error rate. The classification sample error rate is used as the fitness corresponding to the initial particle and the number of iterations is counted. Particles are updated based on the fitness, the number of iterations, and the preset maximum number of iterations to build a fault classification model.

[0105] Step 105: Input the test set into the fault classification model to perform open circuit fault diagnosis, and generate fault diagnosis data corresponding to the charging power module.

[0106] In the embodiment of the present invention, the test set is input into the fault classification model for open circuit fault diagnosis, and the classification result and classification sample error rate are output to obtain the model classification accuracy. The classification result, classification sample error rate and model classification accuracy are used to construct the fault diagnosis data corresponding to the charging power module.

[0107] In an embodiment of the present invention, by acquiring the operation data of the charging power module, the operation data is subjected to variational mode decomposition to generate multiple eigenmode components. Feature extraction is performed based on the correlation coefficients corresponding to the eigenmode components to construct a feature vector set. A feature selection algorithm is used to select features from the feature vectors in the feature vector set to generate a training set and a test set. Based on the particle swarm optimization algorithm, the training set is used to optimize the key parameters of the initial extreme learning machine to construct a fault classification model. The test set is input into the fault classification model for open circuit fault diagnosis to generate fault diagnosis data corresponding to the charging power module. The technical problem that the data-driven diagnosis method has high requirements for data, requires a large amount of computing resources, has a slow fault diagnosis speed, and is prone to errors, resulting in low accuracy of the diagnosis result is solved. By analyzing the open circuit fault of the charging power module, analyzing the signal change characteristics under the open circuit fault of different power switch tubes, extracting signal features through variational mode decomposition and feature selection algorithms, and inputting them into the fault classification model after particle swarm optimization, a fast and accurate diagnosis of the open circuit fault of the power switch tube of the charging power module is achieved, thereby providing a reference for on-site maintenance personnel and improving the maintenance speed and equipment safety.

[0108] See also Figure 5 , Figure 5 This is a flowchart of a method for diagnosing an open circuit fault of a charging power module provided in Embodiment 2 of the present invention.

[0109] Another method for diagnosing an open circuit fault of a charging power module provided by Example 2 of the present invention includes:

[0110] Step 501: Acquire operating data of a charging power module, perform variational modal decomposition on the operating data, and generate a plurality of intrinsic modal components.

[0111] Furthermore, the operation data includes a three-phase input voltage signal, a three-phase input current signal and a DC output voltage signal, and step 501 may include the following sub-steps S11-S15:

[0112] S11. Perform sample division on the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal according to a preset division standard to generate a plurality of samples.

[0113] S12, performing variational modal decomposition on the samples respectively to generate multiple modal components corresponding to the samples.

[0114] S13. An initial constraint model is constructed with the estimated bandwidth sum of all modal components being minimized and the sum of all modal components being equal to the corresponding sample as the optimization goal.

[0115] S14. Using a preset Lagrangian expansion expression to update the initial constraint model, and generate a target constraint model.

[0116] S15. The modal components are iteratively updated by using an alternating direction multiplier algorithm and a preset convergence formula through a target constraint model to generate multiple eigenmodal components.

[0117] The preset division standard is to divide the collected waveform data into one sample every 10 ms with an interval of 2 ms.

[0118] In the embodiment of the present invention, 7 signal waveforms are collected from the normal state of the charging power module to the open circuit fault of each switch tube, including three-phase current signal, three-phase voltage signal and DC output voltage signal, where each signal contains 0.5s waveform data, and the collected waveform data is divided into a sample every 10ms and the interval is 2ms to obtain multiple samples. Then, the variational model decomposition algorithm is used to extract the signal features in the sample, and the specific steps are as follows:

[0119] Step 1: Assume that the original signal f is decomposed into K modal components, each modal component u k (t) are all kept at the center frequency ω k (t)Around.

[0120] Step 2: Construct an initial constrained variational model to find K modal components u for the signal k (t), so that the estimated bandwidth of all modal components is minimized, and the sum of all modes is equal to the constraint of the original input signal. The constraint expression of the initial constrained variational model is as follows:

[0121]

[0122] In the formula, {u k}={u1,u2,...,u k} are K modal components; {ω k}={ω1,ω2,...,ω k} is the frequency center of K modal components; δ(t) is the impulse function; f(t) is the original signal, i.e., the sample; * is the convolution operator; u k (t) is the kth modal component; t is the time variable; K is the preset number of decomposed modes; δ(t) is the unit pulse function; j is the imaginary unit.

[0123] Step 3: In order to solve the above constrained variational problem, the quadratic penalty factor α and the Lagrange operator λ(t) are introduced to convert it into the solution of the unconstrained problem. The initial constraint model is updated using the preset Lagrange expansion expression to obtain the target constraint model. The Lagrange expansion expression, that is, the preset Lagrange expansion expression, is as follows:

[0124]

[0125] In the formula, {u k}={u1,u2,...,u k} are K modal components; {ω k}={ω1,ω2,...,ω k} is the frequency center of K modal components; δ(t) is the impulse function; λ(t) is the Lagrange multiplier; α is the quadratic penalty factor; f(t) is the original signal, i.e., the sample; * is the convolution operator; u k (t) is the kth modal component; t is the time variable; K is the preset number of decomposed modes; δ(t) is the unit pulse function; j is the imaginary unit.

[0126] Step 4: Introduce the alternating direction multiplier algorithm ADMM, by continuously alternating updates and Find the solution to the expansion of the Lagrange operator.

[0127] Step 5: Set the discrimination accuracy ε, iterate until the preset convergence formula is satisfied, and terminate the iteration to obtain K modal components and their corresponding center frequencies, and output the optimal modal component as the intrinsic modal component IMF. The expression of the preset convergence formula is as follows:

[0128]

[0129] In the formula, K is the preset number of decomposition modes; is the kth modal component at the n+1th iteration; is the kth modal component at the nth iteration; n is the current iteration number; ε is the convergence accuracy.

[0130] Step 502: Substitute the samples and data points corresponding to the eigenmode components into the preset correlation coefficient formula respectively to calculate the correlation coefficient corresponding to the eigenmode components.

[0131] In the embodiment of the present invention, each sample signal is decomposed into K modes by variational mode decomposition to obtain multiple intrinsic mode components, and then the correlation coefficient between each intrinsic mode component IMF and the original signal is calculated by a preset correlation coefficient formula. The expression corresponding to the preset correlation coefficient formula is:

[0132]

[0133] In the formula, ρ i is the correlation coefficient between the intrinsic mode component and the original signal, where the original signal refers to the sample corresponding to the intrinsic mode component; y q is the qth data point of the original signal; y0 is the mean of all data points of the original signal; xiq is the qth data point of the i-th eigenmode component; x i0 is the mean value of the i-th eigenmode component; Q is the number of data points contained in the original signal.

[0134] Where ρ i The larger the value is, the higher the similarity between the intrinsic mode component IMF and the original signal is, and the more fault information it contains. Selecting the intrinsic mode component IMF with the largest correlation coefficient for feature extraction can reduce the interference component generated by decomposition iteration error or noise, and improve the accuracy of fault diagnosis.

[0135] Step 503: taking the eigenmode component corresponding to the maximum value in the correlation coefficient as the effective eigenmode component.

[0136] In the embodiment of the present invention, a feature vector with a relatively high weight is selected as a final fault feature quantity by using a feature selection algorithm ReliefF to obtain an effective intrinsic mode component.

[0137] Step 504: perform time domain feature calculation based on the data point data corresponding to the effective eigenmode components to construct a feature vector set.

[0138] Further, the data point data includes the number of data points and a plurality of data points, and step 504 may include the following sub-steps S21-S210:

[0139] S21. Calculate the mean using the number of data points and the data points to generate the mean corresponding to the effective eigenmode component.

[0140] S22, using the mean, the number of data points and the data points to perform variance calculation, and generate the variance corresponding to the effective eigenmode component.

[0141] S23, using the data points to perform peak value calculation, and generating peak values ​​corresponding to effective eigenmode components.

[0142] S24, using the number of data points and the data points to perform kurtosis calculation, and generating kurtosis corresponding to effective eigenmode components.

[0143] S25. Perform root mean square calculation using the number of data points and the data points to generate a root mean square corresponding to the effective eigenmode component.

[0144] S26. Calculate the ratio of the peak value to the root mean square value to generate a peak factor corresponding to the effective eigenmode component.

[0145] S27. Calculate the ratio of the peak value to the mean value to generate a pulse factor corresponding to the effective eigenmode component.

[0146] S28. Calculate the waveform factor using the root mean square and the mean value to generate the waveform factor corresponding to the effective eigenmode component.

[0147] S29. Calculate the margin factor using the peak value, the number of data points, and the data points to generate a margin factor corresponding to the effective intrinsic mode component.

[0148] S210, using mean, variance, peak, kurtosis, root mean square, peak factor, pulse factor, waveform factor and margin factor as feature vectors to construct a feature vector set.

[0149] In the embodiment of the present invention, the intrinsic mode component IMF with the largest correlation coefficient is selected as the effective IMF, that is, the effective intrinsic mode component, denoted as X, and the mean value X of X is solved. mean , Variance X var , peak value X peak , Kurtosis X kurtosis , RMS X rms , peak factor X crest , Pulse Factor X impulse , waveform factor X waveform , margin factor X magin The corresponding calculation formula is as follows:

[0150]

[0151]

[0152] X peak =max{|x i |};

[0153]

[0154]

[0155]

[0156]

[0157]

[0158]

[0159] Where N is the number of X data points; i Represents the i-th data point. Each dimension of each sample is processed above to obtain a feature vector set:

[0160]

[0161] in, Represents the feature vector obtained by feature extraction of the mth dimension of the ith sample.

[0162] Step 505: Use a feature selection algorithm to select features from the feature vectors in the feature vector set to generate a training set and a test set.

[0163] Further, step 505 may include the following sub-steps S31-S34:

[0164] S31. Use a feature selection algorithm to sort the feature vectors in the feature vector set from large to small in importance, and generate a feature vector sequence.

[0165] S32. Select features from the feature vector sequence according to a preset number of features to generate an initial feature set.

[0166] S33, normalizing the initial feature set to generate a target feature set.

[0167] S34. Divide the target feature set into data according to a preset division ratio to generate a training set and a test set.

[0168] The preset division ratio is 8: 2. The preset number of features refers to the number of feature vectors that need to be selected, which is usually set to 10.

[0169] In the embodiment of the present invention, the ReliefF feature selection algorithm is used to select the 10 columns with the largest weights from the feature vectors in the feature vector set to form a new feature vector set, namely the initial feature set, which is normalized as the final feature set, namely the target feature set. The target feature set is divided into a training set and a test set in a ratio of about 8:2.

[0170] Step 506: Based on the particle swarm optimization algorithm, the training set is used to optimize the key parameters of the initial extreme learning machine to build a fault classification model.

[0171] Further, step 506 may include the following sub-steps S41-S48:

[0172] S41. Using the sample dimension, sample number and data classification data corresponding to the training set as the number of input layers, the number of hidden layers and the number of neurons in the output layer of the initial extreme learning machine to generate an intermediate extreme learning machine.

[0173] S42, initializing the connection weights from the input layer to the hidden layer and the bias parameters of the hidden layer of the intermediate extreme learning machine as particles to generate initial particles.

[0174] S43. Update the intermediate extreme learning machine using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine.

[0175] S44, inputting the training set into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer, and generating output connection weights corresponding to the initial particles.

[0176] S45. The classification with the highest output probability corresponding to the output connection weight is used as the classification result predicted by the target extreme learning machine.

[0177] S46. Compare the classification result with the corresponding actual category of the sample to generate the classification sample error rate.

[0178] S47. The classification sample error rate is used as the fitness corresponding to the initial particle and the number of iterations is counted.

[0179] S48. Update particles according to fitness, number of iterations and a preset maximum number of iterations to build a fault classification model.

[0180] In the embodiment of the present invention, Figure 6As shown in the figure, the particle swarm optimization algorithm is used to optimize the extreme learning machine, namely the ELM network. When the ELM network randomly initializes the connection weights and bias parameters, it may cause the hidden layer nodes to fail, resulting in insufficient generalization ability. In order to overcome this defect, the PSO optimization algorithm, namely the particle swarm optimization algorithm, is introduced to optimize the two parameters initialized. In the particle swarm optimization algorithm, the connection weights between the input layer and the hidden layer of the ELM network and the bias parameters of the hidden layer nodes are used as particles, the classification error rate of the extreme learning machine training set is used as the particle fitness function, and the linearly decreasing inertia weight is used to update the particle velocity. Specifically, the number of neurons in the input layer, hidden layer and output layer of the ELM network is set according to the dimension of the input sample feature set, the number of samples and the number of output classifications, and the particle swarm optimization algorithm is used to perform initial optimization on the input connection weights from the input layer to the hidden layer of the ELM network structure and the bias parameters of the hidden layer, that is, to initialize the particle swarm scale, particle velocity and particle position to generate initial particles. Then, the particle position and particle velocity corresponding to the initial particles are used to update the intermediate extreme learning machine to generate the target extreme learning machine. The training set is input into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer, and the output connection weights corresponding to the initial particles are generated. The classification with the highest output probability corresponding to the output connection weights is used as the classification result predicted by the target extreme learning machine. The classification result is compared with the actual category of the corresponding sample, the classification sample error rate is generated, and the particle fitness is calculated. The classification sample error rate is used as the fitness corresponding to the initial particle and the number of iterations is counted. The individual extreme value and the group extreme value are updated through the preset iteration formula, and the speed and position of each particle are updated. The divided training set is used for training, and the particle parameters are continuously iterated and optimized with the classification error rate of the training set as the fitness. After reaching the maximum number of iterations, the optimal particle parameters are output, and the input connection weight ω and the hidden layer bias parameter b of the input layer and the hidden layer are obtained, so as to calculate the output matrix of the hidden layer and the output connection weight β of the hidden layer and the output layer, and obtain the PSO-ELM model, i.e., the fault classification model. Then the test set is input into the fault classification model to verify the performance of the model.

[0181] Further, step S44 may include the following sub-step S441:

[0182] S441. Substituting the training set into a preset least squares solution formula through the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer, and generating output connection weights.

[0183] The default least squares solution formula is:

[0184]

[0185]

[0186] Wherein, β is the output connection weight between the hidden layer and the output layer; H is the output matrix of the hidden layer; T is the output layer; m is the number of categories; K is the number of samples, m is 11 and K is 2755 during training set training; t j is the output of each sample through the target extreme learning machine; g(·) is the activation function of the hidden layer; ω and b are respectively the input connection weight between the input layer and the hidden layer and the hidden layer bias parameter obtained by particle swarm optimization, and x is the input sample; t mj is the probability that the j-th input sample of the output layer belongs to the m-th category; X j =[x 1j ,x 2j ,…,x nj T is the j-th input sample; x n is the n-th input layer node; ω i =[ω i1 ,ω i2 ,…ω in is the connection weight from the input layer node to the i-th hidden layer node; ω L =[ω L1 ,ω L2 ,…ω Ln is the connection weight from the input layer node to the L-th hidden layer node; b L is the bias parameter of the L-th hidden layer neuron; b i is the bias parameter of the i-th hidden layer neuron; L is the number of hidden layer neurons.

[0187] In the embodiment of the present invention, an extreme learning machine ELM is used as a classifier. ELM is a single hidden layer feedforward neural network. The specific implementation process is as follows:

[0188] Step 1: There are K groups of data samples, each sample contains n feature vectors, and the output is m classification. A three-layer ELM network structure is constructed, where the number of feature vectors input by the input layer, the number of neurons in the hidden layer, and the number of categories output by the output layer are n, B, and m respectively. The input matrix and the output matrix are denoted as X and Y:

[0189]

[0190] Step 2: Assume that there are B neurons in the hidden layer (B < K), the threshold is b, the activation function is g(·), ω yx is the connection weight between the x-th input layer node and the y-th hidden layer neuron, and β yz is the connection weight between the y-th hidden layer node and the z-th output layer neuron. b, ω yx and β yz are constituted as follows:

[0191] ​b=[b1,b2,...,b L ] T

[0192]

[0193]

[0194] Step 3: The network output layer T is:

[0195] T=[t1,t2,...,t K ] m×K ;

[0196] The model output t j for:

[0197]

[0198] Simplifying the above formula, we can get:

[0199] Hβ=T′′;

[0200] Step 4: The connection weight β between the hidden layer output matrix H and the output layer can be obtained by the least squares solution of the following formula:

[0201]

[0202] Finally, we get:

[0203] β=H + T′′;

[0204] Among them, H + is the Moore-Penrose generalized inverse of the hidden layer output matrix H; β is the output connection weight between the hidden layer and the output layer; H is the hidden layer output matrix; T is the output layer; m is the number of categories; K is the number of samples. When training the training set, m is 11 and K is 2755; t j is the output of the target extreme learning machine for each sample; g(·) is the hidden layer activation function; ω and b are the input connection weights and hidden layer bias parameters between the input layer and the hidden layer obtained by particle swarm optimization, respectively; x is the input sample.

[0205] The extreme learning machine obtains output weights through analytical methods, which is different from other neural network algorithms that require multiple iterations, so it can achieve a higher diagnostic speed.

[0206] Further, step S48 may include the following sub-steps S481-S486:

[0207] S481. Determine whether the fitness does not meet the preset threshold. If so, execute step S482; if not, execute step S483.

[0208] S482. Use a preset iterative formula to update the particle position and particle velocity to generate target particles.

[0209] S483. Use the initial particles as target particles.

[0210] S484. Determine whether the number of iterations is equal to the preset maximum number of iterations. If so, execute step S485; if not, execute step S486.

[0211] S485. Use the target extreme learning machine corresponding to the target particle as a fault classification model.

[0212] S486: Use the target particle as the initial particle, and jump to execute the step of updating the intermediate extreme learning machine using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine.

[0213] In the embodiment of the present invention, a particle swarm optimization algorithm is used to optimize the key parameters of the extreme learning machine. The specific implementation process is as follows:

[0214] Step 1: Set the fitness function and termination condition, and randomly initialize the position and velocity of the particle. The classification error rate of the actual category and the predicted category of the sample corresponding to the training set is used as the fitness function of PSO. The termination condition is set to reach the preset maximum number of iterations.

[0215] Step 2: Calculate the fitness of the particle in each iteration and update the individual extreme value p best and the group extreme value g best .

[0216] Step 3: Update the particle position and particle velocity according to the preset iterative formula to generate target particles.

[0217] The preset iteration formula is:

[0218] v i,d [k+1]=w k v i,d [k]+c1r1(pbest i,d -x i [k])+c2r2(gbest d -x i [k])

[0219] x i,d [k+1]=x i,d [k]+v i,d [k+1];

[0220] In the formula, x i,d[k] is the position of the dth dimension of the i-th particle at the k-th iteration; v i,d [k] is the velocity of the dth dimension of the ith particle at the kth iteration; r1 and r2 are random numbers between 0 and 1, which obey uniform distribution; c1 and c2 are constants, which are 2 in the experiment and are called individual learning factors and global learning factors, which are used to adjust the step size of the particle moving towards the individual optimal point or the global optimal point; w k is the inertia weight, and the expression of inertia weight is:

[0221]

[0222] In the formula, w max 、w min are the maximum and minimum values ​​of the inertia weight respectively; k is the current number of iterations.

[0223] Step 4: Determine whether the termination condition is met, terminate the iteration if it is met, and output the particle position with the best fitness.

[0224] The particle swarm optimization method is used to iteratively optimize the connection weights between the initial input layer and the hidden layer of the extreme learning machine network and the bias parameters of the hidden layer. The optimized extreme learning machine is used as a fault classification model, and the classification results are output to realize fault diagnosis.

[0225] Step 507: Input the test set into the fault classification model to perform open circuit fault diagnosis, and generate fault diagnosis data corresponding to the charging power module.

[0226] In the embodiment of the present invention, the specific implementation process of step 507 is similar to step 105 and will not be repeated here.

[0227] In the embodiment of the present invention, the charging power module is taken as the research object, the three-phase input current, voltage and DC output voltage are collected as the original signal, and the effective fault feature vector is extracted by variational mode decomposition and selected by the ReliefF algorithm, which utilizes more data information and prevents the dimension explosion problem. In the selection of the classifier, the advantages of the ELM structure being simple and fast training without back propagation are utilized, and the disadvantage that its initialization parameters may lead to poor generalization ability is overcome by particle swarm optimization, so as to achieve fast and accurate fault diagnosis.

[0228] Specifically, since the possibility of multiple power devices being open-circuited at the same time in an actual module is low, the open-circuit fault of a single power device is mainly considered for research. According to the normal situation and the single open-circuit fault of the remaining switch tubes, 15 fault types can be divided. Since the switch tubes in the rectifier and inverter bridge of the post-stage DC-DC circuit can be divided into groups of two according to the same characteristics, the specific open-circuit fault classification is shown in Table 1.

[0229] Table 1 Classification of open circuit faults in charging power module

[0230]

[0231] The simulation model of the charging power module is established using the simulation application package PSIM in the field of power electronics and motor control. When collecting simulation data, the sampling frequency is set to 20kHz, and 7 signal waveforms are collected from the normal state to the open circuit fault of each switch tube, including three-phase current signal, three-phase voltage signal and DC output voltage signal, each of which contains 0.5s waveform data, and the fault is set to occur at the 0.2s. The collected waveform data is divided into a sample every 10ms with an interval of 2ms, so the waveform data of the 7 signals can be divided into 246 samples, each sample contains 7 dimensions, and each dimension contains 200 data points. Since the open circuit experiment was carried out on 14 switches respectively, a total of 14×246=3444 groups of samples can be obtained.

[0232] The sample set is subjected to variational mode decomposition and eigenvector calculation. For each sample, it is decomposed into 6 modal components, and the mean, variance, peak value, kurtosis, root mean square, peak factor, pulse factor, waveform factor, and margin factor of the effective intrinsic modal component with the largest correlation coefficient are calculated as features. According to the above calculations, each sample can be extracted as a 1×636-dimensional feature vector, that is, a 3444×63-dimensional feature vector set is finally obtained. The 10 dimensions with the largest weights are selected from the 63-dimensional feature vectors by the ReliefF algorithm as the new feature vector set, that is, the initial feature set. The new feature vector set is normalized to obtain the final feature set, that is, the target feature set, and the target feature set is divided into a training set and a test set.

[0233] The specific implementation process is as follows:

[0234] Step 1: Decompose the original signal, i.e., the sample, to obtain K modal components. In the experiment, K is set to 6. The initial constraint model is constructed with the sum of the modal components equal to the original signal and the center frequency and modal minimum of each modal component as the optimization goal.

[0235] Step 2: Introduce the Lagrangian multiplier λ and the quadratic penalty factor α to construct the preset Lagrangian expansion expression.

[0236] Step 3: Use the alternating direction multiplier algorithm to iteratively solve the preset Lagrangian expansion expression.

[0237] Step 4: When the convergence condition is met, the iteration ends and the optimal modal component is output as the intrinsic mode function IMF. The convergence accuracy ε is set to 1×10 -7 Set the convergence accuracy ε to 1×10-7 .

[0238] Step 5: Calculate the correlation coefficient ρ between each intrinsic mode component IMF and its corresponding original signal i .

[0239] Step 6: Take the eigenmode component with the largest correlation coefficient as the effective IMF, denoted as X, and solve for the mean value X mean , Variance X var , peak value X peak , Kurtosis X kurtosis , RMS X rms , peak factor X crest , Pulse Factor X impulse , waveform factor X waveform , margin factor X magin index.

[0240] Step 7: According to the original signal, seven signals including three-phase input current, three-phase input voltage and DC output voltage are collected, that is, m = 7. Therefore, the feature vector expression of each sample can be obtained through feature extraction as follows:

[0241] Step 8: Use the ReliefF feature selection algorithm to select the 10 columns with the largest weights from the 63 columns of feature vectors to form a new feature vector set, namely the initial feature set, which is normalized as the final feature set, namely the target feature set. The target feature set is divided into a training set and a test set in a ratio of about 8:2, where the training set includes 2755 samples and the test set includes 689 samples.

[0242] According to the dimension of the input sample feature set, the number of samples and the number of output classifications, the number of neurons in the input layer, hidden layer and output layer of the initial ELM network is set. The particle swarm optimization algorithm is used to perform initial optimization on the connection weights from the input layer to the hidden layer and the bias parameters of the hidden layer of the initial ELM network structure. The divided training set is used for training, and the classification error rate of the training set is used as the fitness to continuously iteratively optimize the particle parameters. After reaching the maximum number of iterations, the optimal particle parameters are output to obtain the PSO-ELM model, i.e., the fault classification model. The test set is then input into the model to verify the performance of the model and generate the fault diagnosis data corresponding to the charging power module.

[0243] The specific implementation process is as follows:

[0244] Step 1: Since the dimension of the sample feature set is 10 and the classification problem to be implemented is 11 categories, the number of neurons in the input layer and output layer is set to 10 and 11 respectively, where the output of the output layer represents the probability of the sample being in each category. In order to make the ELM model converge, the number of hidden layers should be less than the number of samples. Considering the accuracy and convergence speed, the number of hidden layers is set to 100.

[0245] Step 2: Set the input connection weights and hidden layer bias of the ELM network as particles, initialize the position and velocity of the particles, set the termination condition to reach the maximum number of iterations, and set the maximum number of iterations to 100.

[0246] Step 3: Use the parameter information of the current position of the particle to initialize the ELM parameters, input the training set data into the current ELM network, and solve the output connection weight β between the hidden layer and the output layer through the preset least squares solution formula.

[0247] Step 4: Take the classification with the highest probability in the output of the current ELM network model as the classification result predicted by the model, compare it with the actual category of the sample, take the classification error rate of the actual category of the training set sample and the predicted category as the fitness function of PSO, calculate the fitness of each particle and update the individual extreme value and group extreme value of the particle.

[0248] Step 5: Update the particle position and velocity according to the preset iterative formula.

[0249] Step 6: When the maximum number of iterations is reached, the iterative optimization is stopped to obtain the optimal particle position. The parameter information contained corresponds to the connection weights between the extreme learning machine input and the hidden layer and the hidden layer bias parameters. The final ELM network, i.e., the fault classification model, is obtained through training of the training set.

[0250] Step 7: Input the test set into the final ELM network, calculate the model classification accuracy, and construct the fault diagnosis data corresponding to the charging power module.

[0251] In order to verify the superiority of the proposed method, the embodiments of the present invention respectively use VMD-traditional ELM, VMD-PSO-ELM and WPT-PSO-ELM to train and test the same training samples and test samples. The classification results of 689 test sets show that the VMD feature extraction method has a higher accuracy than the wavelet packet change feature extraction method, and the ELM neural network after particle swarm optimization can achieve better classification results than the traditional ELM, as shown in Table 2.

[0252] Table 2 Comparison of accuracy of different algorithms

[0253] algorithm Classification accuracy VMD-Traditional ELM 81.5675％ VMD-PSO-ELM 87.6633％ WPT-PSO-ELM 85.7765％

[0254] Verified by simulation experiments, the fault classification accuracy of the VMD-PSO-ELM algorithm reached 87.6633%, which is 6.0985% and 1.8868% higher than the traditional ELM and WPT-PSO-ELM diagnosis models under the same conditions, achieving a higher accuracy of fault diagnosis.

[0255] See also Figure 7 , Figure 7 This is a structural block diagram of a system for diagnosing an open circuit fault of a charging power module provided in Embodiment 3 of the present invention.

[0256] A third embodiment of the present invention provides a system for diagnosing an open circuit fault of a charging power module, comprising:

[0257] The intrinsic mode component generating module 701 is used to obtain the operating data of the charging power module, perform variational mode decomposition on the operating data, and generate multiple intrinsic mode components.

[0258] The feature vector set construction module 702 is used to extract features based on the correlation coefficients corresponding to the eigenmode components and construct a feature vector set.

[0259] The training set and test set generation module 703 is used to select features from the feature vectors in the feature vector set using a feature selection algorithm to generate a training set and a test set.

[0260] The fault classification model building module 704 is used to optimize the key parameters of the initial extreme learning machine using the training set based on the particle swarm optimization algorithm to build a fault classification model.

[0261] The fault diagnosis data generation module 705 is used to input the test set into the fault classification model to perform open circuit fault diagnosis and generate fault diagnosis data corresponding to the charging power supply module.

[0262] Furthermore, the operation data includes a three-phase input voltage signal, a three-phase input current signal and a DC output voltage signal, and the intrinsic mode component generation module 701 includes:

[0263] The sample generation module is used to perform sample division on the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal according to a preset division standard to generate multiple samples.

[0264] The modal component generation module is used to perform variational modal decomposition on the samples respectively to generate multiple modal components corresponding to the samples.

[0265] The initial constraint model building module is used to build the initial constraint model with the estimated bandwidth sum of all modal components being minimized and the sum of all modal components being equal to the corresponding sample as the optimization target.

[0266] The target constraint model generation module is used to update the initial constraint model using a preset Lagrangian expansion expression to generate a target constraint model.

[0267] The intrinsic mode component generation submodule is used to iteratively update the mode components by adopting the alternating direction multiplier algorithm and the preset convergence formula through the target constraint model to generate multiple intrinsic mode components.

[0268] Optionally, the feature vector set construction module 702 includes:

[0269] The correlation coefficient calculation module is used to substitute the sample and data point data corresponding to the eigenmode component into the preset correlation coefficient formula to calculate the correlation coefficient corresponding to the eigenmode component.

[0270] The default correlation coefficient formula is:

[0271]

[0272] In the formula, ρ i is the correlation coefficient between the intrinsic mode component and the original signal, where the original signal refers to the sample corresponding to the intrinsic mode component; y q is the qth data point of the original signal; y0 is the mean of all data points of the original signal; x iq is the qth data point of the i-th eigenmode component; x i0 is the mean value of the i-th eigenmode component; Q is the number of data points contained in the original signal.

[0273] The effective eigenmode component determination module is used to take the eigenmode component corresponding to the maximum value in the correlation coefficient as the effective eigenmode component.

[0274] The feature vector set construction submodule is used to perform time domain feature calculation based on the data point data corresponding to the effective intrinsic mode components and construct a feature vector set.

[0275] Optionally, the data point data includes the number of data points and a plurality of data points, and the feature vector set construction submodule may perform the following steps:

[0276] The number of data points and the mean of the data points are used to generate the mean corresponding to the effective eigenmode component;

[0277] The variance calculation is performed using the mean, number of data points, and data points to generate the variance corresponding to the effective intrinsic mode component;

[0278] The peak value is calculated using the data points to generate the peak value corresponding to the effective eigenmode component;

[0279] The kurtosis calculation is performed using the number of data points and the data points to generate the kurtosis corresponding to the effective eigenmode component;

[0280] The number of data points and the data points are used to perform RMS calculations to generate the RMS corresponding to the effective eigenmode components;

[0281] Calculate the ratio of peak value to RMS value to generate the peak factor corresponding to the effective eigenmode component;

[0282] Calculate the ratio of the peak value to the mean value to generate the pulse factor corresponding to the effective eigenmode component;

[0283] The waveform factor is calculated using the root mean square and the mean value to generate the waveform factor corresponding to the effective eigenmode component;

[0284] The margin factor is calculated using the peak value, the number of data points and the data points to generate the margin factor corresponding to the effective eigenmode component;

[0285] The mean, variance, peak, kurtosis, root mean square, peak factor, impulse factor, waveform factor and margin factor are used as feature vectors to construct a feature vector set.

[0286] Optionally, the training set and test set generating module 703 includes:

[0287] The feature vector sequence generation module is used to use a feature selection algorithm to sort the feature vectors in the feature vector set from large to small in importance to generate a feature vector sequence.

[0288] The initial feature set generation module is used to select features from the feature vector sequence according to the preset number of features to generate an initial feature set.

[0289] The target feature set generation module is used to normalize the initial feature set to generate the target feature set.

[0290] The training set and test set generation submodule is used to divide the target feature set data according to a preset division ratio to generate a training set and a test set.

[0291] Optionally, the fault classification model building module 704 includes:

[0292] The intermediate extreme learning machine generation module is used to generate an intermediate extreme learning machine by taking the sample dimension, sample number and data classification data corresponding to the training set as the number of input layers, the number of hidden layers and the number of neurons in the output layer of the initial extreme learning machine.

[0293] The initial particle generation module is used to initialize the connection weights from the input layer to the hidden layer and the bias parameters of the hidden layer of the intermediate extreme learning machine as particles to generate initial particles.

[0294] The target extreme learning machine generation module is used to update the intermediate extreme learning machine using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine.

[0295] The output connection weight generation module is used to input the training set into the target extreme learning machine to calculate the connection weight between the hidden layer and the output layer, and generate the output connection weight corresponding to the initial particle.

[0296] The classification result determination module is used to take the classification with the highest output probability corresponding to the output connection weight as the classification result predicted by the target extreme learning machine.

[0297] The classification sample error rate generation module is used to compare the classification result with the corresponding sample actual category and generate the classification sample error rate.

[0298] The module for determining the fitness and number of iterations is used to take the classification sample error rate as the fitness corresponding to the initial particle and count the number of iterations.

[0299] The fault classification model construction submodule is used to update particles according to fitness, number of iterations and preset maximum number of iterations to build a fault classification model.

[0300] Optionally, the output connection weight generation module may perform the following steps:

[0301] Substitute the training set into the preset least square solution formula through the target extreme learning machine to calculate the connection weight between the hidden layer and the output layer, and generate the output connection weight;

[0302] The default least squares solution formula is:

[0303]

[0304]

[0305] In the formula, β is the output connection weight between the hidden layer and the output layer; H is the output matrix of the hidden layer; T is the output layer; m is the number of categories; K is the number of samples. When training the training set, m is 11 and K is 2755; t j is the output of the target extreme learning machine for each sample; g(·) is the hidden layer activation function; ω and b are the input connection weights and hidden layer bias parameters between the input layer and the hidden layer obtained by particle swarm optimization, respectively; x is the input sample; t mj is the probability that the jth input sample of the output layer belongs to the mth category; X j =[x 1j ,x 2j ,…,x nj ] T is the jth input sample; x nis the nth input layer node; ω i =[ω i1 ,ω i2 ,…ω in ] is the connection weight from the input layer node to the i-th hidden layer node; ω L =[ω L1 ,ω L2 ,…ω Ln ] is the connection weight from the input layer node to the Lth hidden layer node; b L is the bias parameter of the Lth hidden layer neuron; b i is the bias parameter of the i-th hidden layer neuron; L is the number of hidden layer neurons.

[0306] Optionally, the fault classification model building submodule may perform the following steps:

[0307] Determine whether the fitness does not meet the preset threshold;

[0308] If yes, the preset iterative formula is used to update the particle position and particle velocity to generate the target particle;

[0309] The preset iteration formula is:

[0310] v i,d [k+1]=w k v i,d [k]+c1r1(pbest i,d -x i [k])+c2r2(gbest d -x i [k])

[0311] x i,d [k+1]=x i,d [k]+v i,d [k+1];

[0312] In the formula, x i,d [k] is the position of the dth dimension of the i-th particle at the k-th iteration; v i,d [k] is the velocity of the dth dimension of the ith particle at the kth iteration; r1 and r2 are random numbers between 0 and 1, which obey uniform distribution; c1 and c2 are constants; w k is the inertia weight, and the expression of inertia weight is:

[0313]

[0314] In the formula, w max 、w min are the maximum and minimum values ​​of the inertia weight respectively; k is the current iteration number;

[0315] If not, the initial particle is used as the target particle;

[0316] Determine whether the number of iterations is equal to the preset maximum number of iterations;

[0317] If so, the target extreme learning machine corresponding to the target particle is used as the fault classification model;

[0318] If not, the target particle is used as the initial particle, and the execution jumps to the step of updating the intermediate extreme learning machine using the particle position and particle velocity corresponding to the initial particle to generate the target extreme learning machine.

[0319] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0320] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing an open circuit fault of a charging power module, characterized in that: include: Acquire operation data of the charging power module, perform variational mode decomposition on the operation data, and generate multiple eigenmode components; Extract features based on the correlation coefficients corresponding to the eigenmode components to construct a feature vector set; Using a feature selection algorithm to select features from the feature vectors in the feature vector set to generate a training set and a test set; Based on the particle swarm optimization algorithm, the training set is used to optimize the key parameters of the initial extreme learning machine to build a fault classification model; Inputting the test set into the fault classification model to perform open circuit fault diagnosis, and generating fault diagnosis data corresponding to the charging power module; The operation data includes a three-phase input voltage signal, a three-phase input current signal and a DC output voltage signal; The step of performing variational mode decomposition on the operating data to generate a plurality of eigenmode components comprises: Performing sample division on the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal according to a preset division standard to generate a plurality of samples; Performing variational modal decomposition on the samples respectively to generate multiple modal components corresponding to the samples; An initial constraint model is constructed with the minimum estimated bandwidth of all the modal components and the sum of all the modal components being equal to the corresponding sample as the optimization goal; The initial constraint model is updated by using a preset Lagrangian expansion expression to generate a target constraint model; Iteratively updating the modal components by using an alternating direction multiplier algorithm and a preset convergence formula through the target constraint model to generate a plurality of eigenmodal components; The step of optimizing key parameters of the initial extreme learning machine using the training set based on the particle swarm optimization algorithm to construct a fault classification model includes: Using the sample dimension, sample number and data classification data corresponding to the training set as the number of input layers, the number of hidden layers and the number of neurons in the output layer of the initial extreme learning machine to generate an intermediate extreme learning machine; Initializing the connection weights from the input layer to the hidden layer and the hidden layer bias parameters of the intermediate extreme learning machine as particles to generate initial particles; The intermediate extreme learning machine is updated using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine; Inputting the training set into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer, and generating output connection weights corresponding to the initial particles; The classification with the highest output probability corresponding to the output connection weight is used as the classification result predicted by the target extreme learning machine; Compare the classification result with the corresponding sample actual category to generate a classification sample error rate; Taking the classification sample error rate as the fitness corresponding to the initial particle and counting the number of iterations; Particles are updated according to the fitness, the number of iterations and a preset maximum number of iterations to build a fault classification model.

2. The method for diagnosing an open circuit fault of a charging power module according to claim 1, characterized in that: The step of extracting features based on the correlation coefficients corresponding to the eigenmode components and constructing a feature vector set comprises: Substituting the sample and data point data corresponding to the eigenmode component into a preset correlation coefficient formula respectively, and calculating the correlation coefficient corresponding to the eigenmode component; The preset correlation coefficient formula is: In the formula, ρ i is the correlation coefficient between the intrinsic mode component and the original signal, where the original signal refers to the sample corresponding to the intrinsic mode component; y q is the qth data point of the original signal; y0 is the mean of all data points of the original signal; x iq is the qth data point of the i-th eigenmode component; x i0 is the mean value of the i-th intrinsic mode component; Q is the number of data points contained in the original signal; Taking the eigenmode component corresponding to the maximum value in the correlation coefficient as the effective eigenmode component; Based on the data point data corresponding to the effective eigenmode components, time domain feature calculation is performed to construct a feature vector set.

3. The method for diagnosing an open circuit fault of a charging power module according to claim 2, characterized in that: The data point data includes the number of data points and multiple data points; the step of performing time domain feature calculation based on the data point data corresponding to the effective intrinsic mode component to construct a feature vector set includes: Performing mean calculation using the number of data points and the data points to generate a mean corresponding to the effective eigenmode component; Performing variance calculation using the mean, the number of data points, and the data points to generate a variance corresponding to the effective intrinsic mode component; Performing peak value calculation using the data points to generate a peak value corresponding to the effective eigenmode component; Calculating kurtosis using the number of data points and the data points to generate kurtosis corresponding to the effective eigenmode component; Performing root mean square calculation using the number of data points and the data points to generate a root mean square corresponding to the effective eigenmode component; Calculating the ratio of the peak value to the root mean square value to generate a peak factor corresponding to the effective eigenmode component; Calculating the ratio of the peak value to the mean value to generate a pulse factor corresponding to the effective eigenmode component; Calculating a waveform factor using the root mean square and the mean to generate a waveform factor corresponding to the effective eigenmode component; Calculating a margin factor using the peak value, the number of data points, and the data points to generate a margin factor corresponding to the effective intrinsic mode component; The mean, the variance, the peak, the kurtosis, the root mean square, the peak factor, the pulse factor, the waveform factor and the margin factor are used as feature vectors to construct a feature vector set.

4. The method for diagnosing an open circuit fault of a charging power module according to claim 1, characterized in that: The step of using a feature selection algorithm to select features from the feature vectors in the feature vector set to generate a training set and a test set includes: Using a feature selection algorithm to sort the feature vectors in the feature vector set from large to small in importance, and generate a feature vector sequence; Performing feature selection on the feature vector sequence according to a preset number of features to generate an initial feature set; Normalizing the initial feature set to generate a target feature set; The target feature set is divided into data according to a preset division ratio to generate a training set and a test set.

5. The method for diagnosing an open circuit fault of a charging power module according to claim 1, characterized in that: The step of inputting the training set into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer to generate the output connection weights comprises: Substituting the training set into a preset least squares solution formula through the target extreme learning machine to calculate the connection weight between the hidden layer and the output layer, and generating an output connection weight; The preset least squares solution formula is: In the formula, β is the output connection weight between the hidden layer and the output layer; H is the output matrix of the hidden layer; T is the output layer; m is the number of categories; K is the number of samples. When training the training set, m is 11 and K is 2755; t j is the output of the target extreme learning machine for each sample; g() is the hidden layer activation function; ω and b are the input connection weights and hidden layer bias parameters between the input layer and the hidden layer obtained by particle swarm optimization, respectively; x is the input sample; t mj is the probability that the jth input sample of the output layer belongs to the mth category; X j =[x 1j ,x 2j ,…,x nj ] T is the jth input sample; x n is the nth input layer node; ω i =[ω i1 ,ω i2 ,…ω in ] is the connection weight from the input layer node to the i-th hidden layer node; ω L =[ω L1 ,ω L2 ,…ω Ln ] is the connection weight from the input layer node to the Lth hidden layer node; b L is the bias parameter of the Lth hidden layer neuron; b i is the bias parameter of the i-th hidden layer neuron; L is the number of hidden layer neurons.

6. The method for diagnosing an open circuit fault of a charging power module according to claim 1, characterized in that: The step of updating particles according to the fitness, the number of iterations and the preset maximum number of iterations to construct a fault classification model includes: Determining whether the fitness does not meet a preset threshold; If yes, the particle position and the particle velocity are updated using a preset iterative formula to generate a target particle; The preset iteration formula is: v i,d [k+1]=w k v i,d [k]+c1r1(pbest i,d -x i [k])+c2r2(gbest d -x i [k]) x i,d [k+1]=x i,d [k]+v i,d [k+1]; In the formula, x i,d [k] is the position of the dth dimension of the i-th particle at the k-th iteration; v i,d [k] is the velocity of the dth dimension of the ith particle at the kth iteration; r1 and r2 are random numbers between 0 and 1, which obey uniform distribution; c1 and c2 are constants; w k is the inertia weight, and the expression of inertia weight is: In the formula, w max 、w min are the maximum and minimum values ​​of the inertia weight respectively; k is the current iteration number; If not, the initial particle is used as the target particle; Determine whether the number of iterations is equal to a preset maximum number of iterations; If yes, the target extreme learning machine corresponding to the target particle is used as a fault classification model; If not, the target particle is used as the initial particle, and the step of updating the intermediate extreme learning machine using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine is executed.

7. A diagnostic system for open circuit fault of charging power module, characterized in that: include: An intrinsic mode component generating module, used for acquiring operation data of the charging power module, performing variational mode decomposition on the operation data, and generating a plurality of intrinsic mode components; A feature vector set construction module, used for extracting features based on the correlation coefficients corresponding to the intrinsic mode components and constructing a feature vector set; A training set and a test set generation module, used to select features from the feature vectors in the feature vector set using a feature selection algorithm to generate a training set and a test set; A fault classification model building module is used to optimize key parameters of the initial extreme learning machine using the training set based on a particle swarm optimization algorithm to build a fault classification model; a fault diagnosis data generating module, used for inputting the test set into the fault classification model to perform open circuit fault diagnosis and generate fault diagnosis data corresponding to the charging power supply module; The operation data includes a three-phase input voltage signal, a three-phase input current signal and a DC output voltage signal; The intrinsic mode component generation module comprises: A sample generation module, used to perform sample division on the three-phase input voltage signal, the three-phase input current signal and the DC output voltage signal according to a preset division standard to generate a plurality of samples; A modal component generation module, used to perform variational modal decomposition on the samples respectively to generate multiple modal components corresponding to the samples; An initial constraint model construction module is used to construct an initial constraint model with the estimated bandwidth sum of all the modal components being minimized and the sum of all the modal components being equal to the corresponding sample as the optimization target; A target constraint model generation module is used to update the initial constraint model using a preset Lagrangian expansion expression to generate a target constraint model; An eigenmode component generation submodule, used to iteratively update the modal components by using an alternating direction multiplier algorithm and a preset convergence formula through the target constraint model to generate multiple eigenmode components; The step of optimizing key parameters of the initial extreme learning machine using the training set based on the particle swarm optimization algorithm to construct a fault classification model includes: Using the sample dimension, sample number and data classification data corresponding to the training set as the number of input layers, the number of hidden layers and the number of neurons in the output layer of the initial extreme learning machine to generate an intermediate extreme learning machine; Initializing the connection weights from the input layer to the hidden layer and the hidden layer bias parameters of the intermediate extreme learning machine as particles to generate initial particles; The intermediate extreme learning machine is updated using the particle position and particle velocity corresponding to the initial particle to generate a target extreme learning machine; Inputting the training set into the target extreme learning machine to calculate the connection weights between the hidden layer and the output layer, and generating output connection weights corresponding to the initial particles; The classification with the highest output probability corresponding to the output connection weight is used as the classification result predicted by the target extreme learning machine; Compare the classification result with the corresponding sample actual category to generate a classification sample error rate; Taking the classification sample error rate as the fitness corresponding to the initial particle and counting the number of iterations; Particles are updated according to the fitness, the number of iterations and a preset maximum number of iterations to build a fault classification model.

Images