A power electronic system fault diagnosis method combining resampling and ensemble learning

Abstract

The application discloses a kind of power electronic system fault diagnosis methods combined with resampling and integrated learning, belong to power electronic equipment fault diagnosis technical field, the method includes real-time sampling electric power electronic system in the current data of one-half fundamental period, current data is normalized, and normalized data is obtained;According to the normalized data, based on fast fourier transform algorithm, the frequency domain characteristics of normalized data are obtained;According to the frequency domain characteristics, the feature vector of normalized data is obtained by feature extraction selector;According to the feature vector of normalized data, the fault class label is obtained using integrated classification model, and the electric power electronic system fault diagnosis is completed.The application solves the technical problems that the original data sample is unbalanced in the existing fault diagnosis method, which leads to inaccurate diagnosis and even wrong judgment of the existing machine learning model, effectively realizes the diagnosis of various fault types of sensors and power devices of electric power electronic system.

Description

Technical Field

[0001] This invention belongs to the field of power electronic equipment fault diagnosis technology, and particularly relates to a power electronic system fault diagnosis method that combines resampling and ensemble learning. Background Technology

[0002] Power electronic converters play an indispensable role in energy conversion systems and are widely used in photovoltaic power generation, railway electric traction transportation, battery charging, aerospace systems, and other fields. Because power electronic equipment is prone to failure and has a high failure rate, power electronic converters are one of the most common sources of failure. Therefore, accurate fault diagnosis is of great significance for fault-tolerant operation control after a failure and for further system maintenance.

[0003] With the rapid development of artificial intelligence and data science technologies, many data-driven fault diagnosis methods have been proposed to address the challenges and difficulties in fault diagnosis of power electronic converter systems, such as multiple fault types and inaccurate mathematical models. However, in most cases, intelligent diagnostic models are trained on a satisfactory dataset, which not only means sufficient samples and minimal noise, but also a balanced distribution of samples from different categories. In reality, power electronic converter systems have very few fault states, and the original historical monitoring datasets are always imbalanced, with more normal operation data samples than fault data samples. Because intelligent learning methods give equal importance to every sample, a few fault samples are easily ignored, leading to poor fault diagnosis performance for a small number of fault samples even with high training accuracy, resulting in misclassification of fault types. Therefore, in power electronic system fault diagnosis, there is an urgent need to design a classifier that can improve the diagnostic accuracy for a small number of fault samples without significantly sacrificing the accuracy of the majority of normal samples, addressing the problem of data imbalance. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a power electronic system fault diagnosis method that combines resampling and ensemble learning. This method solves the technical problem of imbalanced raw data samples in existing fault diagnosis methods, which leads to inaccurate diagnosis or even incorrect judgments by existing machine learning models. It effectively realizes the diagnosis of various fault types of sensors and power devices in power electronic systems.

[0005] To achieve the aforementioned objectives, the present invention employs the following technical solution: a fault diagnosis method for power electronic systems combining resampling and ensemble learning, comprising the following steps:

[0006] S1. Real-time sampling of current data within half of the fundamental cycle of the power electronic system, and per-unit processing of the current data to obtain per-unit data;

[0007] S2. Based on the standardized data, the frequency domain characteristics of the standardized data are obtained using the Fast Fourier Transform algorithm;

[0008] S3. Based on the frequency domain characteristics of the standardized data, the feature vector of the standardized data is obtained through a feature extraction selector;

[0009] S4. Based on the feature vector of the standardized data, use an integrated classification model to obtain fault category labels and complete the fault diagnosis of the power electronic system.

[0010] The beneficial effects of this invention are as follows: This invention effectively utilizes a feature extraction selector to achieve dimensionality reduction of feature vectors, enabling the resampling algorithm to achieve better results under different data balance ratios; the idea of ​​ensemble learning further improves the accuracy and classification ability of the classifier under imbalanced data; it can accurately diagnose the occurrence of faults in power devices and sensors, and realize real-time online identification of faults, discover abnormal problems in power electronic systems, and improve maintenance efficiency.

[0011] Furthermore, the expression for the per-scalarization process in step S1 is:

[0012] x in =x i / max(x1,x2,…x N )

[0013] Where, x in x is the per-unit value of the data. i x represents the actual data value, max(·) is the maximum value function, and x N This is the Nth data point in the data sample.

[0014] The beneficial effects of the above-mentioned further solutions are: to standardize the data, to unify the data under different operating conditions and load conditions, and to facilitate the comparison of the characteristics and parameters of various components of the power system.

[0015] Furthermore, the construction of the feature extraction selector in step S3 includes the following steps:

[0016] A1. Obtain normal data samples and fault data samples of the power electronic system, preprocess and standardize the normal data samples and fault data samples to obtain an initial dataset;

[0017] A2. Based on the initial dataset, obtain the frequency domain characteristics of the initial dataset using the Fast Fourier Transform algorithm;

[0018] A3. Based on the frequency domain characteristics of the initial dataset, the feature weighting algorithm ReliefF is used to evaluate the correlation between each feature attribute and the fault category in the frequency domain characteristics, and to obtain the feature weight of each feature attribute. The expression of the feature weighting algorithm ReliefF is as follows:

[0019]

[0020]

[0021] Where W(A) is the weight of the A-th feature, A is the feature number, R is the sample data, and H is the weight of the A-th feature. j Let j be the j-th nearest neighbor of R, where j is the nearest neighbor number and k is the total number of nearest neighbors selected. Then, diff(A,R,H) j For sample R and sample H j The difference on feature A, p(·) is the prior probability of the class, m is the number of sample samplings, Class(R) is the class label to which R belongs, and M j (C) represents the j-th nearest neighbor sample in the class, C is the number of all labels, diff(A,R1,R2) is the difference between sample R1 and sample R2 on feature A, R1 and R2 are both sample data referencing symbols, and min(·) is the minimum value function.

[0022] A4. Based on the feature weights of each feature attribute, remove feature attributes with weights less than zero to obtain the first feature attribute set;

[0023] A5. Based on the first set of feature attributes, select the m feature attributes with the largest average mutual information to obtain the feature subset of the fault category:

[0024]

[0025]

[0026] Where D(S,c) is the average mutual information between the feature subset S and the fault category c, S is the feature subset, c is the fault category, I(·) is the mutual information measurement result, and z i Let z be the i-th feature attribute, i be the feature attribute number, p(zi) be the marginal probability density function of the i-th feature attribute, p(c) be the marginal probability density function of fault category c, and p(zi,c) be the marginal probability density function of z. i The joint probability density function of c;

[0027] A6. Based on the aforementioned feature subset, select m mutually exclusive feature attributes by adding a minimum redundancy condition to obtain the feature set of maximum relevance minus minimum redundancy:

[0028] mRMR = max(DR)

[0029]

[0030] Where mRMR is the feature set of maximum relevance minus minimum redundancy, R is the result of the minimum redundancy condition, and z j Let j be the j-th feature attribute, where j is the feature attribute number;

[0031] A7. Based on the feature set of maximum relevance - minimum redundancy, the evaluation result is obtained;

[0032] A8. Based on the evaluation results, sort the feature attributes of the first feature attribute set, select the top 20 feature attributes in the sorting, and obtain the feature vector.

[0033] A9. Determine if m is the minimum value under the premise of ensuring that the feature vector test accuracy is greater than 95%. If so, obtain the feature extraction selector; otherwise, adjust the value of m and return to step A5.

[0034] The beneficial effects of the above-mentioned further scheme are as follows: the introduction of the feature extraction selector can extract more feature data, realize the dimensionality reduction of the feature vector, and make the boundaries of different categories of samples in the new feature vector after feature selection clearer, thus laying a good foundation for data resampling.

[0035] Furthermore, the method for constructing the integrated classification model in step S4 includes the following steps:

[0036] B1. Obtain a new feature vector for the initial dataset based on the feature extraction selector;

[0037] B2. Based on the new feature vector, extract the initial dataset to obtain the first dataset;

[0038] B3. The fault data in the first dataset is resampled using the safe-level SMOTE oversampling algorithm to obtain a balanced dataset;

[0039] B4. Based on the balanced dataset, several different RVFL classifiers are trained by adjusting the parameters of the RVFL network model; the expression for the output function of the RVFL network model is:

[0040]

[0041] Where f(X) is the output function value of the RVFL network model, X is the input vector of the RVFL network model, and ω j and b j β represents the weights and biases of hidden nodes between the input and hidden layers, respectively, where g is the activation function type and β is the weights and biases. jThe output weights are J, where J is the number of hidden nodes, j is the hidden node number, N is the number of input data, and x is the number of hidden nodes. j This refers to the j-th data point in the input vector.

[0042] B5. Evaluate the plurality of RVFL classifiers, select the RVFL classifiers whose evaluation values ​​reach the preset values, and use ensemble learning to obtain an ensemble classifier;

[0043] B6. Based on the ensemble classifier, add the output decision of the ensemble classification model to obtain the ensemble classification model.

[0044] The beneficial effects of the above-mentioned further solutions are: using resampling technology to solve the problem of data imbalance, and using multiple RVFL classifiers to classify the data, thereby improving the accuracy of classification.

[0045] Further, step B3 includes the following steps:

[0046] B301. Based on the first dataset, obtain the fault dataset and the normal dataset;

[0047] B302. Based on the aforementioned fault dataset, obtain the safety level ratio:

[0048]

[0049] Among them, S lr For the safety level ratio, S lp S is the number of instances of the k nearest neighbors of sample p in the fault dataset, where p is a sample in the fault dataset. ln Let n be the number of instances of the k nearest neighbors of sample n in the fault dataset, where n is the nearest neighbor of sample p.

[0050] B303. Based on the aforementioned safety level ratio, a new sample of the fault dataset is obtained:

[0051] x new =p + β(np)

[0052] Where, x new For a new sample in the fault dataset, β is the ratio of different safety levels;

[0053] B304. Determine whether the number of new samples in the fault dataset is equal to the number of samples in the normal dataset. If so, update the fault dataset in the first dataset with the new samples of the fault dataset to obtain a balanced dataset. Otherwise, update the fault dataset with the new samples of the fault dataset and return to step B302.

[0054] The beneficial effects of the above-mentioned further scheme are: it balances the dataset and avoids the disadvantage of smote random region synthesis, which may overlap with normal data samples.

[0055] Further, in step B5, the RVFL classifier is evaluated using the F-score and G-means metrics; the expression for the F-score is:

[0056]

[0057] Wherein, F-score is the F-score value, representing the combined value between precision and recall, β1 is a parameter that measures the relative importance of P and R, P is the precision of the RVFL classifier, and R is the recall of the RVFL classifier;

[0058] The expression for G-means is:

[0059]

[0060] The beneficial effects of the above-mentioned further scheme are: it uses two metrics for evaluation to avoid bias in the results, and the application of the metrics takes into account the accuracy and recall of the RVFL classifier, which can effectively avoid the adoption of low-quality RVFL classifiers.

[0061] Furthermore, the expression for the output decision of the ensemble classification model in step B6 is:

[0062] Y o =max(Y c1 ,Y c2 ,Y c3 …,Y cn )

[0063] Y c1 =mean(Y) 11 ,Y 12 ,Y 13 …,Y 1n )

[0064]

[0065] Among them, Y o To integrate the output of the classification model, Y cn Y represents the average probability of fault label cn. c1 Y represents the average probability of fault label c1. R1 Let y be the probability of fault label c1 obtained from the Rth RVFL classifier, mean(·) be the average function, and Y be the output of a single RVFL classifier in the ensemble classifier. i Let y be the output value corresponding to the i-th label. jLet be the output value corresponding to the j-th label, softmax(·) be the normalized exponential function, and exp(·) be the exponential function with the natural constant e as the base.

[0066] The beneficial effect of the above-mentioned further solution is that the decision can calculate the probability of various possible fault labels, making it convenient for staff to view. Attached Figure Description

[0067] Figure 1 This is a flowchart of the method of the present invention.

[0068] Figure 2 This is a topology diagram of a single-phase pulse rectifier in an embodiment of the present invention.

[0069] Figure 3 This diagram illustrates the average test accuracy of the feature vectors after feature sorting using the original feature vectors, the mRMR algorithm, the ReliefF algorithm, and a combination of ReliefF and mRMR algorithms in different feature dimensions, as shown in the embodiments of the present invention.

[0070] Figure 4 This is a flowchart illustrating the integrated classification model testing and diagnostic decision-making method of this invention.

[0071] Figure 5 This is a schematic diagram showing the changes in the converter output grid-side voltage and current, DC-side voltage waveforms, and fault labels output by the integrated classification model before and after an open-circuit fault in an embodiment of the present invention made under a single-phase pulse rectifier experimental prototype system.

[0072] Figure 6 This is a schematic diagram showing the changes in the converter output grid-side voltage and current, DC-side voltage waveforms, and integrated classification model output fault labels before and after an open-circuit fault in an embodiment of the present invention made under a single-phase pulse rectifier experimental prototype system.

[0073] Figure 7 This is a schematic diagram showing the changes in the converter output grid-side voltage and current, DC-side voltage waveforms, and integrated classification model output fault labels before and after a current sensor offset fault, in an embodiment of the present invention made under a single-phase pulse rectifier experimental prototype system.

[0074] Figure 8 This is a schematic diagram showing the changes in the converter output grid-side voltage and current, DC-side voltage waveforms, and integrated classification model output fault labels before and after a current sensor gain failure, in an embodiment of the present invention made under a single-phase pulse rectifier experimental prototype system. Detailed Implementation

[0075] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0076] Example 1

[0077] like Figure 1 As shown, in one embodiment of the present invention, a power electronic system fault diagnosis method combining resampling and ensemble learning includes the following steps:

[0078] S1. Real-time sampling of current data within half of the fundamental cycle of the power electronic system, and per-unit processing of the current data to obtain per-unit data;

[0079] S2. Based on the standardized data, the frequency domain characteristics of the standardized data are obtained using the Fast Fourier Transform algorithm;

[0080] S3. Based on the frequency domain characteristics of the standardized data, the feature vector of the standardized data is obtained through a feature extraction selector;

[0081] S4. Based on the feature vector of the standardized data, use an integrated classification model to obtain fault category labels and complete the fault diagnosis of the power electronic system.

[0082] In this embodiment, the basic idea is to use a safe-level oversampling algorithm (SMOTE) to oversample the minority class, balancing the dataset from a data perspective. The dataset is trained using the fast computation speed of the random vector functional link network (RVFL), generating an RVFL network model. Based on the idea of ​​ensemble learning, the weights and parameters of the network model are modified to obtain diverse RVFL network models, thus solving the problem of identifying minority class fault samples in imbalanced data from an algorithmic perspective. Considering the characteristics of time-domain fault features in practical power electronic converter systems, which are easily submerged by noise and have similar characteristics across different faults, the original dataset is first processed through data preprocessing, feature extraction, and feature selection to obtain low-dimensional, more relevant features. Based on the carefully selected features, safe-level SMOTE is used to oversample and balance the dataset for the minority class. Then, a fast ensemble learning method is designed using the RVFL network model. Different metrics, F-score and G-means, are used to evaluate the RVFL classifier. Finally, the ensemble classification model outputs a decision for probability calculation, accurately identifying fault modes.

[0083] The expression for the per-unitization process in step S1 is:

[0084] x in =x i / max(x1,x2,…x N )

[0085] Where, x in x is the per-unit value of the data. i x represents the actual data value, max(·) is the maximum value function, and x N This is the Nth data point in the data sample.

[0086] In this embodiment, the raw operating data of the power system is typically a time-domain characteristic signal, exhibiting periodicity and temporal sequence. Characteristic data is extracted using one fundamental period of the current waveform as the window length. Considering different operating conditions and load conditions, normal operation data and fault data samples are acquired based on a power electronics system experimental platform. The data samples are then normalized by dividing by the maximum value in each set of sample data.

[0087] The construction of the feature extraction selector in step S3 includes the following steps:

[0088] A1. Obtain normal data samples and fault data samples of the power electronic system, preprocess and standardize the normal data samples and fault data samples to obtain an initial dataset;

[0089] In this embodiment, considering the small number of original samples of power electronic systems, a dataset was created that includes a large number of normal samples, a small number of open-circuit fault samples of power devices, and a small number of sensor fault samples, resulting in a total of 320 data samples. The ratio of normal samples, power device fault samples, and sensor fault samples is 2:1:1.

[0090] A2. Based on the initial dataset, obtain the frequency domain characteristics of the initial dataset using the Fast Fourier Transform algorithm;

[0091] In this embodiment, in order to extract more feature data, the Fast Fourier Transform is used to extract the spectral features in the data samples, and these features are combined with the original time-domain features to form a new feature vector.

[0092] A3. Based on the frequency domain characteristics of the initial dataset, the feature weighting algorithm ReliefF is used to evaluate the correlation between each feature attribute and the fault category in the frequency domain characteristics, and to obtain the feature weight of each feature attribute. The expression of the feature weighting algorithm ReliefF is as follows:

[0093]

[0094]

[0095] Where W(A) is the weight of the A-th feature, A is the feature number, R is the sample data, and H is the weight of the A-th feature. j Let j be the j-th nearest neighbor of R, where j is the nearest neighbor number and k is the total number of nearest neighbors selected. Then, diff(A,R,H) j For sample R and sample H j The difference on feature A, p(·) is the prior probability of the class, m is the number of sample samplings, Class(R) is the class label to which R belongs, and M j (C) represents the j-th nearest neighbor sample in the class, C is the number of all labels, diff(A,R1,R2) is the difference between sample R1 and sample R2 on feature A, R1 and R2 are both sample data referencing symbols, and min(·) is the minimum value function.

[0096] In this embodiment, ReliefF is a feature-weighted algorithm that assigns different weights to features based on the correlation between feature attributes and fault categories. The larger the feature weight, the stronger the classification ability of the feature; otherwise, the classification ability is weaker. Therefore, if the feature weight is less than the weight threshold, which is set to zero in this work, the feature is discarded.

[0097] A4. Based on the feature weights of each feature attribute, remove feature attributes with weights less than zero to obtain the first feature attribute set;

[0098] A5. Based on the first set of feature attributes, select the m feature attributes with the largest average mutual information to obtain the feature subset of the fault category:

[0099]

[0100]

[0101] Where D(S,c) is the average mutual information between the feature subset S and the fault category c, S is the feature subset, c is the fault category, I(·) is the mutual information measurement result, and z i Let p(z) be the i-th feature attribute, where i is the feature attribute number. i Let p(z) be the marginal probability density function of the i-th feature attribute, p(c) be the marginal probability density function of fault category c, and p(z) be the marginal probability density function of fault category c. i c) is z i The joint probability density function of c;

[0102] A6. Based on the aforementioned feature subset, select m mutually exclusive feature attributes by adding a minimum redundancy condition to obtain the feature set of maximum relevance minus minimum redundancy:

[0103] mRMR = max(DR)

[0104]

[0105] Where mRMR is the feature set of maximum relevance minus minimum redundancy, R is the result of the minimum redundancy condition, and z j Let j be the j-th feature attribute, where j is the feature attribute number;

[0106] A7. Based on the feature set of maximum relevance - minimum redundancy, the evaluation result is obtained;

[0107] A8. Based on the evaluation results, sort the feature attributes of the first feature attribute set, select the top 20 feature attributes in the sorting, and obtain the feature vector.

[0108] A9. Determine if m is the minimum value under the premise of ensuring that the feature vector test accuracy is greater than 95%. If so, obtain the feature extraction selector; otherwise, adjust the value of m and return to step A5.

[0109] In this embodiment, to achieve dimensionality reduction of the feature vector, the feature weighting algorithm ReliefF and the max-relevance min-redundance (mRMR) algorithm are combined to remove redundant and irrelevant components from the feature signal. First, the ReliefF algorithm is used to evaluate the correlation between feature attributes and fault categories. Then, feature attributes with weights less than zero are removed, while those with weights greater than zero are retained. Next, the mRMR algorithm is used to further evaluate the correlation between the retained feature attributes and fault categories, as well as the redundancy among features. Based on the evaluation results, the feature attributes are reordered. Finally, the top 20 ranked feature attributes are selected as the new feature vector. By combining the two feature selection algorithms, each attribute in the fault feature vector can be sorted according to its weight, relevance, and redundancy to the classification label. Ultimately, high-quality feature attributes are selected to construct a new feature vector with lower dimensionality, making the boundaries between different categories of samples in the new feature vector clearer and laying a good foundation for data resampling.

[0110] In this embodiment, to verify the effectiveness of the feature selection results, the feature vectors of the original data samples, sorted by the mRMR algorithm, the ReliefF algorithm, and the combination of mRMR and ReliefF algorithms, are compared. By gradually introducing feature vectors of different dimensions, common classification algorithms are trained and tested. Then, the average test accuracy can be calculated to judge the quality of the feature attributes. Finally, under the premise of ensuring high quality of feature vectors (test accuracy greater than 95%), the feature vector with the smallest dimension is selected to complete the most critical feature attribute selection.

[0111] The method for constructing the ensemble classification model in step S4 includes the following steps:

[0112] B1. Obtain a new feature vector for the initial dataset based on the feature extraction selector;

[0113] B2. Based on the new feature vector, extract the initial dataset to obtain the first dataset;

[0114] B3. Resample the fault data in the first dataset using the safe-level SMOTE oversampling algorithm to obtain a balanced dataset, including the following steps:

[0115] B301. Based on the first dataset, obtain the fault dataset and the normal dataset;

[0116] B302. Based on the aforementioned fault dataset, obtain the safety level ratio:

[0117]

[0118] Among them, S lr For the safety level ratio, S lp S is the number of instances of the k nearest neighbors of sample p in the fault dataset, where p is a sample in the fault dataset. ln Let n be the number of instances of the k nearest neighbors of sample n in the fault dataset, where n is the nearest neighbor of sample p.

[0119] B303. Based on the aforementioned safety level ratio, a new sample of the fault dataset is obtained:

[0120] x new =p + β(np)

[0121] Where, x new For a new sample in the fault dataset, β is the ratio of different safety levels;

[0122] B304. Determine whether the number of new samples in the fault dataset is equal to the number of samples in the normal dataset. If so, update the fault dataset in the first dataset with the new samples of the fault dataset to obtain a balanced dataset. Otherwise, update the fault dataset with the new samples of the fault dataset and return to step B302.

[0123] In this embodiment, based on the new feature vectors after feature extraction, the safe-level SMOTE algorithm is used to sample minority class samples of power device faults and sensor faults to obtain new samples and balance the dataset. Resampling techniques are expected to overcome the challenges of imbalanced data problems, including oversampling or undersampling methods. Generally, oversampling methods are more suitable for power electronic systems with smaller datasets because undersampling may lose some important information, thus affecting test accuracy. The safe-level SMOTE algorithm assigns a safety level to each instance before generating synthetic samples. New synthetic instances are created only in safe regions and closer to the maximum safety level. Synthetic samples are randomly generated by selecting the nearest minority neighbors of the minority samples. The calculation method for the safety level is as follows:

[0124] First, define dataset D as the set of all minority class samples to be oversampled, and p as a sample in D. Calculate the k nearest neighbors of sample p, and slp is equal to the number of instances of it in dataset D. Take any nearest neighbor sample and denote it as n. Calculate the k nearest neighbors of sample n, and slp is equal to the number of instances of it in dataset D.

[0125] Then calculate S. lp and S ln The quotient between them is defined as the safety level ratio of S. lr , means as follows:

[0126]

[0127] New sample x new It is based on S lr The range is generated between samples p and n, and is represented as follows:

[0128] x new =p + β(np)

[0129] Wherein, β takes a value according to different safety level ratios. When S lr Equal to ∞ and S lp When S equals 0, no samples are generated; when S... lr Equal to ∞ and S lp When S is not equal to 0, β equals 0, which is equivalent to replicating sample p; when S lr When S equals 1, β is a random number in the range [0 1]; when S... lr When β is greater than 1, it is [0 1 / S lr A random number within the range of ]; when S lr When β is less than 1, β is [1-S lr A random number within the range of 1].

[0130] The above rules are repeated until the required number of oversamples is met, so that the number of minority class samples and normal data samples are equal, resulting in a balanced dataset of 480 samples.

[0131] B4. Based on the balanced dataset, several different RVFL classifiers are trained by adjusting the parameters of the RVFL network model; the expression for the output function of the RVFL network model is:

[0132]

[0133] Where f(X) is the output function value of the RVFL network model, X is the input vector of the RVFL network model, and ω j and b j β represents the weights and biases of hidden nodes between the input and hidden layers, respectively, where g is the activation function type and β is the weights and biases. j The output weights are J, where J is the number of hidden nodes, j is the hidden node number, N is the number of input data, and x is the number of hidden nodes. j This refers to the j-th data point in the input vector.

[0134] B5. Evaluate the plurality of RVFL classifiers, select the classifiers whose evaluation values ​​reach the preset values, and use ensemble learning to obtain an ensemble classifier;

[0135] B6. Based on the ensemble classifier, add the output decision of the ensemble classification model to obtain the ensemble classification model.

[0136] In this embodiment, an RVFL classifier is trained using an RVFL network model based on a balanced dataset. Different RVFL classifiers are obtained by adjusting the parameters of the RVFL network model. The RVFL classifiers are evaluated using different metrics, F-score and G-means. Then, using the idea of ​​ensemble learning, multiple high-performing RVFL classifiers are constructed and combined to obtain an ensemble classifier to complete the learning task.

[0137] Define input data X = [x1, x2, ..., x N Output data Y = [y1, y2, ..., y] N The output function of the RVFL network model can be described as follows:

[0138]

[0139] The RVFL network model randomly generates weights and biases for hidden layer neurons, and calculates the output weights using the Moore-Penrose pseudo-inverse matrix operation for training. While maintaining training accuracy, different RVFL classifiers are trained by adjusting the number of hidden layer nodes and the type of activation function. The RVFL classifiers are evaluated using F-score and G-means metrics, ultimately resulting in an ensemble classifier containing multiple RVFL classifiers.

[0140] In this embodiment, the final decision output of the ensemble classifier is typically determined by a majority voting method, but this method cannot calculate the probability of various possible fault labels. A "softmax" function is introduced to transform the output of a single classifier, resulting in a probability matrix in the range [0 1]. The function is expressed as follows:

[0141]

[0142] Then, the outputs of all RVFL classifiers are calculated to determine the average probability of the same fault label, and the fault label corresponding to the highest probability value is output as the final result Ro.

[0143] In step B5, the RVFL classifier is evaluated using the F-score and G-means metrics; the expression for the F-score is:

[0144]

[0145] Wherein, F-score is the F-score value, representing the combined value between precision and recall, β1 is a parameter that measures the relative importance of P and R, P is the precision of the RVFL classifier, and R is the recall of the RVFL classifier;

[0146] The expression for G-means is:

[0147]

[0148] The expression for the output decision of the ensemble classification model in step B6 is:

[0149] Y o =max(Y c1 ,Y c2 ,Y c3 …,Y cn )

[0150] Y c1 =mean(Y) 11 ,Y 12 ,Y 13 …,Y 1n )

[0151]

[0152] Among them, Y o To integrate the output of the classification model, Y cn Y represents the average probability of fault label cn. c1 Y represents the average probability of fault label c1. R1 Let y be the probability of fault label c1 obtained from the Rth RVFL classifier, mean(·) be the average function, and Y be the output of a single RVFL classifier in the ensemble classifier. i Let y be the output value corresponding to the i-th label. j Let be the output value corresponding to the j-th label, softmax(·) be the normalized exponential function, and exp(·) be the exponential function with the natural constant e as the base.

[0153] In this embodiment, TP and TN represent the number of accurate positive samples and faulty samples output by the ensemble classifier on the test sample set, respectively, and FP and FN represent the number of incorrectly output positive samples and faulty samples output by the RVFL classifier on the test sample set, respectively. Precision and recall are two basic metrics of the classification model, denoted by P = TP / (TP+FP) and R = TP / (TP+FN), respectively. The F-score is a combined value between precision and recall, expressed as follows:

[0154]

[0155] Here, β1 is a parameter that measures the relative importance of P and R. In this invention, β1 equals 1, which means that P and R are equally important.

[0156] G-means is the geometric mean of P and R, expressed as follows:

[0157]

[0158] The F-score and G-means scores are commonly used to evaluate the performance of imbalanced data classification. The higher the F-score and G-means scores, the better the performance of the RVFL classifier.

[0159] Example 2

[0160] This invention provides a power electronic system fault diagnosis method combining resampling and ensemble learning. Taking open-circuit faults in power devices T1 and T3, current sensor offset faults, and gain faults in a single-phase pulse rectifier as examples, it considers different imbalance ratios between normal and fault samples, as well as grid-side voltage variations and load changes, for online diagnosis and localization of different fault modes. The basic circuit topology of a single-phase pulse rectifier is as follows: Figure 2As shown.

[0161] The power electronic system fault diagnosis method proposed in this invention combines resampling and ensemble learning. The input is the grid-side current is; the output of the online fault diagnosis algorithm for single-phase pulse rectifiers is the fault label output by the ensemble classification model.

[0162] First, the raw operating data of the single-phase pulse rectifier is collected. This raw operating data includes the grid-side voltage u. s , grid-side current i s and DC side voltage u dc The fundamental current period is 50Hz, and the sampling frequency is 20kHz. The original data sample consists of 200 dimensions of current data within half the fundamental period under the following conditions: normal operation of the single-phase pulse rectifier, open-circuit fault at T1, open-circuit fault at T3, current sensor offset fault, and gain fault. The grid-side voltage u is also considered. s Within a voltage range of 50V to 70V and a load resistance range of 20Ω to 40Ω, richer normal operation and fault data samples were obtained based on a power electronics system experimental platform. Data samples were normalized by dividing by the maximum value in each sample group and then labeled with fault category tags. Normal operation, open-circuit fault of power device T1, open-circuit fault of T3, current sensor offset fault, and gain fault were defined with fault category tags of 0, 1, 2, 3, and 4, respectively. A dataset was created including a large number of normal samples, a small number of open-circuit fault samples of power devices, and a small number of sensor fault samples, resulting in 320 data samples. The ratio of normal samples, open-circuit fault samples of T1 and T3 power devices, and current sensor offset and gain fault samples was 4:1:1:1:1.

[0163] Furthermore, the spectral features in the data samples are extracted using Fast Fourier Transform (FFT) to extract more frequency domain feature data. The amplitude of the frequency domain components after FFT is calculated as a new feature vector, resulting in 100-dimensional frequency domain feature data. This is then combined with the original 200-dimensional current data samples to form a feature vector with a dimension of 300.

[0164] Furthermore, to achieve dimensionality reduction of the feature vectors and remove redundant and irrelevant components from the feature signals, the ReliefF and mRMR algorithms are combined for feature ranking and verification. First, the ReliefF algorithm is used to evaluate the relationship between feature attributes and fault categories by calculating weights. Then, 200 feature attributes with weights less than zero are removed, retaining the 100 feature attributes with larger weights. Next, the max-relevance min-redundance (mRMR) algorithm is used to further evaluate the correlation between feature attributes and fault categories, as well as the redundancy between features, and the feature attributes are reordered.

[0165] Furthermore, to verify the effectiveness of the feature selection results, the feature vectors of the original data samples, sorted according to the mRMR algorithm, ReliefF algorithm, and a combination of mRMR and ReliefF algorithms, are compared. Common classification algorithms are trained and tested by progressively introducing feature vectors of different dimensions, and then the average test accuracy can be calculated. The results are as follows: Figure 3 As shown, when the feature dimension is greater than 40, the test accuracy of all feature vectors is greater than 95.1%. This means that, without considering any feature selection, the effectiveness of high-dimensional features is acceptable. However, for the most crucial feature selection, the RVFL classifier also achieved an accuracy of 96.6% with 20-dimensional features, compared to the combined mRMR and ReliefF algorithm. Finally, the top 20 ranked feature attributes were selected as the new feature vectors.

[0166] Furthermore, based on the new feature vectors after feature extraction and selection, different imbalance ratios (4:1:1:1:1, 8:1:1:1:1, and 16:1:1:1:1) were set to adjust the number of minority class samples for power device faults and sensor faults, respectively, to obtain imbalanced datasets D1, D2, and D3. The safe-level SMOTE algorithm was used to sample from datasets D1, D2, and D3 to obtain new samples, ultimately ensuring that the number of minority class samples and normal data samples was consistent, resulting in a balanced dataset of 480 samples, constructing balanced datasets D11, D22, and D33. Common classification algorithms were used to train the imbalanced datasets D1, D2, and D3, as well as the balanced datasets D11, D22, and D33, obtaining different classification models. The average test accuracies on the same test set were 0.9511, 0.8455, 0.7909, 0.9841, 0.9034, and 0.8659, respectively. It can be seen that the resampled dataset achieved better classification performance.

[0167] Furthermore, based on the resampled balanced dataset, a classifier was trained using the RVFL network, and its accuracy was tested using different metrics. The classifier was evaluated using F-score and G-means. By adjusting the parameters of the RVFL network model, and training and testing on dataset D11, a single RVFL classifier achieved a test accuracy of 0.9667, with F-scores of 0.9619 and G-means of 0.9768. On dataset D22, a single RVFL classifier achieved a test accuracy of 0.8875, with F-scores of 0.8503 and G-means of 0.8988. On dataset D33, a single RVFL classifier achieved a test accuracy of 0.8381, with F-scores of 0.7214 and G-means of 0.8096.

[0168] Furthermore, by changing the activation function type and the number of hidden layer nodes of the RVFL network model, n different RVFL classifiers are obtained. In this invention, n is 5. Then, using the idea of ​​ensemble learning, multiple high-performing RVFL classifiers are constructed and combined to obtain a trained ensemble classification model and complete the learning task.

[0169] Furthermore, such as Figure 4 As shown, the outputs of all RVFL classifiers are calculated to determine the average probability corresponding to the same fault label. The fault label corresponding to the highest probability value is output as the final result. Based on training and testing on dataset D11, the ensemble RVFL classifier achieves a test accuracy of 0.9810, with F-scores and G-means of 0.9751 and 0.9838, respectively. Based on dataset D22, a single RVFL classifier achieves a test accuracy of 0.9143, with F-scores and G-means of 0.8988 and 0.9103, respectively. Based on dataset D33, a single RVFL classifier achieves a test accuracy of 0.8476, with F-scores and G-means of 0.7522 and 0.8223, respectively.

[0170] Furthermore, current data within half of the fundamental period is sampled in real time, and the data is normalized by referring to the offline training process. The sampling fast Fourier transform algorithm is used for frequency domain feature extraction, and ReliefF and mRMR algorithms are used for feature selection. After constructing the feature vector, it is input into the trained ensemble classification model for diagnostic decision-making and outputs the fault category label.

[0171] Online testing was conducted on the diagnostic algorithm based on the RT-box controller and physical hardware testing platform. The test results for single-phase pulse rectifiers under different power device open-circuit faults and current sensor gain and offset faults are as follows: Figures 5-8 As shown ( Figure 5 For grid-side voltage u s 70V, load resistance R L Diagnostic test results for open-circuit fault in T1 at 40Ω. Figure 6 For grid-side voltage u s 70V, load resistance R L Diagnostic test results for open-circuit fault in T3 at 40Ω. Figure 7 For grid-side voltage u s The voltage is 80V, and the load resistance is R. L Diagnostic test results for current sensor offset fault at 20Ω. Figure 8 For grid-side voltage u s The voltage is 80V, and the load resistance is R. L(Diagnostic test results for current sensor gain fault at 20Ω).

Claims

1. A fault diagnosis method for power electronic systems combining resampling and ensemble learning, characterized in that, Includes the following steps: S1. Real-time sampling of current data within half of the fundamental cycle of the power electronic system, and per-unit processing of the current data to obtain per-unit data; S2. Based on the standardized data, the frequency domain characteristics of the standardized data are obtained using the Fast Fourier Transform algorithm; S3. Based on the frequency domain characteristics of the standardized data, the feature vector of the standardized data is obtained through a feature extraction selector; S4. Based on the feature vector of the standardized data, an ensemble classification model is used to obtain fault category labels, thus completing the fault diagnosis of the power electronic system; the method for constructing the ensemble classification model in step S4 includes the following steps: B1. Obtain a new feature vector for the initial dataset based on the feature extraction selector; B2. Based on the new feature vector, extract the initial dataset to obtain the first dataset; B3. The fault data in the first dataset is resampled using the safe-level SMOTE oversampling algorithm to obtain a balanced dataset; B4. Based on the balanced dataset, several different RVFL classifiers are trained by adjusting the parameters of the RVFL network model; the expression for the output function of the RVFL network model is: in, Let X be the output function value of the RVFL network model, and let X be the input vector of the RVFL network model. ω j and b j These represent the weights and biases of the hidden nodes between the function's input layer and hidden layer, respectively. g It is an activation function type. β j To output weights, J It is the number of hidden nodes. j To hide the node number, N It is the number of input data. x j For the first element in the input vector j One data point; B5. Evaluate the plurality of RVFL classifiers, select the RVFL classifiers whose evaluation values ​​reach the preset values, and use ensemble learning to obtain an ensemble classifier; B6. Based on the ensemble classifier, add the output decision of the ensemble classification model to obtain the ensemble classification model.

2. The power electronic system fault diagnosis method combining resampling and ensemble learning according to claim 1, characterized in that, The expression for the per-unitization process in step S1 is: in, The per-unit value of the data. For the actual data value, It is a function with maximum value. This is the Nth data point in the data sample.

3. The power electronic system fault diagnosis method combining resampling and ensemble learning according to claim 1, characterized in that, The construction of the feature extraction selector in step S3 includes the following steps: A1. Obtain normal data samples and fault data samples of the power electronic system, preprocess and standardize the normal data samples and fault data samples to obtain an initial dataset; A2. Based on the initial dataset, obtain the frequency domain characteristics of the initial dataset using the Fast Fourier Transform algorithm; A3. Based on the frequency domain characteristics of the initial dataset, the feature weighting algorithm ReliefF is used to evaluate the correlation between each feature attribute and the fault category in the frequency domain characteristics, and to obtain the feature weight of each feature attribute. The expression of the feature weighting algorithm ReliefF is as follows: in, Let A be the weight of the A-th feature, where A is the feature number. For sample data, for The j-th nearest neighbor, where j is the nearest neighbor number. For the total number of nearest neighbors selected, For the sample and samples The difference on feature A, Let m be the prior probability of class R, m be the number of sample draws, and Class(R) be the probability of class R. R The class tag to which it belongs. To represent the first in the class j Let C be the number of nearest neighbor samples and C be the total number of labels. For the sample and samples The difference on feature A, and These are all symbols used to refer to sample data. It is a minimum value function; A4. Based on the feature weights of each feature attribute, remove feature attributes with weights less than zero to obtain the first feature attribute set; A5. Based on the first set of feature attributes, select the m feature attributes with the largest average mutual information to obtain the feature subset of the fault category: in, Let S be the average mutual information between the feature subset S and the fault category c, where S is the feature subset and c is the fault category. For the measurement results of mutual information technology, Let i be the i-th feature attribute, where i is the feature attribute number. Let be the marginal probability density function of the i-th feature attribute. Fault Category c The marginal probability density function, for z i and c The joint probability density function; A6. Based on the aforementioned feature subset, select m mutually exclusive feature attributes by adding a minimum redundancy condition to obtain the feature set of maximum relevance minus minimum redundancy: in, The feature set is the one with the highest relevance and the lowest redundancy. R The result is the minimum redundancy condition. Let j be the j-th feature attribute, where j is the feature attribute number; A7. Based on the feature set of maximum relevance - minimum redundancy, the evaluation result is obtained; A8. Based on the evaluation results, sort the feature attributes of the first feature attribute set, select the top 20 feature attributes in the sorting, and obtain the feature vector. A9. Determine if m is the minimum value under the premise of ensuring that the feature vector test accuracy is greater than 95%. If so, obtain the feature extraction selector; otherwise, adjust the value of m and return to step A5.

4. The power electronic system fault diagnosis method combining resampling and ensemble learning according to claim 1, characterized in that, Step B3 includes the following steps: B301. Based on the first dataset, obtain the fault dataset and the normal dataset; B302. Based on the aforementioned fault dataset, obtain the safety level ratio: in, For safety level ratio, For the sample p of k The number of nearest neighbor instances in the fault dataset. p For the samples in the fault dataset, For the sample n of k The number of nearest neighbor instances in the fault dataset. n For the sample p The nearest neighbor sample; B303. Based on the aforementioned safety level ratio, a new sample of the fault dataset is obtained: in, For new samples in the fault dataset, The ratio of different safety levels; B304. Determine whether the number of new samples in the fault dataset is equal to the number of samples in the normal dataset. If so, update the fault dataset in the first dataset with the new samples of the fault dataset to obtain a balanced dataset. Otherwise, update the fault dataset with the new samples of the fault dataset and return to step B302.

5. The power electronic system fault diagnosis method combining resampling and ensemble learning according to claim 1, characterized in that, The evaluation of the RVFL classifier in step B5 specifically involves: The RVFL classifier is evaluated using the F-score and G-means metrics; the expression for the F-score is: in, F-score The F-score is a composite score representing the relationship between precision and recall. P For the accuracy of the RVFL classifier, R The recall rate of the RVFL classifier β 1 is a measure P and R Parameters of relative importance; The expression for G-means is: 。 6. The power electronic system fault diagnosis method combining resampling and ensemble learning according to claim 1, characterized in that, The expression for the output decision of the ensemble classification model in step B6 is: in, To integrate the output of the classification model, This represents the average probability of fault label cn. Fault Label c The average probability of 1 The fault label derived from the R-th RVFL classifier c The probability of 1 Let Y be the average function, and Y be the output of a single RVFL classifier in the ensemble classifier. i Let y be the output value corresponding to the i-th label. j This is the output value corresponding to the j-th label. For normalized exponential functions, It is an exponential function with the natural constant e as its base.

Images