Fault Diagnosis Method for Single-Phase PWM Pulse Rectifier Based on Manifold Learning Algorithm

Abstract

The present invention relates to the technical field of fault diagnosis, and particularly relates to a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm. The method includes: constructing high-dimensional feature vectors corresponding to each operating state based on time-domain feature parameters, frequency-domain feature parameters, and time-frequency domain feature parameters in each operating state; using different manifold learning algorithms to reduce the dimension of the high-dimensional feature vectors to obtain low-dimensional feature vectors corresponding to each manifold learning algorithm; obtaining the optimal manifold learning algorithm and data near neighborhood according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm; using the optimal manifold learning algorithm and data near neighborhood to perform dimensionality reduction to obtain the current low-dimensional feature vector; and inputting the current low-dimensional feature vector into a fault diagnosis model to obtain the fault type of the current single-phase PWM pulse rectifier. The present invention improves the accuracy of fault diagnosis of a single-phase PWM pulse rectifier.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly relates to a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm. Background Art

[0002] The traction converter of an electric locomotive mainly consists of a single-phase PWM pulse rectifier and a three-phase bridge inverter. The single-phase PWM pulse rectifier realizes the voltage stability of the DC link through a traction controller, so as to provide a stable DC power supply for the traction inverter. Due to reasons such as its weak structure, complex control strategy, and harsh operating conditions, the power devices (IGBTs) and power diodes in the rectifier are prone to open-circuit faults, resulting in current distortion, a decrease in power factor, a decrease in the amplitude of the DC link voltage, and an increase in pulses. For the traction inverter, this will cause unbalanced three-phase currents and an increase in harmonic content, thus leading to excessive ripple and causing greater losses to the system, threatening the operation safety of the train. Therefore, there is an urgent need for a fault detection and diagnosis method for the single-phase PWM pulse rectifier in the electric locomotive traction system that can explore fault characteristics and types to improve its stability and reliability.

[0003] The traditional fault diagnosis method for a single-phase PWM pulse rectifier mainly diagnoses based on establishing a mathematical model. This method highly depends on the accuracy of system parameters and the mathematical model. Therefore, the established fault diagnosis model has poor portability and is not universal.

[0004] At the same time, with the continuous improvement of artificial intelligence technology and the rapid development of data analysis technology, data-driven fault diagnosis methods have been continuously proposed and widely concerned. The basic idea is to collect information on the operating state of the device, analyze the fault characteristics of the data and extract characteristic quantities from it, and finally identify the fault type.

[0005] However, due to the existence of a large number of nonlinear components in the single-phase PWM pulse rectifier, the collected data has a complex structure, a large amount of data, a harsh working environment, and a large amount of fault information. Unilaterally performing fault feature analysis and feature quantity extraction may lead to the loss of effective feature quantities or incomplete extraction of fault feature quantities, resulting in inaccurate fault identification. Therefore, it is necessary to extract multiple characteristic quantities during the operation of the single-phase PWM pulse rectifier. However, the original high-dimensional feature data formed by multiple characteristic quantities is large in amount and complex in structure. Therefore, it is necessary to select a suitable manifold learning algorithm. Therefore, it is necessary to construct the extracted characteristic quantities into a high-dimensional feature vector for dimensionality reduction. However, if the selected manifold learning algorithm for dimensionality reduction, the data intrinsic dimension d, and the data neighborhood K are not appropriate, it will affect the fault diagnosis rate of the single-phase PWM pulse rectifier. Summary of the Invention

[0006] To solve the above technical problems, the object of the present invention is to provide a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm, and the specific technical solution adopted is as follows:

[0007] An embodiment of the present invention provides a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm, and the method includes:

[0008] Respectively obtain the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters of the current signal when the single-phase PWM pulse rectifier operates in different operating states;

[0009] Construct high-dimensional feature vectors corresponding to each operating state based on the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters in each operating state;

[0010] Use different manifold learning algorithms to reduce the dimension of the high-dimensional feature vectors to obtain low-dimensional feature vectors corresponding to each manifold learning algorithm;

[0011] Obtain a comparison chart of the dimension reduction effects according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm; based on the comparison chart of the dimension reduction effects, obtain the optimal manifold learning algorithm and the data near neighborhood;

[0012] Use the optimal manifold learning algorithm and the data near neighborhood to reduce the dimension of the high-dimensional feature vectors of the current single-phase PWM pulse rectifier to obtain the current low-dimensional feature vectors;

[0013] Input the current low-dimensional feature vectors into the fault diagnosis model to obtain the fault type of the current single-phase PWM pulse rectifier.

[0014] Preferably, using different manifold learning algorithms to reduce the dimension of the high-dimensional feature vectors to obtain low-dimensional feature vectors corresponding to each manifold learning algorithm includes: determining the intrinsic dimension d according to the number of types of fault types of the single-phase PWM pulse rectifier; setting the value range of the data near neighborhood K, taking values of the data near neighborhood at preset intervals within the value range to obtain different values of the data near neighborhood K; based on the intrinsic dimension d and the different values of the data near neighborhood K, and combining different manifold learning algorithms to process the high-dimensional feature vectors respectively to obtain low-dimensional feature vectors corresponding to each manifold learning algorithm.

[0015] Preferably, a comparison graph of the dimensionality reduction effect is obtained according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm, including: calculating a first index corresponding to different values of the data neighborhood K according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm; establishing a comparison graph of the dimensionality reduction effect based on the first indices corresponding to different values of the data neighborhood K of each manifold learning algorithm, where the vertical axis in the comparison graph of the dimensionality reduction effect represents the value of the first index, and the horizontal axis represents the different values of the data neighborhood K of each manifold learning algorithm.

[0016] Preferably, inputting the current low-dimensional feature vector into a fault diagnosis model to obtain the fault type of the current single-phase PWM pulse rectifier, including:

[0017] Processing the high-dimensional feature vectors corresponding to the respective operating states by using the optimal manifold learning algorithm and data neighborhood to obtain the low-dimensional feature vectors corresponding to the respective operating states; dividing the low-dimensional feature vectors corresponding to the respective operating states into a training set and a test set according to a preset ratio; training the fault diagnosis model by using the training set and the test set, and inputting the current low-dimensional feature vector into the trained fault diagnosis model to output the fault type of the current single-phase PWM pulse rectifier.

[0018] Preferably, the time-frequency domain feature parameters include: processing the current signal by using a DB3 wavelet function to obtain at least two sub-band signal energies; the relative energy ratio of the sub-band signal energy to the total frequency is the time-frequency domain feature parameter.

[0019] Preferably, the different operating states include: the different operating states are the state where the single-phase PWM pulse rectifier operates normally and the operating states when the single-phase PWM pulse rectifier has different types of faults.

[0020] Preferably, based on the comparison graph of the dimensionality reduction effect, the optimal manifold learning algorithm and data neighborhood are obtained, including: the manifold learning algorithm and the data neighborhood K corresponding to the minimum first index in the comparison graph of the dimensionality reduction effect are the optimal manifold learning algorithm and data neighborhood.

[0021] The embodiments of the present invention have at least the following beneficial effects: By collecting the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters of the current signal during the operation of the single-phase PWM pulse rectifier, and analyzing these parameters using a fault diagnosis model, the present invention diagnoses whether a fault occurs during the operation of the single-phase PWM pulse rectifier, overcoming the problem of poor portability in traditional fault diagnosis by establishing a mathematical model and improving the universality of the fault diagnosis model. Moreover, to ensure the complete extraction of the characteristic quantities during the operation of the single-phase PWM pulse rectifier and the identification of a rich variety of fault types, the present invention collects the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters of the current signal during the operation of the single-phase PWM pulse rectifier. The amount of data collected is large, effectively avoiding the problem of human intervention in selecting fault characteristics, ensuring the integrity of the extracted characteristic quantities, and improving the accuracy of diagnosing faults during the operation of the single-phase PWM pulse rectifier. At the same time, when reducing the dimension of the high-dimensional feature vector constructed from the complete and complex data of the topology circuit structure of the single-phase PWM rectifier, by comparing the dimension reduction effects of different manifold learning algorithms, a suitable manifold learning algorithm is obtained to reduce the dimension of the high-dimensional feature vector. On the basis of ensuring the accuracy of fault detection for the single-phase PWM rectifier, the computational amount of the fault diagnosis model is reduced, ensuring the efficiency of diagnosis. Description of the Drawings

[0022] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0023] Figure 1 It is a method flow chart of a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm provided by an embodiment of the present invention;

[0024] Figure 2 It is a schematic diagram of the single-phase PWM pulse rectifier model provided by an embodiment of the present invention;

[0025] Figure 3 It is a comparison diagram of dimension reduction effects provided by an embodiment of the present invention;

[0026] Figure 4 It is a line chart of dimension reduction effects provided by an embodiment of the present invention;

[0027] Figure 5 It is a fault diagnosis block diagram of a single-phase PWM pulse rectifier provided by an embodiment of the present invention;

[0028] Figure 6This is the comparison chart of fault accuracy provided by the embodiments of the present invention. Detailed implementation manners

[0029] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following combines the accompanying drawings and preferred embodiments to specifically describe a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm proposed by the present invention, including its specific implementation manners, structures, features, and effects in detail. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0031] The following specifically describes the specific solution of a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm provided by the present invention with reference to the accompanying drawings.

[0032] Embodiment:

[0033] The main application scenario of the present invention is as follows: When diagnosing the faults of a single-phase PWM pulse rectifier in a locomotive traction converter, in order to ensure the accuracy of subsequent detections, it is necessary to extract characteristic quantities from multiple domains including the time domain, frequency domain, and time-frequency domain for analysis. However, due to the large number and complexity of the extracted data, it is necessary to reduce the dimension of the high-dimensional feature vectors obtained after fusing multi-domain data. Therefore, it is necessary to select a suitable manifold learning algorithm for dimension reduction to obtain low-dimensional data, and then analyze the low-dimensional data through a fault diagnosis model to identify the fault types of the single-phase PWM pulse rectifier.

[0034] Please refer to Figure 1 , which shows the flowchart of a fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm provided by the embodiments of the present invention. The method includes the following steps:

[0035] Step S1, respectively obtain the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters of the current signal when the single-phase PWM pulse rectifier operates in different operating states; construct high-dimensional feature vectors corresponding to each operating state based on the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters in each operating state.

[0036] In order to ensure the integrity of the data of the single-phase PWM pulse rectifier extracted in different operating states, it is necessary to extract features from multiple domains for fusion. In the embodiments of the present invention, the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters of the current signal of the single-phase PWM pulse rectifier are mainly extracted.

[0037] First, establish an open - circuit fault simulation model of the power switch tubes and power diodes of the single - phase PWM pulse rectifier in Matlab / Simulink software, and analyze the fluctuations of the grid - side current under different fault operating conditions of the single - phase PWM pulse rectifier;

[0038] As Figure 2 Shown is the schematic diagram of the single - phase PWM pulse rectifier model. Based on the transient current control strategy, establish an open - circuit fault simulation model of the power switch tubes and power diodes of the single - phase PWM pulse rectifier. And analyze that under different fault operating conditions of the single - phase PWM pulse rectifier, the fluctuation values of the grid - side current are different, and it is concluded that the grid - side current value can reflect the type of fault. Therefore, the grid - side current value is collected when the single - phase PWM pulse rectifier operates in different fault modes;

[0039] Regarding the operating conditions of the single - phase PWM pulse rectifier: it starts to work when the DC - side voltage reaches 2000V and tends to be stable and provides the input voltage for the inverter when it reaches 2800V. Design the DC - side voltage to be different values, and the rectifier operates in different states, including in the fault - free state and in different types of faults. Collect the grid - side current data, that is, the current signal, to form the original sample data set;

[0040] In the embodiment of the present invention, there are a total of 15 types including the fault types of the single - phase PWM pulse rectifier and the normal working state. For the power switch tubes, there are single - tube faults IGBT1, IGBT2, IGBT3, IGBT4; double - tube faults IGBT1 and IGBT2, IGBT1 and IGBT3, IGBT1 and IGBT4, IGBT2 and IGBT3, IGBT2 and IGBT4, IGBT3 and IGBT4. For the power diodes, they are all single - tube faults VD1, VD2, VD3, VD4, including the normal state, a total of 15 types. Among them, for the convenience of subsequent analysis, each fault type needs to be labeled, and the fault types and their corresponding labels are shown in Table 1;

[0041] Table 1 Fault types and classification labels

[0042]

[0043] Perform data pre - processing techniques on the collected current signals to ensure that the extracted features can provide sufficient signal details for the classification model. More precisely, to comprehensively and accurately describe the working - state characteristics of the rectifier under complex working conditions, it is necessary to extract time - domain, frequency - domain, and time - frequency - domain characteristic parameters from the original sample data set to construct a high - dimensional feature vector D H As the feature quantity after multi - domain feature fusion to improve the diagnostic accuracy.

[0044] AsFigure 2 As shown in the schematic diagram of the single-phase PWM pulse rectifier model, the initial value of the DC-side voltage Udc is Udc = 2000V, and the final value is Udc = 2800V. To increase the sample size collected and detect the robustness of the proposed fault diagnosis method to the DC-side voltage, and at the same time further ensure the integrity of the extracted fault features, preferably, in this embodiment, the change interval of the DC-side voltage is designed to be 10V. The sampling frequency is 10kHz, and the sampling time is 3s; thus, the sample size is 30000. The grid-side current signal Ia is collected under 15 fault modes. Each fault type has 81 current signal samples. For 15 faults, there are 81 * 15 = 1215 samples. The size of each sample is 30000, and the sample size, that is, the capacity of the original sample dataset Mo, is 1215 * 30000.

[0045] Furthermore, in engineering applications, the original sample dataset of the equipment is often preprocessed through time-domain feature extraction and frequency-domain feature extraction. Due to the differences in the regularity, sensitivity, and clustering of different characteristic parameters of the current signal, in order to describe the fault information more comprehensively, 10 time-domain characteristic parameters such as the most effective and widely used root mean square value, variance, skewness, etc., and 5 frequency-domain characteristic parameters such as average frequency, center frequency, root mean square frequency, etc. are adopted, as shown in Table 2;

[0046] Table 2 Time-domain and frequency-domain characteristic parameters

[0047]

[0048] Thus, the time-domain characteristic parameters and frequency-domain characteristic parameters of each sample, that is, each current signal, can be obtained. In addition, due to the influence of nonlinear elements in the pulse rectifier, the current signal will have characteristics such as non-stationarity and non-linearity, and resonance phenomena will occur during the power conversion process. Some fault information is not only hidden in the low-frequency signal, but there may be unobvious information hidden in the high-frequency signal. If only the relevant features are analyzed and extracted unidirectionally from the time domain or the frequency domain, the type of the fault cannot be completely identified, and there will still be great risks in the operation of the system. Therefore, it is necessary to extract relevant features from the time-frequency domain to realize the expression of the fault signal. Therefore, an algorithm that can decompose the high-frequency part of the fault signal, the wavelet packet transform, is selected to decompose the low-frequency and high-frequency segments of the fault signal simultaneously to thoroughly excavate the fault information.

[0049] Preferably, in the embodiments of the present invention, the DB3 wavelet function is selected to process the current signal. The implementer can also select other functions to process the electrical signal according to the actual situation. Processing the current signal with the DB3 wavelet function obtains the energy of at least two sub-band signals. Preferably, in the embodiments of the present invention, the input current signal is subjected to 5-layer orthogonal wavelet packet decomposition to obtain filtered signals of 32 sub-bands evenly divided, and the relative energy ratios of the obtained 32 sub-band signal energies to the total frequency are used as time-frequency domain characteristic parameters.

[0050] So far, after each current signal in each sample and each operating state is processed, as Figure 1 shown, 10 time-domain characteristics, 5 frequency-domain characteristics, and 32 time-frequency domain characteristics can be obtained.

[0051] At the same time, for the convenience of subsequent analysis and to ensure the integrity of the data, it is also necessary to fuse the multi-domain data collected and extracted. In the present invention, since the number of samples collected in each operating state is 81, a high-dimensional feature vector D of 1215*47 is thus formed H , the high-dimensional feature vector D H has a capacity of 15*81*47.

[0052] Step S2, use different manifold learning algorithms to reduce the dimension of the high-dimensional feature vector to obtain the low-dimensional feature vector corresponding to each manifold learning algorithm; obtain a comparison graph of the dimensionality reduction effect according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm; based on the comparison graph of the dimensionality reduction effect, obtain the optimal manifold learning algorithm and the data near neighborhood.

[0053] Due to problems such as partial feature redundancy, partial feature conflicts, and increased data dimensions in the construction of multi-domain features, the original high-dimensional feature data formed is large in quantity and complex in data structure, increasing the computational amount of the fault diagnosis model and unable to be directly used as fault recognition features. It is necessary to reduce the dimension of the constructed high-dimensional feature vector D H , so it is necessary to select a suitable manifold algorithm to reduce the dimension of the high-dimensional feature vector.

[0054] Manifold learning algorithms have received extensive attention among many dimensionality reduction algorithms due to their strong ability to extract low-dimensional data structures from non-linear high-dimensional data. However, the application of manifold learning theory is highly related to data structures and data parameters. It is currently mostly used in mechanical fault diagnosis methods and rarely used in the fault diagnosis of power electronic circuits. The fault diagnosis of a single-phase PWM rectifier is actually the fault diagnosis of its circuit topology. Selecting the parameters of manifold learning theory in the fault diagnosis of power electronic circuits according to the empirical method of selecting the parameters of manifold learning theory for mechanical fault diagnosis is likely to lead to problems such as over-dimensionality reduction and incorrect selection of near neighbor points, ultimately resulting in incorrect extraction of fault feature quantities.

[0055] Different manifold learning algorithms are suitable for different data manifold structures, with different advantages and disadvantages, and have different requirements for data structures and parameters. For example, the Isometric Feature Mapping (ISOMAP) algorithm belongs to a global learning algorithm and can well guarantee the distribution of data points during the dimensionality reduction process. However, considering the relationship between global data, the computational complexity is relatively high. The Locally Linear Embedding (LLE) algorithm of local mapping only considers the relationship between local neighborhood points and only ensures that nearby points in the low-dimensional space are still neighboring points, but its algorithm effect is high.

[0056] When reducing the dimensionality of high-dimensional data, there are problems in calculating the intrinsic dimension d of the data and the data neighborhood K. Traditional methods for calculating the intrinsic dimension include maximum likelihood estimation, false nearest neighbors, etc. However, these methods only consider the intrinsic dimension d unidirectionally and do not consider the calculation of the data neighborhood. For the calculation of the data neighborhood K, traditional methods include the K-Nearest Neighbor (KNN) algorithm. However, this type of method cannot obtain a reasonable neighborhood for manifold structures with large curvature and holes. Therefore, it is crucial to find the parameters d and K suitable for the dimensionality reduction of the manifold learning algorithm theory in the fault diagnosis method of the single-phase PWM pulse rectifier. At the same time, it is also crucial to find a suitable manifold learning algorithm to reduce the dimensionality according to the data characteristics of the single-phase PWM pulse rectifier.

[0057] Since the manifold learning algorithm includes a variety of different algorithms, preferably, in the present invention, the Isometric Feature Mapping (ISOMAP) algorithm in the manifold learning theory algorithm, the Locally Linear Embedding (LLE) algorithm representing local mapping, and the Hessian Locally Linear Embedding (Hessian LLE) algorithm are used to reduce the dimensionality of the high-dimensional feature vectors corresponding to the single-phase PWM pulse rectifier. After each algorithm performs dimensionality reduction, a series of low-dimensional feature vectors will be generated. Then, the appropriate manifold learning algorithm, as well as the appropriate values of the intrinsic dimension d and the data neighborhood K, are selected through the clustering criterion.

[0058] During the dimensionality reduction process, since the number of types C of the fault types of the single-phase PWM pulse rectifier is 15, the value of the intrinsic dimension d is d = C - 1, that is, the value of the intrinsic dimension d is 14. At the same time, it is also necessary to determine the value of the data neighborhood K. Preferably, in this embodiment, the value range of the data neighborhood K is set as The implementer can also adjust the value range of the data neighborhood according to the specific situation, where m is the number of samples of each fault type, with a value of 81, and n is the number of all samples, with a value of 1215. The value of the data neighborhood K needs to be taken within this interval. Preferably, the preset interval for taking values in this embodiment is 20. The implementer can also adjust the preset interval for taking values according to the specific situation. At the same time, since the value interval is 20, but The actual calculation result of the value of is set to 401 for the convenience of value taking, that is, 81 ≤ K ≤ 401. When using a manifold learning algorithm to reduce the dimension of high-dimensional feature vectors, the value of the intrinsic dimension d is fixed, and then the high-dimensional feature vectors are reduced in dimension according to the results of the value taking of the data neighborhood K.

[0059] Since there are 17 values for the data neighborhood K, there are 17 corresponding low-dimensional feature vectors obtained after dimension reduction using each manifold learning algorithm. Thus, the low-dimensional feature vectors corresponding to each manifold learning algorithm can be obtained.

[0060] Furthermore, in the case of known classification labels, that is, when determining the optimal clustering, by observing the clustering results of the low-dimensional feature vectors after dimension reduction by each manifold learning algorithm, the suitable dimension reduction algorithm and the data neighborhood K are judged. According to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm, the first index corresponding to different values of the data neighborhood K is calculated, specifically:

[0061] (1) Calculate the within-class average scatter, where Z i represents the class center of class C i ; |C i | represents the number of class samples in class C i , and X represents the samples in class C i .

[0062]

[0063] (2) Calculate the between-class distance, where Z i and Z j respectively represent the class centers of class C i and class C j ,

[0064] d ij = ‖Z i - Z j ‖

[0065] In this formula, the distance between two class centers represents the between-class distance.

[0066] (3) Calculate the first index,

[0067]

[0068] In the formula: k is the number of classifications. Among them, DB k is the first index, and the value of the first index DB k is also the DB Index criterion. Among them, the smaller the value of the first index DB k , the better the clustering effect and the better the dimension reduction effect.

[0069] So far, a first index can be obtained for each value of the different data neighborhood K of each manifold learning algorithm. For example, for the Hessian LLE algorithm, when performing dimensionality reduction, there are 17 values for the data neighborhood K, so there are 17 corresponding first indices.

[0070] Based on the first indices corresponding to the different data neighborhood K values of each manifold learning algorithm, a comparison graph of the dimensionality reduction effects is established, as Figure 3 shown. In the comparison graph of the dimensionality reduction effects, the vertical axis represents the value of the first index, and the horizontal axis represents the different data neighborhood K values of each manifold learning algorithm. Since the dimensionality reduction effects need to be compared, the dimensionality reduction effects of the three dimensionality reduction algorithms can be seen in the graph. From this, it can be seen from the graph that the dimensionality reduction algorithm corresponding to the smallest first index is the Hessian LLE algorithm, indicating that the Hessian LLE algorithm is suitable for the data structure of the high-dimensional feature vectors of the single-phase PWM pulse rectifier, that is, the Hessian LLE algorithm is the optimal manifold learning algorithm. Among them, the LLE algorithm and the ISOMAP algorithm have incorrect dimensionality reduction results when the data neighborhood K is 81 - 161. The main reason is that the value of the data neighborhood K is too small.

[0071] Furthermore, the area with a smaller K value in the comparison graph of the dimensionality reduction effects is locally magnified to obtain a locally magnified line graph of the dimensionality reduction effects, as Figure 4 shown. It should be noted that magnifying the area with a smaller K value in the comparison graph of the dimensionality reduction effects is to clearly illustrate the selection of the optimal data neighborhood K. It can be obtained that when K = 341, the first index is 7.1744, and at this time the first index is the smallest, indicating that under the optimal manifold learning algorithm, when K = 341, the clustering effect is the best. Therefore, in this embodiment, the optimal manifold learning algorithm is the Hessian LLE algorithm. Based on this, when the data neighborhood K takes the value of 341, the dimensionality reduction effect is the best. So far, the optimal manifold learning algorithm and the data neighborhood can be obtained. It should be noted that the embodiments of the present invention only illustrate with three manifold learning algorithms as examples. In the actual implementation process, other manifold learning algorithms not recorded in this embodiment can be added for joint comparison and screening to obtain the manifold learning algorithm most suitable for processing the data of the single-phase PWM pulse rectifier.

[0072] Step S3, use the optimal manifold learning algorithm and the data neighborhood to reduce the dimension of the high-dimensional feature vectors of the current single-phase PWM pulse rectifier to obtain the current low-dimensional feature vectors; input the current low-dimensional feature vectors into the fault diagnosis model to obtain the fault types of the current single-phase PWM pulse rectifier.

[0073] In step S2, the optimal manifold learning algorithm and the data neighborhood K are obtained, which can also be said to be the manifold learning algorithm and the data neighborhood K most suitable for the single-phase PWM pulse rectifier. When it is necessary to perform fault diagnosis on the current single-phase PWM pulse rectifier, the current high-dimensional feature vector is obtained by collecting the current signal of the single-phase PWM pulse rectifier and then processing it. The optimal manifold learning algorithm and the data neighborhood K are used to perform dimensionality reduction processing on the current high-dimensional feature vector to obtain the current low-dimensional feature vector.

[0074] Meanwhile, it is also necessary to construct a fault diagnosis model. Preferably, the fault diagnosis model in the present invention is an SVM support vector machine. Among them, the number of samples collected in step 1 is 1215, and these samples are divided into a training set and a test set according to a preset ratio. The preset ratio is 3:2. Therefore, there are 729 training samples in the training set and 486 test samples in the test set. Among them, there are 15 faults in the 486 test samples, and each fault has 81 test samples. The training set and the test set are used to train the fault diagnosis model to obtain a trained fault diagnosis model. Then, the samples in the test set are used for testing, and it can be obtained that the accuracy rate of fault diagnosis for the single-phase PWM pulse rectifier of high-power AC drive electric locomotives can reach 98.9583%.

[0075] To verify the advantages of the proposed technical solution, the ISOMAP, LLE, and Hessian LLE algorithms are used to perform identification using the fault diagnosis model under the conditions of the same parameters and the same fault sample set. The comparison chart of the accuracy rates of fault diagnosis is as Figure 6 shown, where the fault accuracy rate corresponding to the ISOMAP algorithm is 57.708%, the fault accuracy rate corresponding to the LLE algorithm is 63.75%, and the fault accuracy rate corresponding to the Hessian LLE algorithm is 98.9583%. From this, it can be further determined that the Hessian LLE algorithm is the most suitable algorithm for dimensionality reduction of the data of the PWM pulse rectifier.

[0076] When detecting the current operating state of the PWM pulse rectifier, the obtained current low-dimensional feature vector is input into the trained fault diagnosis model, and the fault type of the current PWM pulse rectifier is output.

[0077] It should be noted that: the above sequence of embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. And the above has described specific embodiments of this specification. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0078] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments.

[0079] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the scope of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fault diagnosis method for single-phase PWM pulse rectifier based on manifold learning algorithm, characterized in that The method includes: Obtaining the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters of the current signal when the single-phase PWM pulse rectifier operates in different operating states respectively; Constructing a high-dimensional feature vector corresponding to each operating state based on the time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency domain characteristic parameters in each operating state; Using different manifold learning algorithms to reduce the dimension of the high-dimensional feature vector to obtain a low-dimensional feature vector corresponding to each manifold learning algorithm; Obtaining a comparison graph of the dimensionality reduction effect according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm; Based on the comparison graph of the dimensionality reduction effect, obtaining the optimal manifold learning algorithm and data neighborhood; Using the optimal manifold learning algorithm and data neighborhood to reduce the dimension of the high-dimensional feature vector of the current single-phase PWM pulse rectifier to obtain the current low-dimensional feature vector; Inputting the current low-dimensional feature vector into the fault diagnosis model to obtain the fault type of the current single-phase PWM pulse rectifier.

2. The fault diagnosis method of a single-phase PWM pulse rectifier based on the manifold learning algorithm according to claim 1, wherein The using different manifold learning algorithms to reduce the dimension of the high-dimensional feature vector to obtain a low-dimensional feature vector corresponding to each manifold learning algorithm includes: determining the intrinsic dimension d according to the number of types of fault types of the single-phase PWM pulse rectifier; setting the value range of the data neighborhood K, and taking values of the data neighborhood at preset intervals within the value range to obtain different values of the data neighborhood K; Based on the intrinsic dimension d and different values of the data neighborhood K, combining different manifold learning algorithms to process the high-dimensional feature vector respectively to obtain a low-dimensional feature vector corresponding to each manifold learning algorithm.

3. A single-phase PWM pulse rectifier fault diagnosis method based on a manifold learning algorithm according to claim 1, characterized in that, The obtaining a comparison graph of the dimensionality reduction effect according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm includes: calculating a first index corresponding to different values of the data neighborhood K according to the clustering results of the low-dimensional feature vectors corresponding to each manifold learning algorithm; Establishing a comparison graph of the dimensionality reduction effect based on the first index corresponding to different values of the data neighborhood K of each manifold learning algorithm, where the vertical axis in the comparison graph of the dimensionality reduction effect represents the value of the first index, and the horizontal axis represents different values of the data neighborhood K of each manifold learning algorithm.

4. A single-phase PWM pulse rectifier fault diagnosis method based on a manifold learning algorithm according to claim 1, characterized in that, The inputting the current low-dimensional feature vector into the fault diagnosis model to obtain the fault type of the current single-phase PWM pulse rectifier includes: Using the optimal manifold learning algorithm and data neighborhood to process the high-dimensional feature vectors corresponding to each operating state to obtain low-dimensional feature vectors corresponding to each operating state; Dividing the low-dimensional feature vectors corresponding to each operating state into a training set and a test set according to a preset ratio; Training the fault diagnosis model using the training set and the test set, and inputting the current low-dimensional feature vector into the trained fault diagnosis model to output the fault type of the current single-phase PWM pulse rectifier.

5. A fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm according to claim 1, characterized in that The time-frequency domain characteristic parameters include: processing the current signal using the DB3 wavelet function to obtain at least two sub-band signal energies; The relative energy ratio of the sub-band signal energy to the total frequency is the time-frequency domain characteristic parameter.

6. A fault diagnosis method for a single-phase PWM pulse rectifier based on a manifold learning algorithm according to claim 1, characterized in that The different operating states include: the normal operating state of the single-phase PWM pulse rectifier and the operating states when different types of faults occur in the single-phase PWM pulse rectifier.

7. A single-phase PWM pulse rectifier fault diagnosis method based on a manifold learning algorithm according to claim 1 or 3, characterized in that, Obtaining the optimal manifold learning algorithm and data neighborhood based on the dimensionality reduction effect comparison graph includes: the manifold learning algorithm and data neighborhood K corresponding to the smallest first index in the dimensionality reduction effect comparison graph are the optimal manifold learning algorithm and data neighborhood.

Images