Fault diagnosis method for series hybrid electric vehicle AC/DC converter

Abstract

A fault diagnosis method for a series hybrid electric vehicle AC/DC (Alternating Current/Direct Current) converter, implementing identifying and diagnosing of an open circuit fault of a power electronic components in an AC/DC converter, and including the following steps: first, establishing a simulation model for a series hybrid electric vehicle AC/DC converter, and selecting a DC bus output current as a fault characteristic; then classifying fault types according to a quantity and locations of faulty power electronic components; next, decomposing the fault characteristic, that is, the DC bus output current by means of fast Fourier transform to different frequency bands, and selecting harmonic ratios of the different frequency bands as fault diagnosing eigenvectors; and finally, identifying the fault types by using a genetic algorithm-based BP (Back Propagation) neural network.

Claims

1. A fault diagnosis method for a series hybrid electric vehicle AC/DC (Alternating Current/Direct Current) converter, adapted to a computer system, the fault diagnosis method comprising the following steps: (1) establishing a simulation model for the series hybrid electric vehicle AC/DC converter, and selecting a DC bus output current as a fault characteristic; (2) classifying fault according to a quantity and locations of faulty power electronic components; (3) decomposing the fault characteristic, that is, the DC bus output current by means of fast Fourier transform to different frequency bands, and selecting harmonic ratios of the different frequency bands as fault diagnosing eigenvectors; and (4) identifying the fault by using a genetic algorithm-based BP (Back Propagation) neural network.

2. The fault diagnosis method for a series hybrid electric vehicle AC/DC converter according to claim 1, wherein the method in the step (3) is specifically: selecting a harmonic ratio of f=30 k Hz as a fault diagnosing eigenvector after comparing a fast Fourier transform analysis result of a DC bus output current waveform in a normal operating state with that of the DC bus output current waveform in a fault state, wherein k=0, 1, 2, 3, . . . , n, and 6n12.

3. The fault diagnosis method for a series hybrid electric vehicle AC/DC converter according to claim 1, wherein the method in the step (4) is specifically: 1) determining a structure of the BP neural network, wherein the BP neural network is constructed as a three-layer network, and there is an approximation relation between a quantity n.sub.2 of hidden-layer neurons and a quantity n.sub.1 of input-layer neurons in the three-layer network:
n.sub.2=2n.sub.1+1, where in the quantity n.sub.1 of the input-layer neurons is an input parameter of the fault diagnosing eigenvector, n.sub.1=n+1, a structure of the three-layer neural network is n.sub.1n.sub.22, and there are (n.sub.1*n.sub.2+2*n.sub.2) weighted values and (n.sub.2+2) thresholds; 2) optimizing an initial weighted value and an initial threshold of the BP neural network by using the genetic algorithm, wherein factors of optimizing the BP neural network by using the genetic algorithm comprise: population initialization, a fitness function, a selection operator, a crossover operator, and a mutation operator, wherein for the population initialization, binary encoding is used for individual encoding, and an input-layer and hidden-layer connection weight, a hidden-layer threshold, a hidden-layer and output-layer connection weight, and an output-layer threshold are comprised; for the fitness function, a norm of an error matrix between a prediction value and an expectation value of a prediction sample is selected as output of a target function, so that a residual between the prediction value and the expectation value is as small as possible when prediction is performed for the BP neural network, and an optimal weighted value and an optimal threshold that enable a value of the target function to be the smallest are obtained; and 3) performing training and prediction for the BP neural network by using the optimized weighted value and the optimized threshold.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a flowchart of a fault diagnosis method according to the present invention.

(2) FIG. 2 is a simulation topology view of an AC/DC converter.

(3) FIG. 3 is a line graph of evolution of a simulation operation result error.

(4) FIG. 4 is a current waveform of DC side in normal running state.

(5) FIG. 5 is a current waveform of DC side in S1 open-circuit fault.

(6) FIG. 6 is a current waveform of DC side in S1 and S4 open-circuit fault.

(7) FIG. 7 is a current waveform of DC side in S3 and S5 open-circuit fault.

DESCRIPTION OF THE EMBODIMENTS

(8) The following describes the technical solutions of the present invention in detail with reference to the accompanying drawings and exemplary embodiments. The following exemplary embodiments are merely used for describing and explaining the present invention, and are not intended to limit the technical solutions of the present invention.

(9) As shown in FIG. 1, FIG. 1 shows a fault diagnosis method for a series hybrid electric vehicle AC/DC converter according to the present invention, including the following steps:

(10) (1) establishing a simulation model for a series hybrid electric vehicle AC/DC converter. Because capacitance existing between a rectifier and an inverter compensates for voltage drop and harmonic changes caused by a fault, normal detection of the fault is affected. In this case, a DC bus output current, that is, a DC current signal I.sub.dc is selected as a fault characteristic. An AC/DC converter is a core device for implementing power conversion and speed adjustment and frequency modulation in an electronic drive system of a series hybrid electric vehicle. A simulation topology structure of the AC/DC converter is shown in FIG. 2, where an AC system source voltage is 220 V, and a frequency is 50 Hz.

(11) (2) classifying fault types according to a quantity and locations of faulty power electronic components. This is specifically as follows:

(12) For an AC/DC converter, a most common power fault is an open circuit fault or single-phase-to-ground fault of a power electronic component. In the present invention, a fault characteristic of open circuit of a power electronic component is focused on and analyzed. For a three-phase voltage-type bridge inverter circuit on the right of the topology structure, an IGBT is used as a switching device, and there are 26 types of switch operating states of six switching devices S1 to S6. The switch operating states are divided into the following types according to locations and a quantity of power electronic components in which the open circuit fault occurs and according to DC bus output current waveforms obtained by means of simulation analysis. As shown in Table 1:

(13) TABLE-US-00001 TABLE 1 Fault types and corresponding faulty waveforms of an AC/DC converter Switch operating Serial Classifi- Faulty state number S1 S2 S3 S4 S5 S6 cation waveform F0 1 T T T T T T G0 as shown in FIG. 4 F1 2 F T T T T T G1 as shown 3 T F T T T T in FIG. 5 . . . . . . . . . . . . . . . . . . . . . 7 T T T T T F F2 8 F T T F T T G2 as shown 9 T F T T F T in FIG. 6 10 T T F T T F 11 F F T T T T G3 as shown 12 F T T T T F in FIG. 7 13 T F F T T T . . . . . . . . . . . . . . . . . . . . . 21 T F T T T F 22 T T T F T F F3 . . . F F F T T T G4 F4 . . . F F F F T T F5 . . . F F F F F T F6 64 F F F F F F

(14) T represents normal operating, F represents an open circuit fault, and F0, F1, F2, F3, F4, F5, and F6 respectively represent a quantity of open circuit faults of 0, 1, 2, 3, 4, 5, and 6. A probability that three or more components concurrently go faulty is smaller than a probability of occurrence of another breaking fault, and is therefore not considered in the present invention. Cases in which all IGBT components in the system have an open circuit fault may be divided into the following types:

(15) G0: A normal operating state, one case;

(16) G1: One component goes faulty, and there are six cases;

(17) G2: Two components (for example, S1 and S4) on arms of a same phase go faulty, and there are three cases;

(18) G3: Two components (for example, S 1 and S6) on different phases go faulty, and there are twelve cases in total;

(19) G4: Three or more components go faulty, and this is not analyzed herein.

(20) (3) decomposing the fault characteristic, that is, the DC bus output current by means of fast Fourier transform to different frequency bands, and selecting signals of 30 k Hz (k=1, 2, 3 . . . ) frequency bands as fault diagnosing eigenvectors. This is specifically as follows:

(21) It can be learned from the analysis in the step (2) that 22 groups of data may be obtained to form training samples of the BP neural network. When fault diagnosing eigenvectors are extracted, 30 Hz is used as a fundamental frequency. Part of DC bus current waveforms in each fault type are selected for performing FFT (Fast Fourier Transform) analysis, and subharmonic contents of f=30 k Hz (k=0, 1, 2, . . . , and 12) are extracted as the fault diagnosing eigenvectors. Training data obtained by performing FFT analysis on the DC bus output currents is shown in the following Table 2:

(22) TABLE-US-00002 TABLE 2 Training data in different fault states Frequency S1 S3 S4 S1, S4 S2, S5 S1, S2 S1, S3 (Hz) Normal breaking breaking breaking breaking breaking breaking breaking 0 2.8597 8.2390 7.1177 7.5581 7.0155 7.9970 9.9290 7.4009 30 0.1669 6.3193 4.8160 4.0456 4.1384 1.3767 3.4836 4.7630 60 3.9712 3.8835 2.2083 2.2223 3.6249 5.5343 0.7156 0.4458 90 0.1122 4.2319 3.0105 2.6492 1.2084 0.4930 0.3539 1.4848 120 2.0621 2.7777 1.9255 1.8339 1.5621 2.5585 0.6082 0.2736 150 0.1089 1.8131 1.3009 1.1428 1.1451 0.8472 0.1884 0.8478 180 1.6674 1.1723 1.1151 1.0331 0.8626 0.7124 0.5259 0.1855 210 0.1105 1.3029 0.6464 0.8501 0.4609 0.6400 0.1503 0.6022 240 1.6739 2.0188 1.1467 1.5259 1.1334 1.8903 0.4749 0.1467 270 0.1133 2.4830 1.4668 1.6632 0.9357 0.8431 0.1412 0.4516 300 1.4735 3.1508 2.1771 2.3908 2.9160 3.6816 0.4133 0.1171 330 0.1119 2.7194 1.7360 1.8667 1.1589 0.9404 0.1148 0.3582 360 1.3163 2.9033 2.2725 2.2030 2.7039 3.5276 0.4037 0.1156

(23) Because four fault modes in total are analyzed in the embodiments of the present invention, the various fault types may be represented in the following forms:

(24) No-fault state: G0 (0, 0);

(25) One component goes faulty: G1 (0, 1);

(26) Two components on a same phase go faulty: G2 (1, 0);

(27) Two components on different phases go faulty: G3 (1, 1).

(28) To test the trained network, the following groups of data are provided as test data of the network, as listed in Table 3:

(29) TABLE-US-00003 TABLE 3 Sample data for testing Frequency S6 S3, S6 S2, S3 S3, S5 (Hz) breaking breaking breaking breaking 0 7.5309 6.4599 8.8885 8.3371 30 4.7168 2.6588 2.3623 1.9969 60 2.3967 3.9707 0.4260 1.0671 90 2.7036 0.8541 0.2546 0.1589 120 1.7640 1.7102 0.5712 0.6593 150 1.0405 0.5784 0.1456 0.1678 180 0.6489 0.8373 0.5201 0.4691 210 0.8494 0.5697 0.1208 0.1601 240 1.2571 1.1196 0.4788 0.3687 270 1.6721 0.5981 0.1074 0.1575 300 2.0927 2.6680 0.4074 0.2988 330 1.7469 1.2904 0.0910 0.1514 360 1.9631 2.8361 0.4047 0.2773

(30) (4) identifying the fault types by using a genetic algorithm-based BP neural network. This is specifically as follows:

(31) Optimizing the BP neural network by using the genetic algorithm mainly includes three sections: BP neural network structure determining, weighted value and threshold optimization by using the genetic algorithm, and BP neural network training and prediction.

(32) 1) determining a structure of the BP neural network.

(33) A mode identifying problem usually occurs during construction of the BP neural network. This problem can be well resolved by using a three-layer network. It can be learned from the step (3) that there are 13 input parameters of the samples, that is, a quantity n.sub.1 of input-layer neurons is 13, and there are two output parameters. It can be calculated according to an approximation relation n.sub.2=2n.sub.1+1 that a quantity n.sub.2 of hidden-layer neurons is 27. Therefore, a structure of the set three-layer neural network is 13272, and there are 13*27+2*27=405 weighted values and 27+2=29 thresholds, and a quantity of parameters optimized by using the genetic algorithm is 405+29=434. In addition, a transfer function of the hidden-layer neuron of the BP neural network uses an S-type tangent function, and a transfer function of an output-layer neuron uses an S-type logarithmic function.

(34) 2) optimizing an initial weighted value and an initial threshold of the BP neural network by using the genetic algorithm.

(35) Generally, weighted values and thresholds of a neural network are random numbers in a range of [0.5, 0.5] that are obtained by performing random initialization. The initialization parameters significantly affect training of the network and cannot be accurately obtained. Training results of the network are the same for same initial weighted values and same initial thresholds. The genetic algorithm is introduced so that an optimal weighted value and an optimal threshold are obtained. The optimizing the BP neural network by using the genetic algorithm is actually optimizing the initial weighted value and the initial threshold of the BP neural network by using the genetic algorithm.

(36) Factors of optimizing the BP neural network by using the genetic algorithm include: population initialization, a fitness function, a selection operator, a crossover operator, and a mutation operator, where for the population initialization, binary encoding is used for individual encoding, and an input-layer and hidden-layer connection weight, a hidden-layer threshold, a hidden-layer and output-layer connection weight, and an output-layer threshold are included; for the fitness function, in the present invention, a norm of an error matrix between a prediction value and an expectation value of a prediction sample is selected as output of a target function, so that a residual between the prediction value and the expectation value is as small as possible when prediction is performed for the BP neural network.

(37) In the present invention, assuming that encoding of each of the weighted value and the threshold is a 10-bit binary number, a binary encoding length of an individual is 4340. For specific operating parameters, refer to Table 4:

(38) TABLE-US-00004 TABLE 4 Setting of operating parameters of a genetic algorithm Largest Quantity of Size of quantity of binary digit Crossover Mutation Gener- popu- gener- bits of proba- proba- ation lation ations variable bility bility gap 40 50 10 0.7 0.01 0.95

(39) 3) performing training and prediction for the BP neural network by using the optimized weighted value and the optimized threshold.

(40) For training and testing of the BP neural network, a process of training data for the neural network is a process of constantly adjusting unknown parameters to obtain a minimum value of a cost function. A training function trains a network by using a Levenberg-Marquardt algorithm, to constantly modify a weighted value and a threshold, so that an output error of the network is the smallest and accuracy of a prediction result is ensured.

(41) According to theories of the genetic algorithm and the BP neural network, a fault diagnosis method of a BP neural network based on the genetic algorithm is implemented by programming in MATLAB software. The part of genetic algorithm uses a Sheffield genetic algorithm toolbox, and the part of BP neural network uses a neural network toolbox of the MATLAB. An operating result is shown in FIG. 3, and output results of the algorithm are optimized weighted value and threshold matrices and a minimum error of a prediction result. It can be learned from FIG. 3 that the minimum error of the prediction result of the BP neural network optimized by using the genetic algorithm is 0.033.

(42) Comparisons between a prediction result of the BP neural network using a random weighted value and a random threshold and a prediction result of a test sample using an optimized weighted value and an optimized threshold are as follows:

(43) TABLE-US-00005 TABLE 5 Comparison between simulation results S6 S3, S6 S2, S3 S3, S5 Simulation breaking breaking breaking breaking error Theoretic output (0, 1) (1, 0) (1, 1) (1, 1) A prediction result of a (0.1874, 0.9891) (0.9230, 0.0902) (0.8474, 0.9770) (0.8987, 0.9487) 0.27308 test sample using a random weighted value and a random threshold A prediction result of a (0.0126, 0.9704) (0.9831, 0.0002) (0.9991, 0.9981) (0.9991, 0.9980) 0.033114 test sample using an optimized weighted value and an optimized threshold

(44) It can be learned from Table 5 that an error generated by a prediction result of the BP neural network using a weighted value and a threshold that are optimized by using the genetic algorithm is apparently smaller than an error generated by a prediction result using a random weighted value and a random threshold. Therefore, by using the neural network that is optimized by using the genetic algorithm, open circuit fault types of all components in an electronic circuit can be effectively and accurately diagnosed.

(45) The present invention is described above by using examples with reference to the accompanying drawings, and has modifications and variations in structure and arrangement. Therefore, all equivalent technical solutions shall also fall within the scope of the present invention, and insubstantial improvements of the ideas and solution of the present invention shall fall within the protection scope of the present invention.