An Online Fault Diagnosis Method for Dual Active Bridge Converters Based on Multi-Source Data Fusion

Abstract

This invention discloses an online fault diagnosis method for dual active bridge converters based on multi-source data fusion, comprising: S1, establishing a simulation model of the dual active bridge converter under test, analyzing the relationship between different open-circuit fault states of the dual active bridge converter and changes in circuit signals, selecting diagnostic signals, and collecting diagnostic signals under different open-circuit fault states of the dual active bridge converter to construct a multi-source data sequence set; S2, spatially reconstructing the multi-source data sequence set to obtain a multi-scale reconstructed data matrix; S3, temporally dividing the multi-scale reconstructed data matrix into two continuous matrices, upscaling the two continuous matrices to a higher dimension, and merging them with the matrices before the upscaling to obtain two high-dimensional matrices; S4, establishing a Koopman model based on the high-dimensional matrix and extracting the modal parameters of the Koopman model; S5, inputting the Koopman modal parameters into a trained wavelet time-scattering convolutional network to obtain the diagnostic results.

Description

Technical Field

[0001] This invention relates to the field of power electronic converter fault diagnosis technology, specifically to an online fault diagnosis method for a dual active bridge converter with multi-source data fusion. Background Technology

[0002] With the rapid development of power electronics technology, dual active bridge converters (DABs) have become key components in various application areas, such as power systems, electric vehicles, and renewable energy systems. This topology is characterized by enabling bidirectional flow of electrical energy, thus allowing for efficient energy conversion and exchange between various energy sources and loads. However, DABs are frequently affected by various internal and external factors during operation, leading to failures. This not only negatively impacts system performance but may also threaten the reliability and safety of the equipment.

[0003] Power semiconductor devices play a crucial role in converters because they are responsible for controlling the flow and conversion of electrical energy. Although the causes of power equipment failures vary, they are all closely related to operating conditions and circuit parameters. Investigations show that power electronic devices such as insulated-gate bipolar transistors (IGBTs) and their gate drivers are prone to failure, especially open-circuit faults. This is because, compared to short-circuit faults, which trigger overcurrent protection, open-circuit faults cannot be better protected by integrated gate driver circuits and are difficult to detect and locate. Summary of the Invention

[0004] The purpose of this invention is to provide an online fault diagnosis method for dual active bridge converters based on multi-source data fusion, which improves the fault tolerance of open-circuit fault diagnosis by integrating multiple fault diagnosis signals.

[0005] To address the aforementioned technical problems, this invention provides an online fault diagnosis method for dual active bridge converters based on multi-source data fusion, comprising:

[0006] S1. Establish a simulation model of the dual active bridge converter under test. By analyzing the relationship between different open-circuit fault states of the dual active bridge converter and changes in circuit signals, select diagnostic signals and collect diagnostic signals under different open-circuit fault states of the dual active bridge converter to construct a multi-source data sequence set.

[0007] S2. Spatial reconstruction of the multi-source data sequence set to obtain a multi-scale reconstructed data matrix;

[0008] S3. Divide the multi-scale reconstructed data matrix into two continuous matrices in time. After raising the two continuous matrices to a higher dimension, merge them with the matrices before the dimension increase to obtain two high-dimensional matrices.

[0009] S4. Establish a Koopman model based on the high-dimensional matrix and extract the modal parameters of the Koopman model;

[0010] S5. Input the modal parameters into the trained wavelet time scattering convolutional network to obtain the diagnostic results.

[0011] According to the above scheme, the diagnostic signals include leakage inductance current, primary side bridge arm midpoint voltage, and secondary side bridge arm midpoint voltage.

[0012] According to the above scheme, step S2 includes: using the number of data sequences from different sources in the diagnostic signal as rows of the multi-scale reconstructed data matrix, and using the length of a single-source data sequence in the diagnostic signal as a column of the multi-scale reconstructed data matrix.

[0013] According to the above scheme, in step S3, the two consecutive matrices are raised to a higher dimension using a deep neural network algorithm with a trainable dictionary.

[0014] According to the above scheme, in step S4, the modal parameters of the Koopman model are extracted through EDMD dynamic mode decomposition.

[0015] According to the above scheme, the deep neural network algorithm for the trainable dictionary is trained through a fully connected neural network.

[0016] This invention also provides a system for online fault diagnosis of dual active bridge converters based on multi-source data fusion, comprising:

[0017] The diagnostic signal selection module is used to analyze the relationship between different open-circuit fault states of the dual active bridge converter and changes in circuit signals based on the established simulation model of the dual active bridge converter under test, and to select diagnostic signals.

[0018] The multi-source data sequence set construction module is used to collect diagnostic signals under different open-circuit fault states of the dual active bridge converter to construct a multi-source data sequence set;

[0019] The multi-scale reconstructed data matrix acquisition module is used to spatially reconstruct a multi-source data sequence set to obtain a multi-scale reconstructed data matrix.

[0020] The high-dimensional matrix acquisition module is used to divide the multi-scale reconstructed data matrix into two continuous matrices in time. After the two continuous matrices are upgraded to high dimensions, they are merged with the matrices before the upgrade to obtain two high-dimensional matrices.

[0021] The Koopman modal parameter extraction module is used to build a Koopman model and extract the modal parameters of the Koopman model.

[0022] The diagnostic result acquisition module is used to input the Koopman mode parameters into the trained wavelet time scattering convolutional network to obtain the diagnostic results.

[0023] According to the above scheme, the diagnostic signals include leakage inductance current, primary side bridge arm midpoint voltage, and secondary side bridge arm midpoint voltage.

[0024] According to the above scheme, the method for the multi-scale reconstruction data matrix acquisition module to acquire the multi-scale reconstruction data matrix includes: taking the number of data sequences from different sources in the diagnostic signal as the rows of the multi-scale reconstruction data matrix, and taking the length of a single-source data sequence in the diagnostic signal as the columns of the multi-scale reconstruction data matrix.

[0025] According to the above scheme, the high-dimensional matrix acquisition module uses a deep neural network algorithm with a trainable dictionary to upscale two consecutive matrices to a higher dimension.

[0026] Beneficial effects

[0027] This invention constructs a multi-source data sequence set based on diagnostic signals under different open-circuit faults, and performs diagnosis based on the multi-source data sequence set, realizing the integration of different fault characteristics, thereby enabling the method to more accurately identify different types of open-circuit faults. By using Koopman modal parameters for analysis, it not only provides the static structure of the dual active bridge converter system, but also captures the dynamic behavior of the system, making the dynamic system analysis more interpretable and practical. The wavelet time scattering convolutional network used, where different orders of wavelet scattering correspond to the extraction of different order features of the signal, can adaptively select appropriate filters to match the spectral characteristics of different signals, and more accurately identify each fault. Attached Figure Description

[0028] Figure 1 This is a flowchart of multi-signal fault diagnosis combining the Koopman operator and wavelet time scattering convolutional network according to Embodiment 1 of the present invention.

[0029] Figure 2 (a) is a simulation schematic diagram of the bidirectional active full-bridge DC-DC converter topology according to Embodiment 1 of the present invention;

[0030] Figure 2 (b) is a circuit diagram of the IGBT driving circuit of Embodiment 1 of the present invention;

[0031] Figure 2 (c) is the output drive waveform diagram of the IGBT drive circuit of Embodiment 1 of the present invention;

[0032] Figure 3 This is the mean square error variation curve of the deep neural network of the trainable dictionary in Embodiment 1 of the present invention.

[0033] Figure 4 (a) is a distribution diagram of modal frequency and attenuation rate, and characteristic roots under fault category F2 in Embodiment 1 of the present invention;

[0034] Figure 4 (b) is a distribution diagram of modal frequency and attenuation rate, and characteristic roots under fault category F3 in Embodiment 1 of the present invention;

[0035] Figure 5 In the middle (a) to (d), the scattering transformation diagrams of lev=1 under fault categories F1 to F4 of Embodiment 1 of the present invention are respectively.

[0036] Figure 6 (a) is a graph showing the change in training accuracy of the wavelet time scattering convolutional network in Embodiment 1 of the present invention;

[0037] Figure 6 (b) is a loss variation curve of the wavelet time scattering convolutional network in Embodiment 1 of the present invention;

[0038] Figure 7 It is the confusion matrix of the test set of the wavelet time-scattering convolutional network in Embodiment 1 of the present invention;

[0039] Figure 8 This is a flowchart of the online fault diagnosis method for dual active bridge converters with multi-source data fusion according to Embodiment 1 of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0041] Example 1:

[0042] See Figure 8 An online fault diagnosis method for a dual active bridge converter with multi-source data fusion includes:

[0043] S1. Establish a simulation model of the dual active bridge converter under test. By analyzing the relationship between different open-circuit fault states and circuit signal changes in the dual active bridge converter, select diagnostic signals and collect diagnostic signals under different open-circuit fault states to construct a multi-source data sequence set; the multi-source data sequence set is represented as follows. ,in ;

[0044] S2. Spatial reconstruction is performed on the multi-source data sequence set to obtain a multi-scale reconstructed data matrix; the multi-scale reconstructed data matrix is ​​represented as follows: ;

[0045] S3. Divide the multi-scale reconstructed data matrix into two continuous matrices in time. After raising the two continuous matrices to a higher dimension, merge them with the matrices before the dimension increase to obtain two high-dimensional matrices.

[0046] S4. Establish a Koopman model based on the high-dimensional matrix and extract the modal parameters of the Koopman model;

[0047] S5. Input the modal parameters into the trained wavelet time scattering convolutional network to obtain the diagnostic results.

[0048] In this embodiment, the simulation circuit diagram of the established dual active bridge converter simulation model is shown below. Figure 2 , Figure 2 (b) The IGBT drive circuit model is used to generate pulse modulation (PWM) signals with a drive voltage of 15V / -8V. It is based on characteristic parameter modeling to achieve extended phase shift control. Figure 2 (c) shows the drive waveform output by the IGBT drive circuit. An open-circuit fault is simulated by disconnecting the IGBT's PWM signal. The fault states and categories of the DAB converter are shown in Table 1.

[0049] Table 1 Fault Status and Category

[0050]

[0051] Furthermore, the simulation model of the dual active bridge converter was established in PSpice.

[0052] Furthermore, the diagnostic signals include leakage inductance current, primary-side bridge arm midpoint voltage, and secondary-side bridge arm midpoint voltage;

[0053] Because the DAB converter has a symmetrical structure, the primary output voltage will be affected by an open-circuit fault in the IGBT at the symmetrical position. With secondary output voltage and the leakage inductance current of the transformer With identical waveforms, it's impossible to accurately locate IGBT faults in symmetrical positions, especially when combined with the bridge arm midpoint voltage. , and leakage current As a diagnostic signal, it is used to study the open-circuit fault diagnosis of all IGBTs in the DAB converter;

[0054] Therefore, extract the voltage at the midpoint of the bridge arm. , and leakage current In this embodiment, 500 points were collected for each sample, and the three time series were spatially reconstructed.

[0055] Further, step S2 includes: using the number of data sequences from different sources in the diagnostic signal as rows of the multi-scale reconstructed data matrix, and using the length of a single-source data sequence in the diagnostic signal as a column of the multi-scale reconstructed data matrix.

[0056] Furthermore, in step S3, the two consecutive matrices are raised to a higher dimension using a deep neural network algorithm with a trainable dictionary; the deep neural network algorithm is trained using a fully connected neural network.

[0057] Multi-scale reconstruction of data matrix The two continuous matrices that are divided along the time dimension are represented as follows:

[0058]

[0059] in , It is the data collection time step place A column vector of dimension;

[0060] In this embodiment, a 3-layer fully connected neural network is trained to train the dictionary function.

[0061]

[0062]

[0063] in, For network weights, For bias, For activation function, It is a function for increasing dimensionality;

[0064] The loss function of a fully connected neural network is defined as follows:

[0065]

[0066] In this embodiment, a deep neural network algorithm is used to achieve the dimensionality increase process. The neural network consists of two hidden layers, each with 50 nodes, and the activation function is the tanh function. The best results are achieved after 66 steps, with optimal validation performance at a mean squared error of [value missing]. hour, Figure 3 The figure shows the mean squared error curve during neural network training. The regression coefficients of the training set, validation set, and test set of this network are all around 0.96, indicating that the output value is highly correlated with the actual value.

[0067] The deep neural network training algorithm for trainable dictionaries is summarized as follows:

[0068] Step 1: Status data acquisition;

[0069] Step 2: Initialize the Koopman operator And the parameters of a fully connected neural network Input training data, and set the maximum training time, batch size, and maximum training error. ;

[0070] Step 3: Fix parameters ,optimization ;

[0071] Step 4: Fix Optimize parameters Repeatedly update through optimization algorithms and parameters (i.e., repeat steps 3-4);

[0072] Step 5: Determine whether the maximum training time has been reached, or whether the training error is satisfied. :

[0073] Yes, end training and save the network parameters of the deep neural network.

[0074] No, proceed to step 3.

[0075] Furthermore, in step S4, the modal parameters of the Koopman model are extracted through EDMD dynamic mode decomposition.

[0076] Two continuous matrices, when raised to a higher dimension, are represented as follows: , By merging the matrices before dimensionality increase, the high-dimensional matrix is ​​represented as follows: , ;

[0077] The finite-dimensional approximate estimation expression for the Koopman operator is:

[0078]

[0079] In the formula It is a generalized inverse, and eigenvalues Approximate values ​​for the Koopman eigenvalues ​​are provided.

[0080] Figure 4The distribution of eigenvalues ​​of the Koopman operator under partially open-circuit fault conditions, as well as the decay rate and frequency of modes, are shown. The higher the energy, the darker the corresponding eigenvalue. Koopman mode analysis not only provides the static structure of the data but also captures the dynamic behavior of the system. In a given time series data, each mode obtained in the calculation is associated with a fixed oscillation frequency and decay / growth rate. Its modal analysis has stronger interpretability and applicability in dynamic system modeling and analysis, especially for the complex dynamic behavior contained in time series data.

[0081] This example uses a wavelet time-scattering convolutional network under different fault conditions. See the scattering transformation diagram. Figure 5 Features are automatically obtained through wavelet scattering transform, where the number of scattering paths is 7, and the scattering sequence is transposed and reshaped.

[0082] For training the wavelet time-scattering convolutional network, see [link to training documentation]. Figure 1 The system includes: the samples are divided into training and test sets in a 7:3 ratio; the classifier based on the scattering convolutional network follows an end-to-end framework; in order to update and optimize the model parameters more effectively, both first-moment and second-moment estimations are used; the learning rate is adaptively adjusted according to the nature of the gradient; the learning rate is changed at specific intervals by using a segmented learning rate plan; the learning rate is reduced after every 5 training epochs to help optimize the weights more slowly in the later stages of training; the mini-batch size is set to 15, which is the number of samples used in each iteration of gradient descent; and the maximum training epoch is 110.

[0083] In wavelet time-scattering convolutional networks, each network group includes convolution, batch normalization (BN), activation, and pooling layers. Finally, a softmax layer is used to classify the input vector for faults. The training set accuracy versus loss curves and the confusion matrix of the test set during the iteration process are shown below. Figure 6 and Figure 7 As shown;

[0084] The DAB converter fault diagnosis model established in this embodiment based on the wavelet time scattering convolutional network can directly process multiple diagnostic signals. It extracts the denoised depth features by performing a series of convolution, modulus operation and downsampling operations through the wavelet scattering path, and inputs them into the scattering convolutional neural network to complete the fault diagnosis.

[0085] The wavelet operator can be represented as:

[0086] ,in, For the largest scale, The scaling function The wavelet mode represents the scattering path; it consists of two parts: the first is the translation invariant. One is called the zero scattering order, and the other is the nonlinear translation covariate. ;

[0087] Repeat the iteration until At the order of 1, we have:

[0088]

[0089]

[0090]

[0091]

[0092] The iterative formula above shows that the... Covariant part of order Through a low-pass filter The average yielded translation invariants After wavelet The convolutional model effect yields covariates Let the wavelet subscript sequence be... , represents a scattering path, Representing the The scattering coefficients of order 0 also represent the number of wavelet filters; the scattering expressions range from order 0 to order 1. All scattering coefficients of order 1 constitute, i.e.:

[0093]

[0094] The obtained multi-scale information is divided into sample sets and input into the classifier. While realizing the automatic acquisition of high-order dynamic information, it can weaken intra-class differences and adaptively reduce the influence of noise. A scattering convolutional network structure is constructed, and the trained fault diagnosis model is tested using a test set to realize fault diagnosis of dual active converters.

[0095] This embodiment proposes an online fault diagnosis method for dual active bridge converters based on multi-source data fusion. It utilizes the Koopman operator and wavelet time scattering network to address the multi-signal problem in converter fault diagnosis. Compared with existing data-driven methods, the proposed method offers higher diagnostic accuracy, lower data requirements, and more stable diagnostic performance. The Koopman operator can directly examine the evolution of correlation functions in the state space from observables while preserving the low-frequency spatiotemporal characteristics of the correlation, thus simplifying the analysis of multi-signal faults. Furthermore, different orders of wavelet scattering correspond to the extraction of different order features of the signal, enabling adaptive selection of appropriate filters to match the spectral characteristics of different signals, and thus more accurately identifying various faults.

[0096] Example 2:

[0097] Based on Embodiment 1, this embodiment provides a system for online fault diagnosis of dual active bridge converters using multi-source data fusion, comprising:

[0098] The diagnostic signal selection module is used to analyze the relationship between different open-circuit fault states of the dual active bridge converter and changes in circuit signals based on the established simulation model of the dual active bridge converter under test, and to select diagnostic signals.

[0099] The multi-source data sequence set construction module is used to collect diagnostic signals under different open-circuit fault states of the dual active bridge converter to construct a multi-source data sequence set;

[0100] The multi-scale reconstructed data matrix acquisition module is used to spatially reconstruct a multi-source data sequence set to obtain a multi-scale reconstructed data matrix.

[0101] The high-dimensional matrix acquisition module is used to divide the multi-scale reconstructed data matrix into two continuous matrices in time. After the two continuous matrices are upgraded to high dimensions, they are merged with the matrices before the upgrade to obtain two high-dimensional matrices.

[0102] The Koopman modal parameter extraction module is used to build a Koopman model and extract the modal parameters of the Koopman model.

[0103] The diagnostic result acquisition module is used to input the Koopman mode parameters into the trained wavelet time scattering convolutional network to obtain the diagnostic results.

[0104] Furthermore, the diagnostic signals include leakage inductance current, primary side bridge arm midpoint voltage, and secondary side bridge arm midpoint voltage.

[0105] Furthermore, the method for obtaining the multi-scale reconstructed data matrix by the multi-scale reconstructed data matrix acquisition module includes: using the number of data sequences from different sources in the diagnostic signal as the rows of the multi-scale reconstructed data matrix, and using the length of a single-source data sequence in the diagnostic signal as the columns of the multi-scale reconstructed data matrix.

[0106] Furthermore, the high-dimensional matrix acquisition module uses a deep neural network algorithm with a trainable dictionary to upscale two consecutive matrices to a higher dimension.

[0107] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0108] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online fault diagnosis of a dual active bridge converter using multi-source data fusion, characterized in that, include: S1. Establish a simulation model of the dual active bridge converter under test. By analyzing the relationship between different open-circuit fault states of the dual active bridge converter and changes in circuit signals, select diagnostic signals and collect diagnostic signals under different open-circuit fault states of the dual active bridge converter to construct a multi-source data sequence set. S2. Spatial reconstruction of the multi-source data sequence set to obtain a multi-scale reconstructed data matrix; S3. Divide the multi-scale reconstructed data matrix into two continuous matrices in time. After raising the two continuous matrices to a higher dimension, merge them with the matrices before the dimension increase to obtain two high-dimensional matrices. S4. Establish a Koopman model based on the high-dimensional matrix and extract the modal parameters of the Koopman model; S5. Input the modal parameters into the trained wavelet time scattering convolutional network to obtain the diagnostic results.

2. The online fault diagnosis method for dual active bridge converters based on multi-source data fusion according to claim 1, characterized in that, Diagnostic signals include leakage inductance current, primary side bridge arm midpoint voltage, and secondary side bridge arm midpoint voltage.

3. The online fault diagnosis method for dual active bridge converters based on multi-source data fusion according to claim 1, characterized in that, Step S2 includes: using the number of data sequences from different sources in the diagnostic signal as rows of the multi-scale reconstructed data matrix, and using the length of a single-source data sequence in the diagnostic signal as a column of the multi-scale reconstructed data matrix.

4. The online fault diagnosis method for dual active bridge converters based on multi-source data fusion according to claim 1, characterized in that, In step S3, the two consecutive matrices are raised to a higher dimension using a deep neural network algorithm with a trainable dictionary.

5. The online fault diagnosis method for dual active bridge converters based on multi-source data fusion according to claim 1, characterized in that, In step S4, the modal parameters of the Koopman model are extracted through EDMD dynamic mode decomposition.

6. The online fault diagnosis method for a dual active bridge converter based on multi-source data fusion according to claim 4, characterized in that, The deep neural network algorithm for the trainable dictionary is trained using a fully connected neural network.

7. A system for online fault diagnosis of a dual active bridge converter using multi-source data fusion, characterized in that, include: The diagnostic signal selection module is used to analyze the relationship between different open-circuit fault states of the dual active bridge converter and changes in circuit signals based on the established simulation model of the dual active bridge converter under test, and to select diagnostic signals. The multi-source data sequence set construction module is used to collect diagnostic signals under different open-circuit fault states of the dual active bridge converter to construct a multi-source data sequence set; The multi-scale reconstructed data matrix acquisition module is used to spatially reconstruct a multi-source data sequence set to obtain a multi-scale reconstructed data matrix. The high-dimensional matrix acquisition module is used to divide the multi-scale reconstructed data matrix into two continuous matrices in time. After the two continuous matrices are upgraded to high dimensions, they are merged with the matrices before the upgrade to obtain two high-dimensional matrices. The Koopman modal parameter extraction module is used to build a Koopman model and extract the modal parameters of the Koopman model. The diagnostic result acquisition module is used to input modal parameters into a trained wavelet time scattering convolutional network to obtain diagnostic results.

8. The online fault diagnosis method system for dual active bridge converters based on multi-source data fusion according to claim 7, characterized in that, Diagnostic signals include leakage inductance current, primary side bridge arm midpoint voltage, and secondary side bridge arm midpoint voltage.

9. The online fault diagnosis method system for dual active bridge converters based on multi-source data fusion according to claim 7, characterized in that, The method for obtaining the multi-scale reconstructed data matrix by the multi-scale reconstructed data matrix acquisition module includes: using the number of data sequences from different sources in the diagnostic signal as the rows of the multi-scale reconstructed data matrix, and using the length of a single-source data sequence in the diagnostic signal as the columns of the multi-scale reconstructed data matrix.

10. The online fault diagnosis method system for dual active bridge converters based on multi-source data fusion according to claim 7, characterized in that, The high-dimensional matrix acquisition module uses a deep neural network algorithm with a trainable dictionary to upscale two consecutive matrices to a higher dimension.

Images