Bidirectional LLC resonant converter fault diagnosis method based on CSA-DBN

Abstract

The invention discloses a bidirectional LLC resonant converter fault diagnosis method based on CSA-DBN, and belongs to the technical field of multi-fault diagnosis. The problems that an existing fault diagnosis method based on DBN is low in diagnosis precision and long in time consumption in the model parameter optimization process are solved. According to the method, the fault features are selected through the random forest classifier, then the RBM is pre-trained based on the selected fault features, the multi-layer pre-trained RBM is stacked to form the DBN, and the hidden layer node number and the learning rate of the DBN are optimized by adopting the crow search algorithm, so that the problem of long time consumption in the parameter initialization process can be solved, and the parameter initialization efficiency is improved. On the basis of the optimized hidden layer node number and learning rate, the DBN is trained by using the constructed fault sample data set, so that the learning and classification capabilities of the trained DBN can be improved, and the fault diagnosis precision of the DBN can be improved. The method can be applied to fault diagnosis of the bidirectional LLC resonant converter.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a fault diagnosis method for bidirectional LLC resonant converters based on CSA-DBN. Background Technology

[0002] Satellites are widely used in scientific exploration, communication, tracking, and navigation. The satellite power system, responsible for supplying power to the satellite's electrical payloads, is a complex, large-scale system with strong nonlinearity and numerous control logic and protection mechanisms. A failure in the satellite power system can lead to the loss of power to certain individual electrical components, loss of payload functionality, or even the failure of the entire satellite. The battery charge and discharge regulator (BCDR) is responsible for storing energy from the bus in the battery or releasing energy from the battery back to the bus. It typically consists of a main circuit, namely a power converter, and a control circuit. The power converter, responsible for step-up and step-down voltage conversion, is a crucial component of the power system, bearing electrical and thermal stresses far greater than those of the control circuit, thus its failure probability is much higher than other parts of the circuit. The bidirectional full-bridge LLC resonant converter, as a novel and key topology in power charge and discharge regulators, has broad application prospects in the field of satellite charge and discharge regulators due to its high reliability at high power and high efficiency in a small size. Charge-discharge regulators are typical power electronic conversion devices. Failures in their components are sudden, with few warning signs beforehand, and are prone to causing severe open-circuit or short-circuit faults, directly leading to power system failure and posing a significant hazard. Therefore, research on data-driven fault diagnosis technology for BCDR main topology bidirectional LLC resonant converters can provide technical support for the digital development of next-generation satellite power systems, aligning with the trend of satellite intelligence.

[0003] Fault diagnosis of satellite charge / discharge regulators mainly focuses on power converters, and fault diagnosis of power converters primarily focuses on circuit components. To date, researchers have conducted research on fault diagnosis and prediction techniques for power electronic circuits, such as Buck and Boost circuits, covering topics such as power device fault feature extraction, device parameter identification, and failure mechanism analysis. However, research on fault diagnosis techniques for BCDR and LLC resonant converters is relatively limited. Methods such as BP neural networks, support vector machines, and extreme learning machines have been widely applied in fault diagnosis and have shown good results. Furthermore, as a deep learning method, Deep Belief Network (DBN) is a commonly used data-driven fault diagnosis method. Compared to shallow networks such as BP, SVM, and ELM, which mainly classify and distinguish a few typical fault modes, DBN can directly take the original signal as input and directly obtain the fault diagnosis result. This solves the problem that when there are many types of fault modes, the fault characteristics of different fault types may have certain similarities, which may lead to the less-than-ideal accuracy of traditional shallow networks such as BP, SVM, and ELM in fault diagnosis. DBN has superior data analysis and processing capabilities and better recognition effect.

[0004] DBN (Deep Binary Neural Network) is a multi-layer neural network model with high-dimensional data training capabilities. It learns and recognizes features from input data by training multi-layer Restricted Boltzmann Machines (RBMs) and has been successfully applied in the field of fault diagnosis. However, the fault diagnosis performance of DBN largely depends on the initial parameter settings, which often rely on experience and require repeated experiments to achieve the desired results, resulting in low efficiency. Therefore, intelligent optimization algorithms can be used to optimize the parameters of DBN fault diagnosis models, improving their diagnostic performance and accuracy to some extent. For example, improved fruit fly optimization algorithms, improved hybrid frog jumping algorithms, and multi-strategy improved quantum differential evolution algorithms have all yielded better DBN parameters, effectively improving the diagnostic accuracy of DBN for various mechanical faults. However, when these methods were used to optimize DBN and experimentally validated for fault diagnosis, it was found that the computational time was long, and there was still room for further improvement in diagnostic accuracy. Summary of the Invention

[0005] The purpose of this invention is to address the problems of low diagnostic accuracy and long model parameter optimization time in existing DBN-based fault diagnosis methods, and to propose a CSA-DBN-based fault diagnosis method for bidirectional LLC resonant converters.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN, the method specifically including the following steps:

[0007] Step 1: Select several voltage signals as fault characteristics in the bidirectional LLC resonant converter. In each fault mode and in the fault-free mode, the time-series voltage values ​​of each selected voltage signal are collected.

[0008] Step 2: Select fault features based on the random forest and the data collected in Step 1;

[0009] Step 3: Standardize the data corresponding to the fault features selected in Step 2 to establish a fault sample dataset;

[0010] Step 4: Divide the fault sample dataset into two parts: a training set and a test set;

[0011] Step 5: Pre-train the RBM using the training set without fault type labels to obtain a pre-trained RBM.

[0012] Step 6: Stack the M layers of pre-trained RBMs to form a pre-trained DBN network;

[0013] Step 7: Use the CSA algorithm to optimize the number of hidden layer nodes and the learning rate of the pre-trained DBN network;

[0014] Step 8: Load the optimized number of hidden layer nodes and learning rate into the pre-trained DBN network, train the pre-trained DBN network using the labeled training set, and determine whether the recognition accuracy of the labeled test set on the trained DBN network converges.

[0015] If the recognition accuracy of the labeled test set on the trained DBN network does not converge, then based on the current trained DBN network parameters and CSA algorithm, return to step seven to continue optimizing the number of hidden layer nodes and the learning rate of the DBN network.

[0016] If the recognition accuracy of the labeled test set converges on the trained DBN network, the final trained DBN network is obtained. Save the parameters of the final trained DBN network and then execute step nine.

[0017] Step 9: In the bidirectional LLC resonant converter to be identified, collect the time-series voltage data corresponding to the fault features selected in Step 2, process the collected time-series voltage data to obtain the fault feature vector to be identified, use the fault feature vector to be identified as the input of the finally trained DBN network, and output the fault identification result of the bidirectional LLC resonant converter through the finally trained DBN network.

[0018] Furthermore, the selected fault characteristics specifically include: input voltage U i The voltage U to ground at the midpoint of bridge arm 1 a The voltage U across Lm1 b Transformer secondary output voltage U c The voltage U across capacitor Cr d The voltage U across Lm2 e Overall output voltage U o .

[0019] Furthermore, the fault modes include short-circuit faults and open-circuit faults in the components inside the bidirectional LLC resonant converter.

[0020] Furthermore, the specific process of step three is as follows:

[0021] For failure mode a:

[0022] Step 3: For any fault feature selected in Step 2, extract the middle part of the data from the time-series voltage data of that fault feature under fault mode a.

[0023] Step 32: Perform wavelet packet decomposition on the extracted data, obtain the wavelet packet energy spectrum, and use the wavelet packet energy corresponding to each data point in the extracted data to form the wavelet packet energy vector of the fault feature.

[0024] Step 33: After traversing each fault feature selected in Step 2 using the methods in Step 31 and Step 32, a fault feature vector is formed by using the wavelet packet energy vector corresponding to each fault feature selected in Step 2. After normalizing the elements in the fault feature vector, a fault sample data is obtained, and a fault type label is added to the fault sample data.

[0025] Furthermore, the value of M is 3 or 5.

[0026] Furthermore, the specific process of step seven is as follows:

[0027] Step 71: Initialize decision variables including the number of hidden layer nodes and the learning rate. Initialize the position constraints of the decision variables, the population size N, and the maximum number of iterations to It. max Flight length fl and self-perceived probability AP;

[0028] Initialize the positions of N crows (Crows):

[0029]

[0030] in, Let represent the initial values ​​of the 1st, 2nd, ..., dth decision variables in the position of the nth crow, where d represents the number of decision variables, and n = 1, 2, ..., N;

[0031] Initialize the memory for N crows:

[0032]

[0033] Step 72: Calculate the fitness function value for each crow based on its initial position;

[0034] Step 73: Initialize the number of iterations k = 0;

[0035] Step 74: The i-th crow randomly selects another crow, denoted as j, and generates a random number r between 0 and 1. j,k The i-th crow according to r j,k Initial update position x′ i,k+1 ;

[0036] Step 75: Calculate the fitness function value corresponding to the initial updated position of each crow, and determine the final updated position x of each crow in this iteration based on the calculated fitness function value. i,k+1 ;

[0037] And according to x i,k+1 Update the memory of N crows;

[0038] Step 76: Determine if the number of iterations satisfies k = It max -1;

[0039] If k = It max If the value is -1, then the position with the largest fitness function value is selected from the positions of each crow after the last iteration, and the values ​​of each decision variable contained in the selected position are used as the optimization result.

[0040] If k = It is not satisfied max If -1, then let k = k + 1, and return to step seven four.

[0041] Furthermore, the specific process of step seven-four is as follows:

[0042]

[0043] in, x represents i,k The values ​​of each decision variable in the data, r i,k A random number between 0 and 1 Represents x′ i,k+1The values ​​of each decision variable in the equation are given by b, which is a randomly generated position within the position constraints of the decision variables, i = 1, 2, ..., N.

[0044] Furthermore, in step seven-five, the fitness function value corresponding to the initial updated position of each crow is calculated, and the final updated position x of each crow in this iteration is determined based on the calculated fitness function value. i,k+1 Specifically:

[0045]

[0046] Where f(·) represents the fitness function value, i = 1, 2, ..., N.

[0047] Furthermore, in step seven-five, according to x i,k+1 Update the memory of N crows, specifically:

[0048]

[0049] in, m i,k+1 =x i,k+1 .

[0050] The beneficial effects of this invention are:

[0051] This invention first selects fault features using a random forest classifier, then pre-trains the RBM based on the selected fault features, and uses stacked multi-layer pre-trained RBMs to form a DBN network. A raven search algorithm is then used to optimize the number of hidden layer nodes and the learning rate of the DBN network, which solves the problem of long parameter initialization time. Based on the optimized number of hidden layer nodes and learning rate, the DBN network is trained using a constructed fault sample dataset, which improves the learning and classification capabilities of the trained DBN network, thereby enhancing the accuracy of fault diagnosis.

[0052] The performance of the model in this invention was tested using a real dataset of multiple samples and multiple faults simulated by actual hardware. The experimental results show that the present invention optimizes the number of nodes in the hidden layer of the DBN network and the learning rate of the model based on the Raven Search algorithm, which can enhance the feature recognition ability of the model and improve the accuracy and efficiency of fault diagnosis. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of a bidirectional LLC resonant converter;

[0054] Figure 2 It is a rectangle plot showing the importance of each feature. Detailed Implementation

[0055] Specific Implementation Method 1: This implementation method describes a fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN. It utilizes a DBN network for model training and testing, and simultaneously optimizes the parameters of the DBN network using a raven search algorithm. The method specifically includes the following steps:

[0056] Step 1, such as Figure 1 The diagram shows a schematic of a bidirectional LLC resonant converter. The bidirectional LLC resonant converter includes a primary circuit, a transformer, and a secondary circuit. Both the primary and secondary sides employ a full-bridge structure, consisting of two H-bridges. The high-voltage end of the transformer is connected to the bus, and the low-voltage end is connected to the battery. The primary side has four NMOS transistors Si1, Si2, Si3, and Si4, with corresponding parasitic capacitances Ci1, Ci2, Ci3, and Ci4, respectively. The secondary side also has four NMOS transistors So1, So2, So3, and So4, with corresponding parasitic capacitances Co1, Co2, Co3, and Co4, respectively. The magnetizing inductor is Lm1, the resonant inductor is Lr, and the resonant capacitor is Cr. The midpoints of bridge arms 1, 2, 3, and 4 are points A, B, C, and D, respectively. Capacitors Ci and Co are output filter capacitors. An auxiliary inductor Lm2 is added between points C and D. When the additional inductance on the secondary side of the transformer is equivalent to the inductance on the primary side and equal to the value of the magnetizing inductance on the primary side, the converter topology is completely symmetrical and can achieve bidirectional energy flow.

[0057] The primary-side circuit includes bridge arm 1, bridge arm 2, inductor Lm1, input filter capacitor Ci, and primary-side windings; bridge arm 1 includes NMOS transistors Si1 and Si2, and bridge arm 2 includes NMOS transistors Si3 and Si4; one end of capacitor Ci is connected to the positive input / output terminal of the primary-side circuit, and the other end of capacitor Ci is connected to the negative input / output terminal of the primary-side circuit; the drains of NMOS transistors Si1 and Si4 are connected to the positive input / output terminal of the primary-side circuit; NMOS transistor Si1... The source of NMOS transistor Si3 is connected to the drain of NMOS transistor Si4; the source of NMOS transistor Si4 is connected to the drain of NMOS transistor Si2; the sources of NMOS transistor Si3 and NMOS transistor Si4 are simultaneously connected to the negative input and output terminals of the primary circuit; one end of inductor Lm1 is connected to the source of NMOS transistor Si1, and the other end of inductor Lm1 is connected to the drain of NMOS transistor Si2; the same-name terminal of the primary winding is connected to the source of NMOS transistor Si1, and the opposite-name terminal of the primary winding is connected to the drain of NMOS transistor Si2.

[0058] The secondary circuit includes bridge arm 3, bridge arm 4, inductor Lm2, output filter capacitor Co, resonant capacitor Cr, resonant inductor Lr, and secondary winding; bridge arm 3 includes NMOS transistor So1 and MOS transistor So2, and bridge arm 4 includes NMOS transistor So3 and NMOS transistor So4; one end of the output filter capacitor Co is connected to the positive input / output terminal of the secondary circuit, and the other end of the output filter capacitor Co is connected to the negative input / output terminal of the secondary circuit; the drains of NMOS transistor So1 and MOS transistor So4 are connected to the positive input / output terminal of the secondary circuit; the source of NMOS transistor So1 and the source of NMOS transistor So4 are connected to the positive input / output terminal of the secondary circuit. The source of transistor So4 is connected to the two ends of inductor Lm2; the source of NMOS transistor So1 is connected to the drain of NMOS transistor So3; the source of NMOS transistor So4 is connected to the drain of NMOS transistor So2; the sources of NMOS transistor So3 and NMOS transistor So2 are simultaneously connected to the negative input and output terminals of the secondary circuit; one end of resonant inductor Lr is connected to the source of NMOS transistor So1, the other end of resonant inductor Lr is connected to one end of resonant capacitor Cr, the other end of resonant capacitor Cr is connected to the opposite-named terminal of the secondary winding, and the same-named terminal of the secondary winding is connected to the source of NMOS transistor So4.

[0059] It should be noted that, under certain circumstances, the input and output relationships at both ends of the transformer can be interchanged.

[0060] To achieve non-invasive power converter fault diagnosis, several voltage signals were selected as fault characteristics within the bidirectional LLC resonant converter. The acquisition locations of each voltage signal used as a fault characteristic are shown in Table 1.

[0061] Table 1 Location of Fault Feature Extraction

[0062] name Corresponding collection location <![CDATA[ Ui ]]> Input voltage <![CDATA[ Ua ]]> Voltage at point A to ground <![CDATA[ Ub ]]> <![CDATA[Voltage at points A and B (L m1 Voltage)]]> <![CDATA[ Uc ]]> Transformer secondary output voltage <![CDATA[ Ud ]]> <![CDATA[Capacitance C r voltage across both ends]]> <![CDATA[ Ue ]]> <![CDATA[Voltage between points C and D (L m2 Voltage)]]> <![CDATA[ Uo ]]> Output voltage

[0063] The selected fault characteristics specifically include: input voltage U i The voltage U to ground at the midpoint of bridge arm 1 a The voltage U across Lm1 b Transformer secondary output voltage U c The voltage U across capacitor Cr d The voltage U across Lm2 e Overall output voltage U o .

[0064] In each fault mode, the timing voltage values ​​of each selected voltage signal are collected. The fault modes include short-circuit faults and open-circuit faults of each component inside the bidirectional LLC resonant converter. In the fault-free mode, the timing voltage values ​​of each selected voltage signal are collected.

[0065] The common failure modes are shown in Table 2:

[0066] Table 2 Structural Failure Modes of Bidirectional Full-Bridge LLC Resonant Converters

[0067]

[0068] Among them, fault number SF1 indicates an open circuit fault in inductor Lm2, fault number SF2 indicates a short circuit fault in inductor Lm2, fault number SF3 indicates an open circuit fault in resonant inductor Lr, fault number SF4 indicates a short circuit fault in resonant inductor Lr, fault number SF5 indicates an open drain fault in NMOS transistor Si1, fault number SF6 indicates a short circuit fault in the gate-drain capacitance of MOS transistor Si1, fault number SF7 indicates a short circuit fault in the gate-source capacitance of MOS transistor Si1, fault number SF8 indicates an open circuit fault in resonant capacitor, fault number SF9 indicates a short circuit fault in resonant capacitor, fault number SF10 indicates an open circuit fault in input capacitor, and fault number SF11 indicates an open circuit fault in output capacitor. Simultaneously, data at various voltage points under some other uncommon fault modes were also collected. The classification labels for these uncommon fault modes were uniformly marked as "other faults," and the classification labels for data under fault-free conditions were marked as "normal." In each fault mode, all other components of the bidirectional LLC resonant converter function normally except for the specified fault. For example, under fault number SF1, all components function normally except for an open-circuit fault in inductor Lm2. Data was acquired under the simulation model operating in the 80V-100V voltage range and with an output power of 2kW. Under the same fault mode, 70,000 data points of voltage signal were collected at each acquisition point, and the collected data was used in subsequent structural fault diagnosis algorithms.

[0069] Step 2: Select fault features based on the data collected in Step 1 and the random forest. Specifically, use the average reduction of impurity (usually Gini impurity or information gain) method in the random forest to perform importance analysis on all fault features, and then select the four fault features with higher importance.

[0070] Step 3: Standardize the data corresponding to the fault features selected in Step 2 to establish a fault sample dataset;

[0071] For fault mode a, the specific process of step three is as follows:

[0072] Step 3: For any fault feature selected in Step 2, extract the middle part of the data from the time-series voltage data of the fault feature under fault mode a, that is, extract the data from the 30,000th to the 40,000th data points from the 70,000 data points.

[0073] Step 32: Perform wavelet packet decomposition on the extracted data, obtain the wavelet packet energy spectrum, and use the wavelet packet energy corresponding to each data point in the extracted data to form the wavelet packet energy vector of the fault feature.

[0074] Step 33: After traversing each fault feature selected in Step 2 using the methods in Step 31 and Step 32, a fault feature vector is formed by using the wavelet packet energy vector corresponding to each fault feature selected in Step 2. After normalizing the elements in the fault feature vector, a fault sample data is obtained, and a fault type label is added to the fault sample data.

[0075] The method in step three is used to process the data under various fault modes and the data under no-fault conditions to obtain the established fault sample dataset.

[0076] Step 4: Divide the fault sample dataset into two parts, a training set and a test set, in an 8:2 ratio;

[0077] Step 5: Pre-train the RBM using the training set without fault type labels to obtain a pre-trained RBM. The pre-training process is an unsupervised greedy layer-by-layer training process.

[0078] Step 6: Stack the M pre-trained RBMs to form a pre-trained DBN network. Each RBM is a pre-trained RBM from Step 5. Initialize the basic parameters of the DBN network (number of network layers, number of network training iterations, activation function type, etc.). In this invention, M is 3 or 5. When M = 3, the number of hidden layer nodes is 2. When M = 5, the number of hidden layer nodes is 4.

[0079] Step 7: Use the CSA algorithm to optimize the number of hidden layer nodes and the learning rate of the pre-trained DBN network; using the CSA algorithm to optimize the DBN network can greatly improve the optimization gradient of the DBN network parameters.

[0080] The specific process of step seven is as follows:

[0081] Step 71: Initialize decision variables including the number of hidden layer nodes and the learning rate (when the DBN network is formed by stacking 3 pre-trained RBM layers, the number of hidden layers is 2; when the DBN network is formed by stacking 5 pre-trained RBM layers, the number of hidden layers is 4), and initialize the decision variables position constraints, population size N, and maximum number of iterations It. max Flight length fl and self-perceived probability AP;

[0082] Initialize the positions of N crows (Crows):

[0083]

[0084] in, Let represent the initial values ​​of the 1st, 2nd, ..., dth decision variables in the position of the nth crow, where d represents the number of decision variables, and n = 1, 2, ..., N;

[0085] Since the crows have no experience in the initial iteration, we assume that their food is hidden at the initial position and initialize the memory of N crows:

[0086]

[0087] Among them, the initial memory (Memory) is the same as the initialization location (Crows);

[0088] Step 72: Calculate the fitness function value of each crow based on the initial position. That is, calculate the recognition accuracy of the DBN network on the training set when the parameter of the DBN network is the position of each crow, and use the calculated accuracy as the fitness function value of each crow.

[0089] Step 73: Initialize the number of iterations k = 0;

[0090] Step 74: The i-th crow randomly selects another crow, denoted as j, and generates a random number r between 0 and 1. j,k The i-th crow according to r j,k Initial update position x′ i,k+1 :

[0091]

[0092] in, x represents i,k The values ​​of each decision variable in the data, r i,k A random number between 0 and 1 Represents x′ i,k+1 In the context of each decision variable value, b is a randomly generated position within the positional constraints of the decision variables, i = 1, 2, ..., N;

[0093] Step 75: Calculate the fitness function value corresponding to the initial updated position of each crow, and determine the final updated position of each crow in this iteration based on the calculated fitness function value.

[0094]

[0095] Where f(·) represents the fitness function value, i = 1, 2, ..., N;

[0096] According to x i,k+1Update the memory of N crows so that the memory is the same as the position after the final update in this iteration:

[0097]

[0098] Step 76: Determine if the number of iterations satisfies k = It max -1;

[0099] If k = It max If the value is -1, then the position with the largest fitness function value is selected from the positions of each crow after the last iteration, and the values ​​of each decision variable contained in the selected position are used as the optimization result.

[0100] If k = It is not satisfied max If the value is -1, then let k = k + 1 and return to step seven four.

[0101] Step 8: Load the optimized number of hidden layer nodes and learning rate into the pre-trained DBN network. Train the pre-trained DBN network using the labeled training set. That is, use the labeled training data to perform supervised backpropagation fine-tuning of the parameters of the entire DBN network. This process fine-tunes the network parameters from the last layer to the previous layer by using known label information, thereby reducing the error between the network output and the actual label, improving the classification accuracy, and further optimizing the DBN network model training results until the set maximum number of iterations is reached. Determine whether the recognition accuracy of the labeled test set on the trained DBN network has converged.

[0102] If the recognition accuracy of the labeled test set on the trained DBN network does not converge, then based on the current trained DBN network parameters and CSA algorithm, return to step seven to continue optimizing the number of hidden layer nodes and the learning rate of the DBN network.

[0103] If the recognition accuracy of the labeled test set converges on the trained DBN network, the final trained DBN network is obtained. Save the parameters of the final trained DBN network and then execute step nine.

[0104] Step 9: In the bidirectional LLC resonant converter to be identified, collect the time-series voltage data corresponding to the fault features selected in Step 2. Process the collected time-series voltage data to obtain the fault feature vector to be identified. Use the fault feature vector to be identified as the input of the finally trained DBN network. Output the fault identification result of the bidirectional LLC resonant converter through the finally trained DBN network. The identification result may be various common types of faults, other types of faults, or normal state (i.e., no fault).

[0105] Experimental Section

[0106] The Random Forest Classifier is used to evaluate the importance of features, such as... Figure 2 As shown, we have drawn a bar chart of feature importance to visually demonstrate the contribution of each feature to the model's predictive ability. It can be seen that the four features [Ui, Uo, Ud, Ue] are more important than other features.

[0107] The test results of the CSA-DBN algorithm of this invention were compared under three feature combinations. The model was trained and tested separately for each of the three feature combinations. Except for the different feature combinations, no changes were made to the model. As can be seen from Table 3, when all seven features were loaded into the model, the test accuracy was 95.39%; when [Ui, Uo, Ua, Ud] was loaded, the test accuracy was 98.07%; and when [Ui, Uo, Ud, Ue] was loaded, the test accuracy was 99.51%.

[0108] Table 3 Comparison of Fault Classification Results with Different Feature Combinations

[0109]

[0110] All three options can meet more than 95% of the hardware requirements, but the feature combination [Ui,Uo,Ud,Ue] can achieve extremely high accuracy and greatly reduce the occurrence of errors. This proves that this feature selection is more suitable for the model designed in this invention, so these four features will continue to be used in subsequent model analysis.

[0111] Next, the CSA-DBN algorithm of this invention is compared with the fault diagnosis algorithms of SVM, DBN, ADAM-DBN and the cyclic learning rate warm-up strategy-DBN. The results are shown in Table 4:

[0112] Table 4 Comparison of Fault Diagnosis Results of Different Algorithms

[0113]

[0114] When the data volume is relatively large and complex, the DBN model structure is better suited to handle this situation compared to SVM. We used three improvement strategies based on DBN, each with its own advantages. For example, compared with the recurrent learning rate warm-up strategy - DBN, the ADAM optimizer can achieve model convergence in a shorter time (98.45%). The recurrent learning rate warm-up strategy can achieve a relatively higher accuracy (98.78%), but it also takes the longest time.

[0115] The CSA-DBN model used in this invention not only optimizes multiple parameters to improve model results, but also, compared to the iterative learning rate warm-up strategy, achieves better test results in fewer iterations, reducing training time. Furthermore, the optimal model parameters obtained through the CSA algorithm can be directly applied to models that do not use the CSA algorithm, further reducing training time. Therefore, the trained DBN model can be directly used for fault identification.

[0116] In summary, the CSA-DBN algorithm proposed in this invention can accurately identify multiple fault modes in bidirectional LLC resonant converters, demonstrating the feasibility and effectiveness of the method. Since the output of the DBN network is related to the setting of the hidden layer, parameter optimization is required in practical engineering applications to effectively identify the causes of power electronic system faults and provide a scientific and reasonable technical means for the system to return to normal.

[0117] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN, characterized in that, The method specifically includes the following steps: Step 1: Select several voltage signals as fault characteristics in the bidirectional LLC resonant converter. In each fault mode and in the fault-free mode, the time-series voltage values ​​of each selected voltage signal are collected. Step 2: Select fault features based on the random forest and the data collected in Step 1; Step 3: Standardize the data corresponding to the fault features selected in Step 2 to establish a fault sample dataset; Step 4: Divide the fault sample dataset into two parts: a training set and a test set; Step 5: Pre-train the RBM using the training set without fault type labels to obtain a pre-trained RBM. Step 6: Stack the M layers of pre-trained RBMs to form a pre-trained DBN network; Step 7: Use the CSA algorithm to optimize the number of hidden layer nodes and the learning rate of the pre-trained DBN network; Step 8: Load the optimized number of hidden layer nodes and learning rate into the pre-trained DBN network, train the pre-trained DBN network using the labeled training set, and determine whether the recognition accuracy of the labeled test set on the trained DBN network converges. If the recognition accuracy of the labeled test set on the trained DBN network does not converge, then based on the current trained DBN network parameters and CSA algorithm, return to step seven to continue optimizing the number of hidden layer nodes and the learning rate of the DBN network. If the recognition accuracy of the labeled test set converges on the trained DBN network, the final trained DBN network is obtained. Save the parameters of the final trained DBN network and then execute step nine. Step 9: In the bidirectional LLC resonant converter to be identified, collect the time-series voltage data corresponding to the fault features selected in Step 2, process the collected time-series voltage data to obtain the fault feature vector to be identified, use the fault feature vector to be identified as the input of the finally trained DBN network, and output the fault identification result of the bidirectional LLC resonant converter through the finally trained DBN network.

2. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 1, characterized in that, The selected fault characteristics specifically include: input voltage U i The voltage U to ground at the midpoint of bridge arm 1 a The voltage U across Lm1 b Transformer secondary output voltage U c The voltage U across capacitor Cr d The voltage U across Lm2 e Overall output voltage U o .

3. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 2, characterized in that, The fault modes include short-circuit faults and open-circuit faults in the components inside the bidirectional LLC resonant converter.

4. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 3, characterized in that, The specific process of step three is as follows: For failure mode a: Step 3: For any fault feature selected in Step 2, extract the middle part of the data from the time-series voltage data of that fault feature under fault mode a. Step 32: Perform wavelet packet decomposition on the extracted data, obtain the wavelet packet energy spectrum, and use the wavelet packet energy corresponding to each data point in the extracted data to form the wavelet packet energy vector of the fault feature. Step 33: After traversing each fault feature selected in Step 2 using the methods in Step 31 and Step 32, a fault feature vector is formed by using the wavelet packet energy vector corresponding to each fault feature selected in Step 2. After normalizing the elements in the fault feature vector, a fault sample data is obtained, and a fault type label is added to the fault sample data.

5. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 4, characterized in that, The value of M is 3 or 5.

6. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 5, characterized in that, The specific process of step seven is as follows: Step 71: Initialize decision variables including the number of hidden layer nodes and the learning rate. Initialize the position constraints of the decision variables, the population size N, and the maximum number of iterations to It. max Flight length fl and self-perceived probability AP; Initialize the positions of N crows (Crows): in, Let represent the initial values ​​of the 1st, 2nd, ..., dth decision variables in the position of the nth crow, where d represents the number of decision variables, and n = 1, 2, ..., N; Initialize the memory for N crows: Step 72: Calculate the fitness function value for each crow based on its initial position; Step 73: Initialize the number of iterations k = 0; Step 74: The i-th crow randomly selects another crow, denoted as j, and generates a random number r between 0 and 1. j,k The i-th crow according to r j,k Initial update position x′ i,k+1 ; Step 75: Calculate the fitness function value corresponding to the initial updated position of each crow, and determine the final updated position x of each crow in this iteration based on the calculated fitness function value. i,k+1 ; And according to x i,k+1 Update the memory of N crows; Step 76: Determine if the number of iterations satisfies k = It max -1; If k = It max If the value is -1, then the position with the largest fitness function value is selected from the positions of each crow after the last iteration, and the values ​​of each decision variable contained in the selected position are used as the optimization result. If k = It is not satisfied max If -1, then let k = k + 1, and return to step seven four.

7. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 6, characterized in that, The specific process of step seven-four is as follows: in, x represents i,k The values ​​of each decision variable in the data, r i,k A random number between 0 and 1 Represents x′ i,k+1 The values ​​of each decision variable in the equation are given by b, which is a randomly generated position within the position constraints of the decision variables, i = 1, 2, ..., N.

8. The fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 7, characterized in that, In step seven-five, the fitness function value corresponding to the initial updated position of each crow is calculated, and the final updated position x of each crow in this iteration is determined based on the calculated fitness function value. i,k+1 Specifically: Where f(·) represents the fitness function value, i = 1, 2, ..., N.

9. A fault diagnosis method for a bidirectional LLC resonant converter based on CSA-DBN according to claim 8, characterized in that, In step seven five, according to x i,k+1 Update the memory of N crows, specifically: in, m i,k+1 =x i,k+1 .

Images