Transformer fault diagnosis model construction method based on deep learning network
By building a transformer fault diagnosis model based on deep learning network, using dissolved gas characteristic information in transformer oil in nuclear power plant, optimizing the hyperparameters of DBN network, the misjudgment and misjudgment problems in the fault diagnosis of main transformer of nuclear power plant are solved, the diagnostic accuracy is improved, and the equipment is safe.
Patent Information
- Application Number
- CN202510334414.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-20
- Publication Date
- 2025-07-08
AI Technical Summary
The existing technology has misjudgment and misjudgment in the fault diagnosis of main transformers of nuclear power plants. The shallow machine learning method lacks feature mining capabilities in the era of power big data, which affects the accuracy of fault diagnosis, resulting in equipment damage and unplanned downtime.
The transformer fault diagnosis model based on deep learning network is adopted. By building a DBN network, combining Logistic chaotic mapping, improved DBO algorithm and Levy flight strategy, the hyperparameters of the DBN network are optimized, the transformer deep learning model is constructed, and the failure identification is used to use the dissolved gas characteristic information in the transformer oil of the nuclear power plant for fault identification.
It improves the accuracy of fault diagnosis evaluation, effectively prevents leakage and misjudgment of nuclear power transformers, and ensures safe and stable operation of the equipment.
Smart Images
Figure CN120277484A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of transformers, and in particular to a method for constructing a transformer fault diagnosis model based on a deep learning network. Background Art
[0002] As the only key equipment in the conventional island electrical equipment of a nuclear power plant that is subject to nuclear safety supervision, the main transformer of a nuclear power plant is not only a hub for power transmission between the nuclear power plant and the external power grid, but also the first external power source to ensure nuclear safety. Therefore, it is crucial to ensure its reliable operation. Therefore, establishing an accurate and reliable fault diagnosis model to timely detect potential faults in the main transformer of a nuclear power plant and perform targeted maintenance based on its health status is of great significance for ensuring the safe and stable operation of the nuclear power plant and the external power grid.
[0003] Dissolved Gas Analysis (DGA) can perform on-line monitoring and off-line manual oil sampling under the normal operation state of the transformer, and is not affected by the complex external electromagnetic field. It has now become the preferred measure for transformer fault diagnosis. At present, the operation and maintenance personnel of nuclear power plants usually use traditional methods such as the IEC three-ratio method and the Duval triangle method to diagnose transformer faults, which has played a positive role in discovering potential transformer faults. However, in engineering practice, problems such as incomplete coding, overly absolute boundaries, and poor fault recognition effects have also emerged.
[0004] With the development of artificial intelligence technology, many machine learning methods such as expert systems, artificial neural networks, and support vector machines have been applied in the field of fault diagnosis and have achieved certain results. However, they all belong to shallow machine learning methods and have disadvantages such as lack of feature mining ability and insufficient learning ability in the era of big power data, which affect the further improvement of the accuracy of transformer fault diagnosis, are prone to misjudgment and missed judgment, and pose a threat to the safe and stable operation of the transformer. Summary of the Invention
[0005] In order to solve the above problems, the present invention provides a method for constructing a transformer fault diagnosis model based on a deep learning network, which can improve the evaluation accuracy of fault diagnosis and effectively prevent equipment damage and unplanned shutdown losses caused by missed judgment and misjudgment of nuclear power transformers.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0007] A method for constructing a transformer fault diagnosis model based on a deep learning network includes the following steps:
[0008] S1. Obtain the data of the dissolved gases in the transformer oil and form the original data of DGA;
[0009] S2. Construct the original data into a fault feature set of the transformer. Through normalizing and non-linearly reducing the dimension of the fault feature set, an input feature vector is extracted and used as sample data.
[0010] S3. Divide the sample data into a pre-training set, a tuning set, and a test set.
[0011] S4. Establish a DBN network and initialize the parameters of the DBN network.
[0012] S5. Improve the DBO algorithm, and the improved DBO algorithm is used for hyperparameter optimization of the DBN network to construct a deep learning model of the transformer.
[0013] S6. Obtain the training set and the tuning set in step S3, and train the deep learning model of the transformer through the training set and the tuning set to obtain a trained deep learning model of the transformer.
[0014] S7. Diagnose the test set in step S3 through the trained deep learning model of the transformer, and evaluate the trained deep learning model of the transformer. Adjust the trained deep learning model of the transformer according to the evaluation results to obtain a fault diagnosis model.
[0015] Further, in step S2, construct the fault feature set of the transformer according to the characteristic gases corresponding to different fault types in the transformer oil of the nuclear power plant, and according to the dissolved gas content and ratio of different fault types in the transformer oil of the nuclear power plant.
[0016] Further, in step S2, the method for normalizing the fault feature set is as follows:
[0017]
[0018] where X i is the concentration value of the i-th gas; x i is the normalized concentration value of the i-th gas, X i,max is the maximum value of the i-th gas; X i,min is the minimum value of the i-th gas.
[0019] Use the KPCA algorithm for non-linear dimensionality reduction of the normalized fault feature set. Through the non-linear mapping kernel function map the normalized fault feature set X = (x1, x2, x3…, x n ) from the low-dimensional space to the high-dimensional space G, and use the principal component analysis method in the high-dimensional space G to extract the principal components of the data, and use the newly generated comprehensive variables as the input feature vector.
[0020] Furthermore, in step S5, the method for constructing the transformer deep learning model includes the following steps:
[0021] S5.1 Initialize the population position of the DBN network through Logistic chaotic mapping, and set the population size and the search ranges of the number of iterations, the number of times of DBN backpropagation training, the number of neurons in each hidden layer, and the learning rate;
[0022] S5.2 Construct a fitness function using the fault recognition error rate, and update the individual positions of the rolling dung beetle, the breeding dung beetle, the small dung beetle, and the thief dung beetle in sequence through the improved DBO algorithm, and calculate the fitness value after the position update of the dung beetle;
[0023] S5.3 Sort the fitness values, and use the individual position of the dung beetle corresponding to the minimum fitness value as the optimal parameter of the DBN network to construct the transformer deep learning model.
[0024] Furthermore, in step S5.1, the method for initializing the population position of the DBN network through Logistic chaotic mapping is as follows:
[0025] X i+1 = μX i (1 - X i ) Formula (2)
[0026] where X i is the mapping function value at the i-th iteration, μ is the mapping parameter, and when μ ∈ [3.569, 4], the Logistic mapping distribution trajectory exhibits completely chaotic characteristics.
[0027] Furthermore, in step S5.2, the individual position of the rolling dung beetle is updated using the improved sine method strategy:
[0028]
[0029] where δ = rand(1); ST ∈ (0.5, 1]; x i (t) is the i-th position component of individual X at the t-th iteration; p i (t) is the i-th component of the best individual position variable at the t-th iteration; r2 is a random number in the interval [0, 2π]; r3 is a random number in the interval [-2, 2].
[0030] Furthermore, in step 5.2, the value set by the self-adaptive variable inertia weight strategy, the self-adaptive variable inertia weight ω t linearly decreases as the number of iterations increases as follows:
[0031]
[0032] Among them, ω t is a self-adaptive variable inertia weight;
[0033] The value of r1 is set in a non-linear decreasing mode, and the cosine function between 0 and π is used to determine the change of the value of r1:
[0034] Among them, w max is the maximum value of w t ; w min is the minimum value of w t ; t represents the current iteration number; T max represents the maximum iteration number.
[0035] Furthermore, in step S5.3, the sorting of the fitness values is used to determine the optimal position of the dung beetle individuals through Levy flight:
[0036]
[0037] Among them, is point-to-point multiplication; α is the step size control quantity, and Levy(λ) is the search path where the step size follows the Levy distribution;
[0038] A greedy rule is added during Levy flight to determine whether to update the target position; and the current optimal fitness value is updated through iteration to obtain the position of the dung beetle individual corresponding to the minimum fitness value.
[0039] Furthermore, in step S5.3, it is judged whether the maximum iteration number is reached. When the maximum iteration number is not reached, it returns to continue iterative update. When the maximum iteration number is reached, the transformer deep learning model is obtained.
[0040] Furthermore, in step S6, the group size, iteration number, DBN backpropagation training number, the number of neurons in each hidden layer, and the optimization range of the learning rate are given to the transformer deep learning model, and the transformer deep learning model is trained through the training set and the tuning set.
[0041] The beneficial effects of the present invention are:
[0042] Based on the gas generation mechanism of dissolved gases in large power transformers of nuclear power plants, by analyzing the characteristic gases corresponding to different fault types in the transformer oil of nuclear power plants, as well as the content and ratio of dissolved gases of different fault types in the transformer oil of nuclear power plants, the fusion characteristic gas method and ratio method are used to fully extract the fault characteristic information of the transformer, construct a 16-dimensional transformer fault characteristic set, and use the KPCA algorithm to perform non-linear feature extraction and dimensionality reduction on the fault characteristic matrix to form the input feature vector of the diagnostic model; combined with various strategies such as Logistic chaotic mapping, improved sine strategy, Levy flight and greedy rules to comprehensively improve the DBO algorithm, so as to balance the global search and local development capabilities of the algorithm, effectively improve the convergence performance and global optimization ability of the algorithm; use the improved DBO algorithm to comprehensively optimize the number of reverse training times, the number of neurons in each hidden layer and the learning rate of the DBN deep learning network, and assign the optimized parameters to the DBN network, thereby constructing a fault diagnosis model, which can effectively identify six operating states: partial discharge, low-energy discharge, high-energy discharge, medium-low temperature overheating, high temperature overheating and normal, improve the evaluation accuracy of fault diagnosis, and effectively prevent equipment damage and unplanned shutdown losses caused by missed and misjudged of nuclear power transformers. Description of the Drawings
[0043] Figure 1 It is a flow chart of a method for constructing a transformer fault diagnosis model based on a deep learning network according to a preferred embodiment of the present invention.
[0044] Figure 2 It is a Pareto chart of kernel principal component analysis of the KPCA algorithm of a method for constructing a transformer fault diagnosis model based on a deep learning network according to a preferred embodiment of the present invention.
[0045] Figure 3 It is a schematic diagram of the DBN network of a method for constructing a transformer fault diagnosis model based on a deep learning network according to a preferred embodiment of the present invention.
[0046] Figure 4 It is a flow chart of the improved DBO algorithm of a method for constructing a transformer fault diagnosis model based on a deep learning network according to a preferred embodiment of the present invention.
[0047] Figure 5 It is a confusion matrix diagram of the fault diagnosis result of a method for constructing a transformer fault diagnosis model based on a deep learning network according to a preferred embodiment of the present invention. Detailed Embodiments
[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.
[0050] Please also refer to Figures 1 to 5 , a method for constructing a transformer fault diagnosis model based on a deep learning network in a preferred embodiment of the present invention includes the following steps:
[0051] S1. Obtain the data of dissolved gases in transformer oil and form the original data of DGA. In this embodiment, the original data can be obtained by collecting the data of dissolved gases in transformer oil of nuclear power plants at home and abroad.
[0052] S2. Construct the fault feature set of the transformer, and through normalizing processing and non-linear dimensionality reduction processing on the fault feature set, extract and obtain the input feature vector, and use the input feature vector as the sample data.
[0053] In step S2, according to the characteristic gases corresponding to different fault types in the transformer oil of nuclear power plants, and according to the content and ratio of dissolved gases corresponding to different fault types in the transformer oil of nuclear power plants, construct the fault feature set of the transformer.
[0054] The insulation system of large oil-immersed transformers in nuclear power plants consists of insulating oil and insulating paper (board). Under the action of factors such as temperature and electric field, the C-H bonds, C-C bonds, and C-O bonds in the insulating material will break, and a small amount of gases such as hydrogen and low-molecular hydrocarbons will be formed through a series of chemical reactions. When there are latent faults inside the transformer, the above-mentioned cracking effect will be strengthened, and the generation rate of these gases will increase. The gas components corresponding to different fault types are shown in Table 1.
[0055] Table 1 Gas components corresponding to different fault types
[0056]
[0057] As can be seen from Table 1, the components and contents of the dissolved gases in oil can reflect the nature and type of transformer faults to a certain extent. However, the fault information carried is limited, and it is impossible to identify the type of transformer fault only based on the content of the above gases.
[0058] The ratio of gas content is related to the type of transformer fault to a certain extent. Therefore, the dissolved gas content and ratio in transformer oil can reflect the type of transformer fault from different aspects. To fully extract the characteristic information of the transformer, the input feature vector of the transformer shown in Table 2 is selected in this embodiment to improve the effect of fault recognition.
[0059] Table 2 Input feature vector of transformer
[0060] Table 2 Input feature vector of transformer
[0061]
[0062] In step S2, the normalization processing method for the fault feature set is as follows:
[0063]
[0064] Among them, X i is the concentration value of the i-th gas; x i is the normalized concentration value of the i-th gas, X i,max is the maximum value of the i-th gas; X i,min is the minimum value of the i-th gas;
[0065] The KPCA algorithm is used for non-linear dimensionality reduction processing of the normalized fault feature set. Through the non-linear mapping kernel function the normalized fault feature set X = (x1, x2, x3…, x n ) is mapped from the low-dimensional space to the high-dimensional space G, and the principal component analysis method is used to extract the principal components of the data in the high-dimensional space G. The newly generated comprehensive variable is used as the input feature vector.
[0066] Since the 16-dimensional DGA fault feature set has a large number of data dimensions, there is information redundancy, resulting in a high complexity of the diagnostic model, which affects the classification speed and accuracy of the model. The KPCA algorithm is used for non-linear dimensionality reduction processing of the DGA fault feature set, thereby improving the convergence speed and recognition accuracy of the transformer fault diagnosis model. At the same time, due to the characteristics of wide fluctuation range and large data difference in the concentration of each characteristic gas, in order to improve the diagnostic accuracy of the model and reduce the influence of singular data on the model, the characteristic gas is normalized before feature extraction. As Figure 2 shown, it is the Pareto chart of kernel principal component analysis of the KPCA algorithm.
[0067] S3. Divide the sample data into a pre-training set, a tuning set, and a test set.
[0068] S4. Establish a DBN network and initialize the parameters of the DBN network. The schematic diagram of the DBN network is as Figure 3 shown.
[0069] S5. Improve the DBO algorithm, and the improved DBO algorithm is used for hyperparameter optimization of the DBN network to construct a transformer deep learning model.
[0070] In step S5, the method for constructing the transformer deep learning model includes the following steps:
[0071] S5.1 Initialize the population position of the DBN network through Logistic chaotic mapping, and set the search range of the population size, the number of iterations, the number of times of DBN backpropagation training, the number of neurons in each hidden layer, and the learning rate.
[0072] In step S5.1, the method for initializing the population position of the DBN network through Logistic chaotic mapping is:
[0073] X i+1 = μX i (1 - X i ) Formula (2)
[0074] where X i is the mapping function value at the i-th iteration, μ is the mapping parameter, and when μ ∈ [3.569, 4], the Logistic mapping distribution trajectory shows completely chaotic characteristics.
[0075] The DBO algorithm initializes the population position by generating random numbers, which easily leads to uneven distribution of the positions of dung beetle individuals, resulting in insufficient population diversity and a decline in the global search ability of the DBO. In this embodiment, Logistic chaotic mapping is used to improve the initial position of the dung beetle population, so that the positions of dung beetle individuals are more evenly distributed in the search space.
[0076] S5.2 Construct a fitness function using the fault recognition error rate, and update the individual positions of the rolling dung beetle, the breeding dung beetle, the small dung beetle, and the thief dung beetle in turn through the improved DBO algorithm, and calculate the fitness value after the position update of the dung beetle.
[0077] In step S5.2, the improved sine method strategy is used to update the individual position of the rolling dung beetle:
[0078]
[0079] where δ = rand(1); ST ∈ (0.5, 1]; x i(t) is the i-th position component of individual X in the t-th iteration; p i (t) is the i-th component of the best individual position variable in the t-th iteration; r2 is a random number in the interval [0, 2π]; r3 is a random number in the interval [-2, 2].
[0080] In step 5.2, the self-adaptive variable inertia weight ω is set by the self-adaptive variable inertia weight strategy t It linearly decreases as the number of iterations increases to:
[0081]
[0082] where ω t is the self-adaptive variable inertia weight;
[0083] A non-linear decreasing mode is adopted to set the value of r1, and the cosine function between 0 and π is used to determine the change of the value of r1:
[0084] where w max is w t 's maximum value; w min is w t 's minimum value; t represents the current iteration number; T max represents the maximum number of iterations.
[0085] In this embodiment, to balance the global exploration and local development capabilities of the DBO algorithm, when the rolling dung beetle encounters an obstacle and needs to reposition a new route, an improved sine method strategy is introduced to replace the tangent dancing strategy in the DBO algorithm, improving the defect that the position update in the DBO algorithm is too random.
[0086] S5.3 Sort the fitness values, and use the dung beetle individual position corresponding to the minimum fitness value as the optimal parameter of the DBN network to construct a transformer deep learning model.
[0087] In step S5.3, the sorting of the fitness values determines the optimal position of the dung beetle individual through Levy flight:
[0088]
[0089] where is the point-to-point multiplication; α is the step size control quantity, and Levy(λ) is the search path where the step size follows the Levy distribution;
[0090] When performing Levy flight, a greedy rule is added to determine whether to update the target position; and the current optimal fitness value is updated through iteration to obtain the dung beetle individual position corresponding to the minimum fitness value.
[0091] To prevent the DBO algorithm from falling into the local optimum problem due to the shrinking search range in the later stage of iteration, the Levy flight strategy is introduced to fully increase the diversity of the dung beetle population and expand the search range, enabling the DBO algorithm to jump out of the local optimum problem and search for the global optimum position. Although Levy flight helps to jump out of the local optimum problem, it cannot guarantee that the fitness of the new position generated after Levy flight is necessarily better than that of the original position. In this embodiment, a greedy rule can be added after Levy flight to determine whether to update the target position.
[0092] The improved DBO algorithm of this embodiment is as Figure 4 shown.
[0093] S6. Obtain the training set and the tuning set in step S3, and train the transformer deep learning model through the training set and the tuning set to obtain the trained transformer deep learning model.
[0094] In step S6, the search ranges of the swarm size, the number of iterations, the number of times of DBN backpropagation training, the number of neurons in each hidden layer, and the learning rate are assigned to the transformer deep learning model, and the transformer deep learning model is trained through the training set and the tuning set.
[0095] S7. Diagnose the test set in step S3 through the trained transformer deep learning model, and evaluate the trained transformer deep learning model. Adjust the trained transformer deep learning model according to the evaluation results to obtain the fault diagnosis model. As Figure 5 shown, it is the confusion matrix diagram of the fault diagnosis results of the fault diagnosis model of this embodiment. In this embodiment, the actual transformer fault data of nuclear power plants is used to evaluate the model performance, and its accuracy rate reaches 97%.
Claims
1. A method for constructing a transformer fault diagnosis model based on a deep learning network, characterized in that, It includes the following steps: S1. Obtain the data of dissolved gases in transformer oil and form the original data of DGA. S2. Construct the fault feature set of the transformer from the original data. Through normalizing and non-linearly reducing the dimension of the fault feature set, an input feature vector is extracted and used as sample data. S3. Divide the sample data into a pre-training set, a tuning set, and a test set. S4. Establish a DBN network and initialize the parameters of the DBN network. S5. Improve the DBO algorithm, and the improved DBO algorithm is used for hyperparameter optimization of the DBN network to construct a transformer deep learning model. S6. Obtain the training set and the tuning set in step S3, and train the transformer deep learning model with the training set and the tuning set to obtain a trained transformer deep learning model. S7. Use the trained transformer deep learning model to diagnose the test set in step S3, and evaluate the trained transformer deep learning model. Adjust the trained transformer deep learning model according to the evaluation results to obtain a fault diagnosis model.
2. The method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 1, wherein: In step S2, construct the fault feature set of the transformer according to the characteristic gases corresponding to different fault types in the nuclear power plant transformer oil, and according to the dissolved gas content and ratio of different fault types in the nuclear power plant transformer oil.
3. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 1, characterized in that: In step S2, the method for normalizing the fault feature set is: Among them, X i is the concentration value of the i-th gas; x i is the normalized concentration value of the i-th gas, and X i,max is the maximum value of the i-th gas; X i,min is the minimum value of the i-th gas; The KPCA algorithm is used for non - linear dimensionality reduction of the normalized fault feature set. Through the non - linear mapping kernel function the normalized fault feature set X=(x1, x2, x3…, x n ) is mapped from a low - dimensional space to a high - dimensional space G, and the principal components of the data are extracted using the principal component analysis method in the high - dimensional space G. The newly generated comprehensive variables are used as input feature vectors.
4. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 1, characterized in that: In step S5, the method for constructing the transformer deep learning model includes the following steps: S5.1 Initialize the population position of the DBN network through Logistic chaotic mapping, and set the search ranges of the population size, the number of iterations, the number of times of DBN backpropagation training, the number of neurons in each hidden layer, and the learning rate. S5.2 Construct a fitness function using the fault recognition error rate. Update the individual positions of the rolling dung beetle, the breeding dung beetle, the small dung beetle, and the thief dung beetle in turn through the improved DBO algorithm, and calculate the fitness value after the position update of the dung beetle. S5.3 Sort the fitness values, and use the individual position of the dung beetle corresponding to the minimum fitness value as the optimal parameters of the DBN network to construct a transformer deep learning model.
5. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 4, characterized in that: In step S5.1, the method for initializing the population position of the DBN network through Logistic chaotic mapping is: X i+1 = μX i (1 - X i ) Formula (2) Among them, X i is the mapping function value at the i-th iteration, μ is the mapping parameter, and when μ ∈ [3.569, 4], the Logistic mapping distribution trajectory exhibits completely chaotic characteristics.
6. The method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 4, characterized in that: In step S5.2, the individual position of the rolling dung beetle is updated using an improved sine method strategy: where, δ = rand(1); ST ∈ (0.5, 1]; x i (t) is the i-th position component of individual X in the t-th iteration; p i (t) is the i-th component of the best individual position variable in the t-th iteration; r2 is a random number on the interval [0, 2π]; r3 is a random number on the interval [-2, 2].
7. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 6, characterized in that: In step 5.2, the value set by the self-adaptive variable inertia weight strategy, the self-adaptive variable inertia weight ω t Linearly decreases as the number of iterations increases to: where ω t is a self-adaptive variable inertia weight; Set the value of r1 in a non-linearly decreasing pattern, and use the cosine function between 0 and π to determine the change of the value of r1. Among them, w max is the maximum value of w t ; w min is the minimum value of w t ; t represents the current iteration number; T max represents the maximum number of iterations.
8. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 4, characterized in that: In step S5.3, the sorting of the fitness values determines the optimal position of the dung beetle individual through Levy flight: Among them, is a point-to-point multiplication; α is a step size control quantity, and Levy(λ) is a search path where the step size follows a Levy distribution; Add a greedy rule during Levy flight to determine whether to update the target position; and iteratively update the current optimal fitness value to obtain the individual position of the dung beetle corresponding to the minimum fitness value.
9. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 8, characterized in that: In step S5.3, it is judged whether the maximum number of iterations is reached. If the maximum number of iterations is not reached, return to continue iterative update. If the maximum number of iterations is reached, the transformer deep learning model is obtained.
10. A method for constructing a transformer fault diagnosis model based on a deep learning network according to claim 4, characterized in that: In step S6, the optimization ranges of the group size, the number of iterations, the number of times of DBN back-training, the number of neurons in each hidden layer, and the learning rate are assigned to the transformer deep learning model, and the transformer deep learning model is trained by the training set and the tuning set.