Converter transformer vibration data sample generation method

Abstract

The invention discloses a converter transformer vibration data sample generation method, which comprises the steps of constructing a correlation between winding vibration and current based on a winding electromagnetic vibration theory, constructing a correlation between iron core vibration and voltage based on an iron core magnetostriction effect, and obtaining the correlation between winding vibration and box vibration and the correlation between iron core vibration and box vibration; according to the association relationship, constructing a converter transformer vibration data generation model based on a dimension attention mechanism-long and short term time sequence network, and performing training; and inputting voltage and current of a target working condition and historical vibration data under the corresponding working condition by using the trained model, and generating a vibration data sample similar to the vibration of the target working condition.

Description

Technical Field

[0001] The present invention relates to a method for generating and enhancing vibration data samples of a converter transformer based on a dimensional attention mechanism (DAM) and a long short time series network (LSTNet), and belongs to the technical field of power systems. Background Art

[0002] The vibration detection technology of converter transformer can sensitively reflect the internal mechanical structure fault. By observing the changes of multiple characteristic parameters of the vibration signal, the internal mechanical structure state can be judged, which is particularly sensitive to early latent mechanical degradation faults. Therefore, the box vibration signal can be used as an important way to identify the operating state of the internal structure of the transformer, and has gradually become a key means of diagnosing early latent mechanical faults of converter transformers.

[0003] The vibration of the converter transformer is mainly manifested as the vibration of the box body caused by the core vibration and the winding vibration, which is generally a coupled vibration signal transmitted through multiple paths of transformer oil and solid connectors. The core vibration mainly comes from the magnetostrictive effect of the silicon steel sheet in the core, and the winding vibration is mainly caused by the electromagnetic force generated by the interaction between the winding current and the leakage magnetic field. Traditionally, it is believed that the frequency of the winding vibration and the core vibration is twice the power supply frequency, the amplitude of the core vibration fundamental frequency is positively correlated with the square of the voltage, and the amplitude of the winding vibration fundamental frequency is positively correlated with the square of the current. According to the load power adjustment requirements of the DC transmission system, the converter transformer will switch between different operating conditions. The changes in voltage and current cause the core vibration and winding vibration characteristics to change, affecting the box vibration distribution, which indicates that there is a mapping relationship between voltage, current and box vibration.

[0004] At present, the research on the vibration change law of converter transformers under the influence of voltage and current mainly starts from the electromagnetic field theory, and is calculated by establishing an electromagnetic-vibration theoretical numerical model and finite element simulation. Although the numerical model can effectively reflect the correlation between voltage, current and vibration in the time-frequency domain, the vibration will be affected by the pads, struts, insulating cardboard and other structures during the transmission to the box, so the numerical model cannot reflect the nonlinearity generated by the vibration during the transmission process. In addition, the internal and external structures of the converter transformer are complex, and it is difficult to reconstruct its real structure in the simulation, which makes the simulation results have a large error in amplitude compared with the measured value, and the data cannot provide rich multi-operating condition samples for state identification, which makes it difficult to improve the identification accuracy.

[0005] Although the converter transformers in operation have been operating under full-load conditions for a long time, sometimes, according to the regular maintenance and power adjustment requirements of the converter station equipment itself, the converter transformers will also operate under complex conditions such as no-load charging and closing, on-load tap changer operation, and converter unlocking and power transmission. Long-term operation under full voltage, full current and full load conditions, under which vibration data collection is relatively easy and a relatively rich data recognition sample can be constructed; however, vibration data under complex conditions such as no-load charging and closing, converter unlocking and power transmission are difficult to obtain, resulting in a small number of vibration data samples for the corresponding complex conditions. The imbalance of sample data directly affects the accuracy of vibration multi-feature recognition method for identifying the operating status of the converter transformer in operation, and also indirectly affects the promotion and application of this recognition method in vibration measurement. It is difficult to reconstruct the complex internal and external mechanical structure of the converter transformer in the current electromagnetic vibration theoretical numerical model and finite element simulation calculation. The conventional deep learning generation model does not deeply explore the spatiotemporal coupling relationship of the box vibration. The influence of the diversity of working conditions is not considered in the vibration data sample generation model, making it difficult to use the vibration sequence generated by voltage and current calculation to construct new vibration feature data samples. Summary of the invention

[0006] In view of this, from the data-driven perspective, the present invention proposes a method for generating vibration data samples of converter transformers based on DAM-LSTNet. By combining the long and short term time series network (LSTNet) with the dimensional attention mechanism (DAM), more abundant vibration data samples are provided for the operating state recognition method, thereby improving the accuracy of converter transformer fault recognition under the condition of unbalanced sample data.

[0007] A method for generating vibration data samples of a converter transformer comprises: constructing a correlation between winding vibration and current based on winding electromagnetic vibration theory, constructing a correlation between core vibration and voltage based on the magnetostrictive effect of the core, and obtaining the correlation between winding vibration, core vibration and box vibration; according to the correlation, constructing a converter transformer vibration data generation model based on a dimensional attention mechanism-long-short term time series network and training the model; using the trained model, inputting the voltage and current of a target working condition and the historical vibration data under the corresponding working condition, and generating vibration data samples similar to the vibration of the target working condition.

[0008] Based on the rich voltage, current and vibration data under multiple operating conditions, the present invention adopts the Long and Shortterm Time Series Network (LSTNet) to extract the local correlation between voltage, current and vibration, and uses the loop-jump module to capture the complex long-term dependency between the two. At the same time, in order to capture the association law of voltage and current with vibration data of different operating conditions, the Dimensional Attention Mechanism (DAM) is introduced to optimize multiple input variables, and a DAM-LSTNet vibration data generation model is proposed to provide richer vibration data samples for the operating state recognition method, and improve the accuracy of converter transformer fault recognition under the condition of unbalanced sample data.

[0009] The DAM-LSTNet model proposed in this invention can generate a vibration sequence similar to the historical data of the target working condition by inputting the voltage, current data and historical vibration data of the target working condition, increase the recognition data samples of the characteristic spectrum library of the same working condition, and enhance the generalization ability of the recognition model. In addition, it is of great significance to improve the accuracy of the operating state recognition method, ensure the safe and stable operation of the converter transformer, and improve the reliability of the DC transmission system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the vibration conduction path of the converter transformer according to an embodiment of the present invention.

[0011] Figure 2 It is a diagram of the DAM-LSTNet vibration data generation model architecture of an embodiment of the present invention.

[0012] Figure 3 It is a diagram of the processing principle of the dimensional attention mechanism of an embodiment of the present invention.

[0013] Figure 4 3 is a comparison diagram of the generated vibration and the measured vibration under different input conditions of the embodiment of the present invention.

[0014] Figure 5 It is a comparative distribution diagram of the vibration data generation results of different models. DETAILED DESCRIPTION

[0015] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods and examples. It should be understood that the purpose of providing the examples is only to illustrate and not to impose any limitation.

[0016] The specific implementation method of the present invention is mainly based on the data-driven perspective. Based on the rich voltage, current and vibration data under multiple operating conditions, a long and short term time series network (LSTNet) is used to extract the local correlation between voltage, current and vibration, and a loop-jump module is used to capture the complex long-term dependency between the two. At the same time, in order to capture the association law of voltage and current with vibration data under different operating conditions, a dimensional attention mechanism (DAM) is introduced to optimize multiple input variables, and a vibration data generation model based on DAM-LSTNet is proposed. Finally, different experimental control groups are designed to analyze the vibration sequence generation effect. Based on this, an embodiment of the present invention proposes a method for generating vibration data samples for converter transformers based on DAM-LSTNet, and the method includes the following steps 1 to 7:

[0017] Step 1: Transmission and numerical calculation of vibration between commutation transformer winding and core:

[0018] The vibration of the converter transformer is mainly manifested as the vibration of the box body caused by the core vibration and the winding vibration. The core vibration mainly comes from the magnetostrictive effect of the silicon steel sheet in the core. Based on the law of electromagnetic induction, the numerical model of the core vibration and voltage can be derived from the macro level for the analysis of the core vibration characteristics. The winding vibration is mainly caused by the electromagnetic force generated by the interaction between the winding current and the leakage magnetic field. Therefore, this step constructs the correlation between the winding vibration and the current based on the winding electromagnetic vibration theory, and the correlation between the core vibration and the voltage based on the core magnetostrictive effect, and obtains the correlation between the winding vibration and the core vibration and the box body vibration, which is used as a theoretical basis for the construction of the vibration data generation model.

[0019] The grid-side voltage of the converter transformer is generally a sinusoidal power frequency voltage u(t)=Usinωt, which contains almost no harmonic components. Therefore, the core vibration can be expressed as:

[0020]

[0021] Among them, a c represents the core vibration, L is the length of the steel silicon sheet in the core, ΔL represents the change in the length of the steel silicon sheet under the magnetostrictive effect, λ s is the saturation magnetostriction coefficient, U is the voltage, N 1 is the number of turns of the primary winding, A c is the effective cross-sectional area of ​​the core, B s is the saturation flux density;

[0022] The relationship between winding vibration and current is: the winding vibration amplitude is positively correlated with the square of the current, and the winding vibration frequency is twice the current frequency. Therefore, the winding vibration can be expressed as:

[0023] a w ∝I 2 cos2ωt (2)

[0024] Among them, a w Represents winding vibration, I is current, and ω is power frequency.

[0025] When the converter transformer is online, the vibration of the winding and the core is transmitted to the surface of the box through multiple paths. The specific transmission paths are as follows: Figure 1 As shown. The winding is tightly wound on the iron core, and the transmission process of the winding vibration and the iron core vibration will affect each other. Part of the core vibration is directly transmitted to the box through the transformer oil and the solid connectors connected to the upper and lower iron yokes, and the other part is transmitted to the winding, causing the winding vibration. When the winding vibrates, it will also cause the core vibration and then be transmitted through the solid connector, and the remaining part of the vibration is directly transmitted to the box through the transformer oil. Therefore, it can be seen that the surface vibration of the converter transformer box is a coupled vibration after the multi-source vibration is transmitted through multiple paths.

[0026] Step 2: Establishment of converter transformer vibration data generation model:

[0027] The winding vibration generated by the converter transformer under the action of current and the core vibration generated by the voltage are transmitted to the transformer box, which together cause the box vibration. Based on this theoretical basis, the embodiment of the present invention constructs the following Figure 2 The vibration data generation model DAM-LSTNet shown in the figure mainly uses the dimensional attention mechanism (DAM) to optimize the long-term and short-term time series network (LSTNet) structure, and calculates the weight matrix of each dimensional input variable (the input variable is voltage, current and historical vibration data under the corresponding situation) and output vibration (referring to the box vibration, the winding and core vibration is transmitted to the box to form the box vibration) through the fully connected layer, and updates the weight through the loss function. The weight matrix and the feature matrix (composed of voltage, current and historical vibration data) are multiplied element by element to achieve parameter weight adjustment under the optimal strategy, thereby improving the vibration sequence generation effect. Finally, according to the trained vibration data generation model DAM-LSTNet, the voltage, current and historical vibration data under the corresponding working conditions associated with the vibration are input to generate a vibration sequence similar to the target working condition vibration.

[0028] Step 3: Calculate DAM-LSTNet model parameters:

[0029] The purpose of this step is to re-extract the features of the input variables and then calculate the model parameters, which mainly includes four aspects: feature information extraction, dimensional attention mechanism processing, LSTM network processing, autoregressive linear fitting and result output.

[0030] 3.1 Feature Information Extraction

[0031] Please refer to Figure 2 The DAM-LSTNet model of the embodiment of the present invention has a multivariate input (voltage, current, vibration data). The feature information extraction module first extracts the historical vibration data of the same time interval through the loop-jump module to form a feature matrix, and then performs a convolution operation on the feature matrix to obtain a multidimensional information matrix.

[0032] The loop-jump module is a part of the known LSTNet network structure. In a preferred embodiment of the present invention, the loop-jump module is used to synchronously extract two-way features to obtain the feature matrix W 1 , W 2 ; The feature matrix W 1 Contains the periodic correlation of vibration data, the characteristic matrix W 2 is the multidimensional parameter matrix of voltage and current input variables. Next, the characteristic matrix W 1 , W 2 Perform convolution operation to obtain the multidimensional information matrix W 1 ', W 2 ', to ensure the consistency of the dimensions of the previous and next data, set the convolution kernel size to 1, W 1 'With W 2 ' is expressed as follows:

[0033]

[0034] In formula (3), W k is the convolution kernel, D is the dimension of the multidimensional information matrix, b k is the bias term to realize the connection between convolutional layers.

[0035] 3.2. Dimensional attention mechanism processing:

[0036] refer to Figure 3 , the dimensional attention mechanism extracts t k The input vector at time and The fitted weight vector obtained after processing by the fully connected layer constitutes a matrix, and the weight values ​​of each dimension are processed by softmax normalization to form a new weight matrix W weights ; Matrix W 1 'With W 2 'After transposition, multiply each element by the weight matrix to obtain the matrix W for dimensional attention processingDAM The mathematical formula for dimensional attention processing is as follows:

[0037]

[0038] In formula (4), {X 1},{X 2},{X 3},…{X D} represents the multidimensional information matrix W 1 ', W 2 'In each dimension sequence, for the multidimensional information matrix W 1 'For dimensional attention processing, W 1 Substitute the dimensional sequences of ' into formula (4) to obtain the multidimensional information matrix W 2 'For dimensional attention processing, W 2 ' is substituted into formula (4); {Y} is the training line loss rate sequence; F dense It means that after being processed by the fully connected layer, the number of neurons in the first layer of the fully connected layer is the matrix dimension D, ⊙ means the corresponding multiplication of the two matrix elements one by one, and softmax is the normalized exponential function.

[0039] 3.3, LSTM network processing:

[0040] In the embodiment of the present invention, the DAM-LSTNet model uses the LSTM network as a recurrent module to convert the matrix W processed by the dimensional attention mechanism into DAM After inputting into LSTM, a vibration sequence is generated. The loop-jump module in LSTM is a loop network that introduces continuous jumps in a cycle, which can extract the long-term correlation of the historical vibration sequence by setting continuous jumps in cycle T. The DAM-LSTNet model of the embodiment of the present invention adopts two LSTMs, namely LSTM_1 and LSTM_2. The matrices obtained by the two-way dimensional attention processing enter LSTM_1 and LSTM_2 respectively for adaptive feature extraction.

[0041] 3.4. Autoregressive linear fitting and result output:

[0042] In order to reduce the impact of non-periodic changes in signals on the generation of network results, autoregressive analysis is used to add linear components to the LSTM output vibration sequence, and the final generated vibration data samples are summed and output through the fully connected layer.

[0043] In the LSTNet structure of the embodiment of the present invention, a linear component is added to the output result, and a classic linear AR model is used as a linear component to form an autoregressive linear module. The mathematical expression of the autoregressive linear module is:

[0044]

[0045] In formula (5), is the output of the autoregressive linear model, and b ar is the model coefficient, q ar is the input matrix window size.

[0046] The output of the autoregressive linear module The final generated result is obtained by summing the outputs of LSTM_1 and LSTM_2 through the fully connected layer:

[0047]

[0048] In formula (6), Represents the sum of the outputs of LSTM_1 and LSTM_2, which is the vibration sequence before adding the linear component.

[0049] It can be seen that in the model calculation process of step 3, not only multi-dimensional time series processing is performed, but also the short-term and long-term correlations between voltage, current and vibration sequences are extracted by integrating the dimensional attention mechanism and the computational advantages of long-term and short-term time series networks, and the model parameters are adaptively adjusted based on the causal transmission law of voltage & current-vibration, providing a theoretical reference for multi-condition vibration data generation and model parameter selection.

[0050] The DAM-LSTNet model of the embodiment of the present invention fits the mapping law of voltage, current and vibration. Through the calculation in step 3, the neural network adaptive feature extraction fixes the model parameters. Subsequently, only the voltage and current of the target working condition and the historical vibration data under the corresponding working condition need to be input to generate vibration data samples similar to the vibration of the target working condition. The similarity can be evaluated by the error evaluation index of the generated sequence.

[0051] Step 4: Collect and construct vibration data of multiple operating conditions during operation

[0052] The data comes from the voltage, current and vibration measurement data of the LY-A phase converter transformer of the ±800kV converter station under 6 different operating conditions. Condition 1 is the no-load charging closing condition, conditions 2 to 4 are the no-load operating stages 10min, 20min and 24min after charging closing, condition 5 is the converter unlocking condition, and condition 6 is the full-load operating condition. The sampling frequency is 10kHz, the sampling time of each condition is 6s, and there are 60,000 sample points in total. Among them, 50,000 samples are used as the training set of the model, and the remaining 10,000 samples are used as the test set. The division of the training set and the test set is shown in Table 1 and Table 2. Conditions 1 to 6 cover most of the operating conditions of the converter transformer in operation, providing a good data basis for the generation and effect analysis of the vibration data of the full cycle condition of the converter transformer.

[0053] Table 1. Training and test sets of voltage and current input variable data of the model

[0054]

[0055]

[0056] Table 2 Model vibration output value training set and test

[0057]

[0058] Step 5: Generate effect evaluation indicators:

[0059] In order to evaluate the effectiveness of the vibration data sample generation model DAM-LSTNet according to the embodiment of the present invention, the R-Square coefficient (R 2 ), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used as evaluation indicators for the generation effect of vibration data samples of the entire model.

[0060]

[0061]

[0062] In formulas (7) to (10), y i is the actual vibration value collected, is the generated vibration value, y i is the actual vibration average value, and n is the length of the vibration sequence. RMSE, MAE, and MAPE are negative indicators. The smaller their values ​​are, the smaller the error between the generated value and the true value is. 2 It is an indicator to evaluate the closeness between the generated curve and the true curve, and its value ranges from 0 to 1; R 2 The closer the coefficient is to 1, the closer the generated curve is to the true curve, and the better the model vibration sample enhancement effect is.

[0063] Step 6: Model generation test design and effect comparison under multiple working conditions:

[0064] In the converter transformer under different operating conditions, the measured box vibration has winding vibration and core vibration in different coupling superposition states, so the correlation between voltage, current and vibration can be extracted based on multiple operating conditions. When voltage and current are used as input variables in different combinations, the accuracy of DAM-LSTNet model generation will change. Therefore, the converter transformer vibration data generation experiment is carried out under the input conditions of voltage single variable, current single variable, and voltage and current multivariable fusion. Taking the measurement point 5 under operating condition 1 as an example, the error analysis of vibration data generation effect is carried out through evaluation indicators.

[0065] Under the voltage single variable input condition, the input variables in the training set and test set samples of the model are only the grid side voltage, valve side voltage and neutral point voltage, and the vibration data of measuring point 5 under working condition 1 is the output. Under the current single variable input condition, the input variables in the training set and test set samples of the model are only the grid side current, valve side current and neutral point current, and the rest of the model settings are consistent with the voltage single variable input. Under the current and voltage multivariate input conditions, the training set and test set samples of the model contain both voltage data and current data, and the vibration data of measuring point 5 under working condition 1 is the output. Figure 4 The comparison between the measured vibration and the generated vibration under different input variable conditions in a time length of 0.02s (200 sample points) is shown.

[0066] The vibration sequence generation effect of different working conditions is evaluated based on the vibration data generation model. Table 3 shows the average values ​​of R-Square coefficient, root mean square error, mean absolute error, and mean absolute percentage error under different input conditions for 5 repeated experiments.

[0067] Table 3 Evaluation indexes of vibration data generation effect under different input conditions

[0068]

[0069] Taking the average value of the evaluation index as an example, R 2 The coefficients are increased by 15.91% and 9.57% compared with voltage input and current input, respectively. The average errors of RMSE, MAE and MAPE are reduced by 77.59%, 80.19% and 77.41% compared with voltage input, and by 74.07%, 78.84% and 78.78% compared with current input. Therefore, from the comparison of the single-input and multi-input vibration data generation effects, it can be seen that the model with voltage and current as multi-variable input has a good vibration data generation effect.

[0070] Step 7: Comparison between the DAM-LSTNet vibration data sample generation model and the traditional generation model:

[0071] In the vibration data generation effect comparison experiment of the DAM-LSTNet model, LSTM, GRU, CNN, and BP models were used as the control group. The training set and test set data were all from the multi-condition vibration of measuring point 5. The model was repeatedly trained 5 times and the corresponding multi-condition generated vibration sequence was output. The vibration data generation effects of different models are shown in Figure 2. Figure 5 shown.

[0072] Figure 5 The box plot is combined with the scatter plot, where the scatter plot represents the generation effect of the vibration model of 5 repeated experiments, and the green represents the optimal generation result. In the box plot, the solid square is the average of the 5 results, the upper and lower sides of the dark blue box represent the Q1 (first quartile) and Q3 (third quartile) of the decision results, the red line segment represents the median, and the brown line segments on the upper and lower edges are calculated by Q1, Q3 and the interquartile range.

[0073] From the results, the DAM-LSTNet model based on voltage and current multi-input proposed in the embodiment of the present invention has a better vibration data generation effect. In the comparative experiment, the evaluation index R 2 , RMSE, MAE, and MAPE are all the best results.

[0074] In summary, the present invention takes into account the diversity of input variables and operating conditions. Since the vibration data of normal operating conditions are relatively abundant, it is relatively easy to collect on-site. However, the vibration data under complex operating conditions such as no-load closing and converter unlocking are difficult to obtain, resulting in less sample data of the vibration characteristic map of complex operating conditions. When identifying the operating state, due to insufficient sample data of complex operating conditions, low recognition accuracy or even misidentification may occur. The present invention extracts the dependency between voltage, current and vibration variables based on the feature extraction advantage of the neural network model, and effectively reflects the causal law between related voltage & current-vibration variables. The LSTNet model structure is optimized by using the dimensional attention module DAM, and the weight matrix of each dimensional input variable and output vibration is calculated through the fully connected layer. The weight is updated through the loss function, and the parameter weight adjustment under the optimal strategy is realized by multiplying the weight matrix and the feature matrix element by element, thereby improving the vibration sequence generation effect. Finally, according to the trained model, the voltage, current and other data associated with the vibration are input to generate a vibration sequence similar to the target operating condition vibration, so as to enhance the operating condition vibration sample data.

[0075] The present invention compares and analyzes the error effects of vibration data under different operating conditions and measuring point locations, and uses ±800kV converter transformer voltage & current-vibration data to perform multiple repeated algorithm experiments. By analyzing and comparing the error effects and applicability of different generation methods, the effectiveness and correctness of the method proposed in this paper are proved. The effectiveness of the vibration data generation model is verified by designing different control group experiments and vibration data generation evaluation indicators. Based on the DAM-LSTNet model, by inputting the voltage, current data and historical vibration data of the target working condition, a vibration sequence similar to the historical data of the target working condition can be generated, and the recognition data samples of the characteristic spectrum library of the same working condition can be increased, thereby enhancing the generalization ability of the recognition model. In addition, it is of great significance to improve the accuracy of the operating state recognition method, ensure the safe and stable operation of the converter transformer, and improve the reliability of the DC transmission system.

[0076] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for generating converter transformer vibration data samples, It is characterized in that include: Based on the theory of winding electromagnetic vibration, the correlation between winding vibration and current is constructed, and based on the magnetostrictive effect of the iron core, the correlation between iron core vibration and voltage is constructed to obtain the correlation between winding vibration, iron core vibration and box vibration; According to the association relationship, a converter transformer vibration data generation model based on a dimensional attention mechanism-long-short term time series network is constructed and trained; By using the trained model, the voltage and current of the target working condition and the historical vibration data under the corresponding working condition are input to generate vibration data samples similar to the vibration of the target working condition.

2. The method for generating converter transformer vibration data samples according to claim 1, It is characterized in that The correlation between the winding vibration and the current is that the winding vibration amplitude is positively correlated with the square of the current, and the winding vibration frequency is twice the current frequency; the correlation is that the winding vibration generated by the converter transformer under the action of the current and the core vibration generated by the voltage are transmitted to the transformer box, jointly causing the box vibration.

3. The method for generating converter transformer vibration data samples according to claim 2, It is characterized in that The correlation between the winding vibration and the current is expressed as: a w ∝I 2 cos2ωt Among them, a w represents the vibration of the winding, I is the current, and ω is the power frequency; The correlation between the core vibration and the voltage is expressed as: Among them, a c represents the core vibration, L is the length of the steel silicon sheet in the core, ΔL represents the change in the length of the steel silicon sheet under the magnetostrictive effect, λ s is the saturation magnetostriction coefficient, U is the voltage, N 1 is the number of turns of the primary winding, A c is the effective cross-sectional area of ​​the core, B s is the saturation flux density.

4. The method for generating converter transformer vibration data samples according to claim 1, It is characterized in that The model includes: A feature information extraction module, used to extract features from multi-variable inputs to obtain a corresponding multi-dimensional information matrix; wherein the multi-variable inputs include voltage input, current input, and historical vibration sequence input of the converter transformer; A dimensional attention mechanism module, used for performing dimensional attention processing on the multidimensional information matrix; An LSTM module, for generating a vibration sequence using the matrix processed by the dimensional attention mechanism module as input; The autoregressive linear module is used to add linear components to the generated vibration sequence to generate the final vibration data samples.

5. The method for generating converter transformer vibration data samples according to claim 4, It is characterized in that The feature information extraction module extracts features from the multivariate input to obtain a corresponding multidimensional information matrix, including: Extract historical vibration data of the same time interval through a loop-jump module to obtain a first characteristic matrix and a second characteristic matrix; wherein the first characteristic matrix contains periodic correlation of vibration data, and the second characteristic matrix is ​​a multidimensional parameter matrix of voltage and current input variables; Convolution operations are performed on the first characteristic matrix and the second characteristic matrix respectively to obtain a first multi-dimensional information matrix and a second multi-dimensional information matrix.

6. The method for generating converter transformer vibration data samples according to claim 5, It is characterized in that The step of the dimensional attention mechanism module performing dimensional attention processing on the multidimensional information matrix includes: Multiply the transpose of the first multidimensional information matrix by the weight matrix element by element to obtain a first dimensional attention processing matrix; Multiply the transpose of the second multidimensional information matrix by the weight matrix element by element to obtain a second dimensional attention processing matrix; The weight matrix represents the contribution to vibration generation, and the influence weight value is assigned through model adaptive training.

7. The method for generating converter transformer vibration data samples according to claim 6, It is characterized in that The LSTM module includes: a first LSTM and a second LSTM, which are used to process the first dimensional attention processing matrix and the second dimensional attention processing matrix respectively.

8. The method for generating converter transformer vibration data samples according to claim 7, It is characterized in that The autoregressive linear module takes the multivariable input as input, performs a function linearization calculation, and sums the obtained output result with the output of the first LSTM and the second LSTM through a fully connected layer to obtain the final vibration data sample.

9. The method for generating converter transformer vibration data samples according to claim 1, It is characterized in that The training process of the converter transformer vibration data generation model includes: The dimensional attention mechanism module is used to optimize the long- and short-term time series network structure. The weight matrix of each dimensional input variable and output vibration is calculated through the fully connected layer. The weight is updated through the loss function. The weight matrix and the feature matrix are multiplied element by element to achieve parameter weight adjustment under the optimal strategy. Among them, the input variables of each dimension refer to voltage, current and historical vibration data under corresponding conditions; the output vibration refers to the box vibration formed by the transmission of winding vibration and core vibration to the box; the characteristic matrix is ​​composed of voltage, current and historical vibration data.

10. The method for generating converter transformer vibration data samples according to claim 1, It is characterized in that The method further includes: evaluating the effect of the model on generating the vibration data samples by using an R-Square coefficient, a root mean square error, a mean absolute error and / or a mean absolute percentage error.

Images