A current transformer fault early warning method

Abstract

The present invention proposes a current transformer fault warning method, which relates to the field of fault prediction. The fault warning method comprises collecting current transformer operating status information, preprocessing the collected operating status information data, constructing a current transformer fault warning model, and training and testing the current transformer fault warning model; the current transformer fault warning model consists of a long-term trend analysis module, a conditional enhancement modeling module, a bidirectional LSTM module integrating an attention mechanism and a convolutional layer, and a self-adjusting fusion module; the long-term trend analysis module models the long-term trend in the time series; the conditional enhancement modeling module models the seasonal component in the time series; the bidirectional LSTM module integrating an attention mechanism and a convolutional layer captures the short-term and long-term dependencies in the time series; and the self-adjusting fusion module automatically adjusts the contribution of each model output to the final prediction result. The method accurately predicts the future fault conditions of the current transformer.

Description

Technical Field

[0001] The invention belongs to the field of fault prediction, and in particular relates to a current transformer fault early warning method. Background Art

[0002] In the operation and maintenance of power systems, current transformers are one of the key devices, and their reliability is directly related to the safety and stability of the entire system. However, since current transformers are in a high-voltage, high-current operating environment for a long time, they are subject to multiple stresses such as electrical, thermal, and mechanical stresses, making them prone to failure. Traditional current transformer fault detection methods often rely on regular maintenance and simple single-variable monitoring, which have problems such as detection lag and inaccurate fault location, and cannot meet the needs of modern power systems for efficient and accurate early warning.

[0003] In order to overcome the shortcomings of traditional methods, multivariate time series data and deep learning technology have been gradually introduced into the field of current transformer fault warning in recent years. Multivariate time series data can fully reflect the operating status of current transformers, including multiple key parameters such as current, voltage, temperature, vibration, insulation resistance, etc. Through the comprehensive analysis of these data, the changing trend of equipment status can be captured more accurately, thereby improving the accuracy of fault detection.

[0004] The present invention collects a large amount of historical data, combines the advantages of multivariate time series and deep learning, and proposes a current transformer fault warning model. The warning model has powerful feature extraction and pattern recognition capabilities and can effectively process complex multivariate time series data. The warning model can automatically learn potential fault modes and features from historical data without relying on expert experience or manually designed features, thereby improving the intelligence level of fault warning. Summary of the invention

[0005] The main purpose of the present invention is to provide a current transformer fault warning method, aiming to propose a current transformer fault warning model, the transformer fault warning model is composed of a long-term trend analysis module, a conditional enhancement modeling module, a bidirectional LSTM module integrating an attention mechanism and a convolutional layer, and a self-adjusting fusion module; wherein, the long-term trend analysis module models the long-term trend in the time series, so that the model ignores the interference of short-term fluctuations and noise on the overall prediction; the conditional enhancement modeling module models the seasonal component in the time series, and accurately captures complex and non-stationary seasonal patterns by combining seasonal feature reference, denoising stage and seasonal signal restoration technology; the bidirectional LSTM module integrating an attention mechanism and a convolutional layer extracts local features and periodic patterns through the convolutional layer, the attention mechanism focuses on key time steps, and the bidirectional LSTM combines forward and reverse information flows to comprehensively capture short-term and long-term dependencies in the time series; the self-adjusting fusion module enables the warning model to automatically adjust the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolutional layer to the final result.

[0006] To achieve the above object, the technical solution of the present invention is: a current transformer fault early warning method, the method comprising:

[0007] S1. Collecting operating status information of the current transformer, wherein the operating status information includes electrical parameters, temperature parameters, mechanical parameters and insulation resistance;

[0008] S2. Preprocessing the collected operation status information data, wherein the preprocessing includes data cleaning and data normalization operations to improve the overall quality of the data set;

[0009] S3. A current transformer fault warning model is proposed. The specific method includes:

[0010] S31. Build a long-term trend analysis module, which models the long-term trend in the time series so that the model ignores the interference of short-term fluctuations and noise on the overall prediction;

[0011] S32. Construct a conditional enhancement modeling module to model the seasonal component in the time series, and accurately capture complex and non-stationary seasonal patterns by combining seasonal feature reference, denoising stage and seasonal signal restoration technology;

[0012] S33. Construct a bidirectional LSTM module that integrates the attention mechanism and the convolutional layer. The convolutional layer extracts local features and periodic patterns, the attention mechanism focuses on key time steps, and the bidirectional LSTM combines forward and reverse information flows to fully capture short-term and long-term dependencies in the time series.

[0013] S34, construct a self-adjusting fusion module, the early warning model automatically adjusts the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolution layer to the final result;

[0014] S4, training the current transformer fault warning model, wherein the training includes setting hyperparameters, and as the number of iterations increases, the accuracy of the fault warning model on the validation set becomes higher and higher;

[0015] S5. Testing the current transformer fault warning model. The test is to use a test set to test the accuracy of the current transformer fault warning model to determine whether the fault warning model can be used in an actual environment.

[0016] Furthermore, in step S31, the long-term trend analysis module first processes the input data into the trend component of the time series by modeling the long-term trend in the time series. The processing formula is as follows:

[0017] ;

[0018] In the formula, is the original sequence, is the trend part of the forecast window, It is a moving average operation;

[0019] Then the long-term trend analysis module uses two modeling paths to model the trend part, and finally uses the dynamic weighted fusion mechanism for fusion; modeling path one uses the long-term trend path modeling to capture the global change characteristics of the data, and the formula operation is:

[0020] ;

[0021] In the formula, is the weight matrix used to extract linear changes, is a bias term used to adjust the center of the model output. To capture the global change characteristics;

[0022] The second modeling path is to use complex trend path modeling to capture complex trends through flexible nonlinear transformations. The formula is:

[0023] ;

[0024] In the formula, is the weight matrix used to extract nonlinear changes, is the bias term, is a nonlinear change function, is a logarithmic transformation, and the calculation formula is , To capture the complex trend changes;

[0025] The feature interaction mechanism is then used to enhance the long-term trend path With complex trend paths The feature interaction between them uses the attention mechanism for interaction, and the operation formula is:

[0026] ;

[0027] In the formula, is the interactive feature, for The transpose operation, is the scaling constant, is the activation function; the feature interaction mechanism dynamically aggregates the features of the two trend modeling paths;

[0028] Then, the dynamic weighted fusion mechanism is used to dynamically adjust the contribution of the two paths according to the characteristics of the input data. First, the statistical information of the historical window is used to generate the weighted coefficient. The formula for generating the weighted coefficient is:

[0029] ;

[0030] In the formula, MLP is a multi-layer perceptron network. is the mean and variance of the historical window, through Ensure that the weight is Between, and satisfy normalization;

[0031] Eventually the dynamic weight and The final trend estimation result is applied to the long-term trend path and interaction characteristics. The fusion formula is:

[0032] ;

[0033] In the formula, and is a dynamic weight, and the trend is simple. Larger, the results are more inclined to model the long-term trend path If the trend is complex Larger, the results are more dependent on the interaction features ; The outputs of long-term trend path modeling and complex trend path modeling are adjusted by weights to obtain the final trend estimation result.

[0034] Furthermore, in step S31, the long-term trend analysis module can adapt to both simple linear trends and complex nonlinear changes by combining long-term trend path modeling with complex trend path modeling; in addition, the dynamic weighting mechanism of the long-term trend analysis module ensures that the model can adaptively adjust the modeling strategy according to the characteristics of the input data; the feature interaction mechanism generates a comprehensive expression of the long-term trend path and the complex trend path, captures the potential correlation between the features, and the dynamic weighted fusion flexibly adjusts the contribution of the features according to the characteristics of the input data.

[0035] Furthermore, in step S32, the conditional enhanced modeling module is composed of a feature combination strategy and a dynamic contribution balance; the feature combination strategy is used to dynamically integrate the statistical characteristics of the reference seasonal data and the state of the current denoising stage to generate conditional inputs to guide the model to gradually restore the harmonic seasonal forecast of the target window; the dynamic contribution balance adjusts the contribution ratio of the conditional input and the restored seasonal signal according to the denoising stage and data characteristics to ensure that the model can balance the rough trend and detail recovery at each stage.

[0036] Furthermore, in step S32, the feature combination strategy first divides the historical window into Divide into multiple segments , ,..., , each piece contains H time points, and then the statistical information of each piece is calculated from the historical window to obtain the mean of each piece and variance , the calculation formulas for the mean and variance are as follows:

[0037] , , where is the number of fragments, is the mean of the historical window, indicating the center of historical seasonal fluctuations. is the variance of the historical window, indicating the magnitude of seasonal fluctuations;

[0038] Then, a multi-layer perceptron is used to aggregate the features of each fragment to generate global historical statistical features, thereby generating conditional input. The aggregation formula is as follows:

[0039] , , where is the mean prediction of the target window, For the variance prediction of the target window, adjust the condition input according to the denoising status and use the current denoising step number and noise scheduling parameters Dynamically adjust the condition input, the formula is as follows:

[0040] , , where is Gaussian noise, used to simulate uncertainty;

[0041] The dynamic contribution balance is then combined and the final output is generated according to the denoising step. And historical features generate weights, the calculation formula is as follows:

[0042] ;

[0043] In the formula, is the weight of the current restored seasonal signal, The weight of the condition input; used to dynamically adjust the condition input With the current reduction season signal Contribution ratio;

[0044] According to the weighted combination condition input and restored seasonal signals, the final output is the dynamically combined harmonic seasonal forecast. The calculation formula is:

[0045] ;

[0046] In the formula, is the intermediate result of denoising in the denoising process, is the output after dynamic combination; the output of each denoising step will be used as the input of the next step, and the complete harmonic seasonal forecast will be gradually restored through multiple iterations.

[0047] Furthermore, in step S33, the convolution operation is performed by sliding a convolution kernel and inputting time series data to extract local features, and the formula is as follows:

[0048] ;

[0049] In the formula, is the eigenvalue at the tth moment in the input sequence, is the output feature after the convolution operation, is the weight of the convolution kernel, is the ReLU activation function, is the bias term of the convolution layer; the size of the convolution kernel is k, then the convolution operation will slide a window of length k in the input sequence, extract the features of the area by weighted summing the data in each window, and perform convolution calculation at each time t;

[0050] The self-attention mechanism then assigns different weights to different parts of the input sequence, thereby focusing on the most meaningful information, where the query matrix in the self-attention mechanism is , key matrix and the value matrix From the same matrix ,Right now , , ; The calculation formula of the attention mechanism at time t is as follows:

[0051] ;

[0052] In the formula, is the scaling factor; the final output of the self-attention mechanism contains , each The input weighted by the self-attention weight , the output of the self-attention mechanism is used as the input of the subsequent bidirectional LSTM;

[0053] Two-Way From positive With reverse It is composed of two independent The difference between the layers is that they process data in different orders. From left to right in the sequence, that is, from time step arrive Processing, gradually updating its hidden state from the past to the present, reverse From right to left in the sequence, that is, from time step arrive Processing, gradually updating its hidden state from the future to the past, in a bidirectional In, two The layers learn the dependencies of time series data in the forward and reverse directions respectively. Each has its own independent parameters and calculation process; Starting from the first step of the sequence and processing the input sequence backward step by step, the calculation process is as follows:

[0054] ;

[0055] In the formula, For a positive The output, The hidden state of the previous moment, positive As one of the inputs at the current moment, is the cell state at the previous moment;

[0056] Reverse Starting from the last time step of the sequence, the input sequence is processed step by step forward. The calculation process is as follows:

[0057] ;

[0058] In the formula, For a layer of reverse The output, The hidden state at the next moment, reverse As one of the inputs at the current moment, is the cell state at the next moment;

[0059] In one-way In the paper, the forget gate and cell state are improved. First, the forget gate is improved by introducing a multi-scale mechanism to dynamically adjust the forgetting rate at different time steps, and use a multi-layer perceptron to dynamically calculate the forgetting rate at different time scales. The formula is as follows:

[0060] ;

[0061] In the formula, To improve the output of the forget gate, is the weight matrix, is the bias term, for The hidden state of the moment, is the forgetting weight calculated by multi-scale features, is an adjustable coefficient that helps the model finely control the forgetting process. is the activation function; secondly, the cell state is improved to use feedback cell state, introducing a feedback mechanism so that the cell state is not only affected by the current input, but also updated in combination with the past cell state. The formula is as follows:

[0062] ;

[0063] In the formula, To provide feedback on the cell state at the previous moment and further enhance the memory ability of the cell state, is the output of the input gate, which determines the current input How much information is written to the cell state, for Feedback cell status at all times, for Feedback cell status at all times;

[0064] The outputs of the forward and reverse LSTM are concatenated at each time step, and the concatenation formula is:

[0065] ;

[0066] In the formula, is the matrix concatenation operation, Two-Way Final output:

[0067] Finally, after the last fully connected layer, the bidirectional fusion attention mechanism and the convolutional layer are calculated. The output value of the module, the fully connected calculation formula is as follows:

[0068] ;

[0069] In the formula, is the weight of the fully connected layer, is the bias term of the fully connected layer, To integrate the attention mechanism with the bidirectional convolutional layer The final output of the module.

[0070] Furthermore, in step S33, the bidirectional LSTM that integrates the attention mechanism and the convolutional layer combines the advantages of the convolutional layer, the attention mechanism and the bidirectional LSTM, and can better process complex time series data; the convolutional layer is used to extract local features to help the model capture short-term dependencies and local patterns, which is particularly important for time series with periodicity or trends; the attention mechanism further enhances the focusing ability of the model, enabling it to automatically identify and focus on the most critical parts of the sequence, especially in long sequence data; the bidirectional LSTM enhances the model's understanding of long- and short-term dependencies in the sequence by simultaneously considering forward and reverse contextual information, thereby comprehensively improving the model's expressive power.

[0071] Furthermore, in step S34, the self-adjusting fusion module automatically adjusts the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolutional layer to the final result. The self-adjusting fusion module performs weight learning during model training and dynamically adjusts the influence of each module under real-time warning conditions. The calculation formula of the self-adjusting fusion module is as follows:

[0072] ;

[0073] In the formula, , , is the weight parameter, obtained during model training, where , This is the final prediction result of the fault warning model.

[0074] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0075] In the present invention, a current transformer fault warning model is constructed, and the transformer fault warning model consists of a long-term trend analysis module, a conditional enhancement modeling module, a bidirectional LSTM module integrating an attention mechanism and a convolutional layer, and a self-adjusting fusion module; wherein, the long-term trend analysis module models the long-term trend in the time series, so that the model ignores the interference of short-term fluctuations and noise on the overall prediction; the conditional enhancement modeling module models the seasonal component in the time series, and accurately captures the complex and non-stationary seasonal pattern by combining the seasonal feature reference, the denoising stage and the seasonal signal restoration technology; the bidirectional LSTM module integrating the attention mechanism and the convolutional layer extracts local features and periodic patterns through the convolutional layer, the attention mechanism focuses on the key time steps, and the bidirectional LSTM combines forward and reverse information flows to comprehensively capture the short-term and long-term dependencies in the time series; the self-adjusting fusion module enables the warning model to automatically adjust the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolutional layer to the final result. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 The present invention is a flow chart of the steps of a current transformer fault early warning method.

[0077] Figure 2 The present invention is a flow chart of the steps of a current transformer fault warning model in a current transformer fault warning method.

[0078] Figure 3 This is the structure diagram of the long-term trend analysis module of the current transformer fault warning model.

[0079] Figure 4 This is the structural diagram of the conditional enhancement modeling module in the current transformer fault warning model.

[0080] Figure 5 This is the structure diagram of the bidirectional LSTM module that integrates the attention mechanism and the convolution layer in the current transformer fault warning model.

[0081] Figure 6 This is a comparison chart of the predicted value and the actual value of the current transformer fault warning model. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0083] See also Figure 1-Figure 6 The present invention provides a technical solution: a current transformer fault warning method, the method steps including collecting current transformer operating status information, preprocessing the collected operating status information data, building a current transformer fault warning model, training the current transformer fault warning model, and testing the current transformer fault warning model.

[0084] Please refer to Figure 1 As shown, a current transformer fault warning method in an embodiment of the present application, the specific steps are as follows:

[0085] S1. Collecting operating status information of a current transformer, wherein the operating status information includes electrical parameters, temperature parameters, mechanical parameters, and insulation resistance.

[0086] Furthermore, in step S1, the electrical parameters include the current value obtained by the current sensor, the voltage value and the power value obtained by the voltage sensor, the temperature parameters include the coil temperature inside the transformer, the casing temperature and the ambient temperature value, the mechanical parameters are the vibration value obtained by the vibration sensor and the noise value obtained by the noise sensor, the insulation resistance parameter is the insulation resistance value obtained by the insulation resistance tester, and the acquisition frequency is set to once every 5 minutes.

[0087] S2. Preprocessing the collected operation status information data, wherein the preprocessing includes data cleaning and data normalization operations to improve the overall quality of the data set.

[0088] Furthermore, in step S2, data cleaning is to process missing values ​​of the collected data, and the method used for missing values ​​is linear interpolation; in time series data, missing values ​​are processed The missing values ​​are estimated by neighboring values, and the linear interpolation formula used is as follows:

[0089] ;

[0090] In the formula, To calculate the missing values ​​using linear interpolation, and For missing values The values ​​of the two time steps before and after ;

[0091] The data normalization operation is to scale the data set after data cleaning to a fixed range between 0 and 1. The data normalization adopts the maximum-minimum normalization method for scaling, and the formula used is as follows:

[0092] ;

[0093] In the formula, is the value calculated by the maximum-minimum normalization method, is the value to be normalized, indicating the No. values, , Respectively The minimum and maximum values ​​of the column data are obtained; the data after final preprocessing are divided into a training set and a test set in a ratio of 8:2, the training set is used to train the current transformer fault warning model, and the test set is used to perform performance testing on the current transformer fault warning model.

[0094] S3. Build a current transformer fault warning model. The steps are as follows: Figure 2 As shown, it specifically includes: a long-term trend analysis module, a conditional enhancement modeling module, a bidirectional LSTM module that integrates the attention mechanism and the convolutional layer, and a self-adjusting fusion module. The specific implementation methods are as follows.

[0095] S31. Construct a long-term trend analysis module. By modeling the long-term trend in the time series, the model ignores the interference of short-term fluctuations and noise on the overall prediction.

[0096] Furthermore, in step S31, the long-term trend analysis module first processes the input data into the trend component of the time series by modeling the long-term trend in the time series. The processing formula is as follows:

[0097] ;

[0098] In the formula, is the original sequence, is the trend part of the forecast window, It is a moving average operation;

[0099] Then the long-term trend analysis module uses two modeling paths to model the trend part, and finally uses the dynamic weighted fusion mechanism for fusion, such as Figure 3 As shown; Modeling path 1 is to use long-term trend path modeling to capture the global change characteristics of the data. The formula operation is:

[0100] ;

[0101] In the formula, is the weight matrix used to extract linear changes, is a bias term used to adjust the center of the model output. To capture the global change characteristics;

[0102] The second modeling path is to use complex trend path modeling to capture complex trends through flexible nonlinear transformations. The formula is:

[0103] ;

[0104] In the formula, is the weight matrix used to extract nonlinear changes, is the bias term, is a nonlinear change function, is a logarithmic transformation, and the calculation formula is , To capture the complex trend changes;

[0105] The feature interaction mechanism is then used to enhance the long-term trend path With complex trend paths The feature interaction between them uses the attention mechanism for interaction, and the operation formula is:

[0106] ;

[0107] In the formula, is the interactive feature, for The transpose operation, is the scaling constant, is the activation function; the feature interaction mechanism dynamically aggregates the features of the two trend modeling paths;

[0108] Then, the dynamic weighted fusion mechanism is used to dynamically adjust the contribution of the two paths according to the characteristics of the input data. First, the statistical information of the historical window is used to generate the weighted coefficient. The formula for generating the weighted coefficient is:

[0109] ;

[0110] In the formula, MLP is a multi-layer perceptron network. is the mean and variance of the historical window, through Ensure that the weight is Between, and satisfy normalization;

[0111] Eventually the dynamic weight and The final trend estimation result is applied to the long-term trend path and interaction characteristics. The fusion formula is:

[0112] ;

[0113] In the formula, and is a dynamic weight, and the trend is simple. Larger, the results are more inclined to model the long-term trend path If the trend is complex Larger, the results are more dependent on the interaction features ; The outputs of long-term trend path modeling and complex trend path modeling are adjusted by weights to obtain the final trend estimation result.

[0114] S32. Construct a conditional enhancement modeling module to model the seasonal components in time series, and accurately capture complex and non-stationary seasonal patterns by combining seasonal feature reference, denoising stage and seasonal signal restoration technology.

[0115] Furthermore, in step S32, the feature combination strategy first divides the historical window into Divide into multiple segments , ,..., , each piece contains H time points, and then the statistical information of each piece is calculated from the historical window to obtain the mean of each piece and variance , the calculation formulas for the mean and variance are as follows:

[0116] , , where is the number of fragments, is the mean of the historical window, indicating the center of historical seasonal fluctuations. is the variance of the historical window, indicating the magnitude of seasonal fluctuations;

[0117] Then, a multi-layer perceptron is used to aggregate the features of each segment to generate global historical statistical features, thereby generating conditional inputs, such as Figure 4 As shown, the aggregation formula is as follows:

[0118] , , where is the mean prediction of the target window, For the variance prediction of the target window, adjust the condition input according to the denoising status and use the current denoising step number and noise scheduling parameters Dynamically adjust the condition input, the formula is as follows:

[0119] , , where is Gaussian noise, used to simulate uncertainty;

[0120] The dynamic contribution balance is then combined and the final output is generated according to the denoising step. And historical features generate weights, the calculation formula is as follows:

[0121] ;

[0122] In the formula, is the weight of the current restored seasonal signal, The weight of the condition input; used to dynamically adjust the condition input With the current reduction season signal Contribution ratio;

[0123] According to the weighted combination condition input and restored seasonal signals, the final output is the dynamically combined harmonic seasonal forecast. The calculation formula is:

[0124] ;

[0125] In the formula, is the intermediate result of denoising in the denoising process, is the output after dynamic combination; the output of each denoising step will be used as the input of the next step, and the complete harmonic seasonal forecast will be gradually restored through multiple iterations.

[0126] S33. Construct a bidirectional LSTM module that integrates the attention mechanism and the convolutional layer. The convolutional layer extracts local features and periodic patterns. The attention mechanism focuses on key time steps. The bidirectional LSTM combines forward and reverse information flows to fully capture short-term and long-term dependencies in the time series.

[0127] Furthermore, in step S33, Figure 5 As shown in the figure, the convolution operation is to extract local features by sliding a convolution kernel and inputting time series data. The formula is as follows:

[0128] ;

[0129] In the formula, is the eigenvalue at the tth moment in the input sequence, is the output feature after the convolution operation, is the weight of the convolution kernel, is the ReLU activation function, is the bias term of the convolution layer; the size of the convolution kernel is , then the convolution operation will slide a length of The window is weighted by summing the data in each window to extract the features of the area. Perform convolution calculation;

[0130] The self-attention mechanism then assigns different weights to different parts of the input sequence, thereby focusing on the most meaningful information, where the query matrix in the self-attention mechanism is , key matrix and the value matrix From the same matrix ,Right now , , ; The calculation formula of the attention mechanism at time t is as follows:

[0131] ;

[0132] In the formula, is the scaling factor; the final output of the self-attention mechanism contains , each The input weighted by the self-attention weight , the output of the self-attention mechanism is used as the input of the subsequent bidirectional LSTM;

[0133] Two-Way From positive With reverse It is composed of two independent The difference between the layers is that they process data in different orders. From left to right in the sequence, that is, from time step arrive Processing, gradually updating its hidden state from the past to the present, reverse From right to left in the sequence, that is, from time step arrive Processing, gradually updating its hidden state from the future to the past, in a bidirectional In, two The layers learn the dependencies of time series data in the forward and reverse directions respectively. Each has its own independent parameters and calculation process; Starting from the first step of the sequence and processing the input sequence backward step by step, the calculation process is as follows:

[0134] ;

[0135] In the formula, For a positive The output, The hidden state of the previous moment, positive As one of the inputs at the current moment, is the cell state at the previous moment;

[0136] Reverse Starting from the last time step of the sequence, the input sequence is processed step by step forward. The calculation process is as follows:

[0137] ;

[0138] In the formula, For a layer of reverse The output, The hidden state at the next moment, reverse As one of the inputs at the current moment, is the cell state at the next moment;

[0139] In one-way In the paper, the forget gate and cell state are improved. First, the forget gate is improved by introducing a multi-scale mechanism to dynamically adjust the forgetting rate at different time steps, and use a multi-layer perceptron to dynamically calculate the forgetting rate at different time scales. The formula is as follows:

[0140] ;

[0141] In the formula, To improve the output of the forget gate, is the weight matrix, is the bias term, for The hidden state of the moment, is the forgetting weight calculated by multi-scale features, is an adjustable coefficient that helps the model finely control the forgetting process. is the activation function; secondly, the cell state is improved to use feedback cell state, introducing a feedback mechanism so that the cell state is not only affected by the current input, but also updated in combination with the past cell state. The formula is as follows:

[0142] ;

[0143] In the formula, To provide feedback on the cell state at the previous moment and further enhance the memory ability of the cell state, is the output of the input gate, which determines the current input How much information is written to the cell state, for Feedback cell status at all times, for Feedback cell status at all times;

[0144] The outputs of the forward and reverse LSTM are concatenated at each time step, and the concatenation formula is:

[0145] ;

[0146] In the formula, is the matrix concatenation operation, Two-Way Final output:

[0147] Finally, after the last fully connected layer, the bidirectional fusion attention mechanism and the convolutional layer are calculated. The output value of the module, the fully connected calculation formula is as follows:

[0148] ;

[0149] In the formula, is the weight of the fully connected layer, is the bias term of the fully connected layer, To integrate the attention mechanism with the bidirectional convolutional layer The final output of the module.

[0150] S34. Construct a self-adjusting fusion module. The early warning model automatically adjusts the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module that integrates the attention mechanism and the convolutional layer to the final result.

[0151] Furthermore, in step S34, the self-adjusting fusion module automatically adjusts the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolutional layer to the final result. The self-adjusting fusion module performs weight learning during model training and dynamically adjusts the influence of each module under real-time warning conditions. The calculation formula of the self-adjusting fusion module is as follows:

[0152] ;

[0153] In the formula, , , is the weight parameter, obtained during model training, where , This is the final prediction result of the fault warning model.

[0154] S4. Training the current transformer fault warning model, wherein the training includes setting hyperparameters. As the number of iterations increases, the accuracy of the fault warning model on the validation set becomes higher and higher.

[0155] Furthermore, in step S4, the hyperparameters of the model during training are set as follows: the learning rate is set to 0.0002, the number of iterations Epochs is 1000, the batch size batch_size is 128, the optimizer uses gradient descent SGD, and to prevent the model from overfitting, the Dropout rate is set to 0.3.

[0156] S5. Testing the current transformer fault warning model. The test is to use a test set to test the accuracy of the current transformer fault warning model to determine whether the fault warning model can be used in an actual environment.

[0157] Furthermore, in step S5, in order to make the model have a higher prediction accuracy in the actual environment, it is very necessary to use the test set to train the trained samples, such as Figure 6 As shown in the figure above, the early warning model uses the values ​​of the previous 60 minutes to predict the failure situation in the next 30 minutes. Figure 6 The following part shows the comparison between the predicted value of the fault warning model and the true value, the predicted value of the recurrent neural network and the predicted value of the LSTM network. Figure 6 It can be found that the current transformer fault warning model proposed in this application has high prediction accuracy and can be applied in practical environments.

[0158] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A current transformer fault early warning method, characterized in that: The following steps are involved: S1. Collecting the operating status information of the current transformer, including electrical parameters, temperature, mechanical properties and insulation resistance information; S2. Preprocess the collected operation status information data, clean and normalize the collected operation status data to ensure the accuracy and consistency of the data, thereby improving the data quality and ensuring the reliability of subsequent analysis; S3. A current transformer fault warning model is proposed. The specific method includes: S31. Build a long-term trend analysis module to model the long-term trend in the time series, so that the model can remove the interference of short-term fluctuations and noise and focus on long-term trend changes; S32. Construct a conditional enhancement modeling module to model the seasonal component in the time series, and accurately capture complex and non-stationary seasonal patterns by combining seasonal feature reference, denoising stage and seasonal signal restoration technology; S33. Construct a bidirectional LSTM module that integrates the attention mechanism and the convolutional layer. The convolutional layer extracts local features and periodic patterns, the attention mechanism focuses on key time steps, and the bidirectional LSTM combines forward and reverse information flows to fully capture short-term and long-term dependencies in the time series. S34, construct a self-adjusting fusion module, the early warning model automatically adjusts the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolution layer to the final result; S4. Train the current transformer fault warning model. During the training process, perform iterative optimization according to the set hyperparameters. By gradually increasing the number of training times, improve the prediction accuracy of the fault warning model on the validation set, and ensure that the model gradually converges to the optimal state. S5. Test the current transformer fault warning model. Use the test set to test the trained current transformer fault warning model, evaluate its prediction performance in the actual environment, and verify whether it has practical application capabilities.

2. A current transformer fault early warning method according to claim 1, characterized in that: In step S1, the electrical parameters include the current value obtained by the current sensor, the voltage value obtained by the voltage sensor and the power value, the temperature parameters include the coil temperature inside the transformer, the shell temperature and the ambient temperature value, the mechanical parameters are the vibration value obtained by the vibration sensor and the noise value obtained by the noise sensor, the insulation resistance parameter is the insulation resistance value obtained by the insulation resistance tester, and the acquisition frequency is set to once every 5 minutes.

3. A current transformer fault early warning method according to claim 2, characterized in that: In step S31, the long-term trend analysis module models the long-term trend in the time series by first processing the input data into the trend component of the time series. The processing formula is as follows: In the formula, X T is the original sequence, is the trend part of the prediction window, and Avg is the moving average operation; Then the long-term trend analysis module uses two modeling paths to model the trend part, and finally uses the dynamic weighted fusion mechanism for fusion; modeling path one uses the long-term trend path modeling to capture the global change characteristics of the data, and the formula operation is: Where W linear is the weight matrix used to extract linear changes, b linear is a bias term used to adjust the center of the model output. To capture the global change characteristics; The second modeling path is to use complex trend path modeling to capture complex trends through flexible nonlinear transformations. The formula is: Where W nonlinear is the weight matrix used to extract nonlinear changes, b nonlinear is the bias term, is a nonlinear change function, is a logarithmic transformation, and the calculation formula is g(x)=log(x+∈), To capture the complex trend changes; The feature interaction mechanism is then used to enhance the long-term trend path With complex trend paths The feature interaction between them uses the attention mechanism for interaction, and the operation formula is: Where, T fusion is the interactive feature, T nonlinear The transpose operation, is the scaling constant, and softmax is the activation function; the feature interaction mechanism dynamically aggregates the features of the two trend modeling paths; Then, the dynamic weighted fusion mechanism is used to dynamically adjust the contribution of the two paths according to the characteristics of the input data. First, the statistical information of the historical window is used to generate the weighted coefficient. The formula for generating the weighted coefficient is: In the formula, MLP is a multi-layer perceptron network. is the mean and variance of the historical window. Softmax is used to ensure that the weight is between [0, 1] and satisfies normalization. Finally, the dynamic weights λ1 and λ2 are applied to the long-term trend path and the interaction characteristics to obtain the final trend estimation result. The fusion formula is: f θ (XT)=λ1·T linear (X T )+λ2·T fusion In the formula, λ1 and λ2 are dynamic weights. If the trend is simple, λ1 is larger, and the result is more inclined to long-term trend path modeling. If the trend is complex, λ2 is larger and the result depends more on the interaction feature T fusion ;f θ (X T ) is the output of long-term trend path modeling and complex trend path modeling, and the final trend estimation result is obtained through weight adjustment.

4. A current transformer fault early warning method according to claim 3, characterized in that: In step S32, the feature combination strategy first divides the historical window into Divide into multiple segments x L，1 、x L，2 ,...,x L，n , each piece contains H time points, and then the statistical information of each piece is calculated from the historical window to obtain the mean μ of each piece L and variance The calculation formulas for the mean and variance are as follows: In the formula, i is the number of fragments, μ L is the mean of the historical window, indicating the center of historical seasonal fluctuations. is the variance of the historical window, indicating the magnitude of seasonal fluctuations; Then, a multi-layer perceptron is used to aggregate the features of each fragment to generate global historical statistical features, thereby generating conditional input. The aggregation formula is as follows: μ T =MLP μ (μ L ), In the formula, μ T is the mean prediction of the target window, σ T is the variance prediction of the target window, adjusting the conditional input according to the state of denoising, using the current denoising step number k and the noise scheduling parameter α k Dynamically adjust the condition input, the formula is as follows: Where z is Gaussian noise, which is used to simulate uncertainty; The dynamic contribution balance is then combined with the final output generation, and the weights are generated according to the denoising step k and historical features. The calculation formula is as follows: In the formula, ρ1 is the weight of the current restored seasonal signal, and ρ2 is the weight of the conditional input; Used to dynamically adjust conditional input With the current reduction season signal Contribution ratio; According to the weighted combination condition input and restored seasonal signals, the final output is the dynamically combined harmonic seasonal forecast. The calculation formula is: In the formula, is the intermediate result of denoising process, X θ is the output after dynamic combination; the output of each denoising step will be used as the input of the next step, and the complete harmonic seasonal forecast will be gradually restored through multiple iterations.

5. A current transformer fault early warning method according to claim 4, characterized in that: In step S33, the convolution operation is performed by sliding a convolution kernel and inputting time series data to perform calculations, thereby extracting local features. The formula is as follows: In the formula, x t is the eigenvalue at the tth moment in the input sequence, is the output feature after the convolution operation, W conv is the weight of the convolution kernel, ReLU(·) is the ReLU activation function, b conv is the bias term of the convolution layer; the size of the convolution kernel is k, then the convolution operation will slide a window of length k in the input sequence, extract the features of the window area covered by the convolution kernel by weighted summing of the data in each window, and perform convolution calculation at each time t; Then, different weights are assigned to different parts of the input sequence through the self-attention mechanism, so as to focus on the most meaningful information. The query matrix Q, key matrix K, and value matrix V in the self-attention mechanism come from the same matrix X, that is, The calculation formula of the attention mechanism at time t is as follows: In the formula, is the scaling factor; the final output of the self-attention mechanism contains [a1, a2, ..., a T ], each a t The input x is weighted according to the self-attention weight t , the output of the self-attention mechanism is used as the input of the subsequent bidirectional LSTM; Bidirectional LSTM consists of forward LSTM and reverse LSTM, which are two independent LSTM layers. The difference is that they process data in different orders. The forward LSTM processes from left to right of the sequence, that is, from time step t=1 to t=T, and gradually updates its hidden state from the past to the present. The reverse LSTM processes from right to left of the sequence, that is, from time step t=T to t=1, and gradually updates its hidden state from the future to the past. In the bidirectional LSTM, the two LSTM layers learn the dependency of time series data in the forward and reverse directions respectively. The LSTM in each direction has its independent parameters and calculation process; the forward LSTM starts from the first step of the sequence and gradually processes the input sequence backwards. The calculation process is as follows: In the formula, is the output of a layer of forward LSTM, The hidden state of the previous moment is used as one of the inputs of the current moment by the forward LSTM. is the cell state at the previous moment; The reverse LSTM starts from the last time step of the sequence and processes the input sequence step by step. The calculation process is as follows: In the formula, is the output of a layer of reverse LSTM, The hidden state of the next moment, the reverse LSTM takes it as one of the inputs of the current moment, is the cell state at the next moment; In the unidirectional LSTM, the forget gate and cell state are improved. First, the forget gate is improved by introducing a multi-scale mechanism to dynamically adjust the forgetting rate at different time steps, and use a multi-layer perceptron to dynamically calculate the forgetting rate at different time scales. The formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f +γ·g t ); In the formula, f t To improve the output of the forget gate, W f is the weight matrix, b f is the bias term, h t-1 is the hidden state at time t-1, g t is the forgetting weight calculated by multi-scale features, γ is an adjustable coefficient that helps the model finely control the forgetting process, and σ is the activation function; secondly, the cell state is improved to use feedback cell state, introducing a feedback mechanism so that the cell state is not only affected by the current input, but also updated in combination with the past cell state. The formula is as follows: C t =Feedback(C t-1 )+f t ·C t-1 +i t ·c t ; In the formula, Feedback(·) is the feedback of the cell state at the previous moment, which further enhances the memory ability of the cell state. t is the output of the input gate, which determines the current input c t How much information is written to the cell state, C t-1 is the feedback cell state at time t-1, C t is the feedback cell state at time t; The outputs of the forward and reverse LSTM are concatenated at each time step, and the concatenation formula is: Where [·] is the matrix concatenation operation, The final output of the bidirectional LSTM; Finally, the output value of the bidirectional LSTM module that integrates the attention mechanism and the convolutional layer is calculated after passing through the last fully connected layer. The fully connected calculation formula is as follows: Where W out is the weight of the fully connected layer, b out is the bias term of the fully connected layer, y t It is the final output of the bidirectional LSTM module that integrates the attention mechanism and the convolutional layer.

6. A current transformer fault early warning method according to claim 5, characterized in that: In step S34, the self-adjusting fusion module automatically adjusts the contribution of the final trend estimation result of the long-term trend analysis module, the seasonal prediction of the conditional enhancement modeling module, and the output result of the bidirectional LSTM module integrating the attention mechanism and the convolutional layer to the final result. The self-adjusting fusion module performs weight learning during model training and dynamically adjusts the influence of each module under real-time warning conditions. The calculation formula of the self-adjusting fusion module is as follows: In the formula, λ1, λ2, and λ3 are weight parameters obtained during model training, where λ1+λ2+λ3=1, and y′ is the final prediction result of the fault warning model.

Images