Transformer electric-heat-vibration signal prediction and anomaly detection method

Abstract

The invention discloses a transformer electric-thermal-vibration signal prediction and anomaly detection method, which comprises the following steps of S1, accurately quantifying delay time among transformer electric-thermal-vibration signals by using a cross-correlation analysis method, re-calibrating a data set according to the delay time, aligning different signals on a time axis, and normalizing data; s2, establishing a transformer data prediction model based on a Transform-KAN network, taking the aligned transformer current, load rate, oil temperature, environment temperature and vibration fundamental frequency amplitude data as model input, and predicting transformer operation data; s3, sharing an encoder in the prediction model, learning a normal data rule, and reconstructing a transformer operation signal; and S4, detecting anomalies from two aspects of reconstruction errors and prediction errors by using a joint optimization transformer anomaly detection method. According to the method, the delay time among the electric-thermal-vibration signals can be accurately quantified, the prediction precision of the state data of the transformer is improved, and the reliability of abnormal detection of the transformer is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer operation data prediction and abnormality detection, and specifically relates to a transformer electric-thermal-vibration signal prediction and abnormality detection method. Background Art

[0002] Transformers are important electrical equipment in power systems, and their operating status is directly related to the safety and reliability of power systems. In actual operation, transformers are easily disturbed by factors such as load and environment, and comprehensive and accurate monitoring and evaluation of equipment status are challenging. The dynamic changes of transformer electrical, thermal, and vibration signals contain rich physical information. The relationship and characteristic analysis of these signals can reveal the internal operating mechanism of the transformer and are of great significance for predicting its operating status and performing abnormality detection.

[0003] At present, the use of data-driven methods to model and analyze transformer states has become a research hotspot. However, due to the significant time series correlation, nonlinear characteristics and multivariate coupling characteristics of transformer operation signals, most existing methods consider the correlation between different signal fields of the transformer, but ignore the time series characteristics of the signal, making it difficult to quantify the precise delay time between signals; most of the current complex time series data prediction methods are insufficient in capturing long-term dependencies, and are prone to gradient vanishing or gradient explosion; in addition, the current anomaly detection for multivariate time series data is mostly based on prediction and reconstruction, which cannot take into account the unpredictability and time dependency of time series at the same time. Summary of the invention

[0004] The purpose of the present invention is to provide a transformer electro-thermal-vibration signal prediction and anomaly detection method, which can accurately quantify the delay time between electro-thermal-vibration signals, improve the accuracy of transformer state data prediction, and enhance the reliability of transformer anomaly detection.

[0005] To achieve the above object, the present invention provides a transformer electric-thermal-vibration signal prediction and abnormality detection method, comprising the following steps:

[0006] S1. Use the cross-correlation analysis method to accurately quantify the delay time between the transformer's electrical, thermal, and vibration signals, recalibrate the data set based on the delay time, align different signals on the time axis, and normalize the data;

[0007] S2. Using the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN, a transformer data prediction model based on the Transformer-KAN network is established. The aligned transformer current, load rate, oil temperature, ambient temperature and vibration fundamental frequency amplitude data are used as model input to predict the transformer operation data.

[0008] S3, the encoder in the shared prediction model, learns the normal data rules and reconstructs the transformer operation signal;

[0009] S4. Utilize the jointly optimized transformer anomaly detection method to detect anomalies from both the reconstruction error and prediction error.

[0010] As a further solution of the present invention: in step S1, the delay time between the electric-thermal-vibration signals of the transformer is accurately quantified by using a cross-correlation analysis method. Specifically, the current of the transformer is used as the reference input time series x(t), and the transformer oil temperature and vibration fundamental frequency amplitude are used as the output signal y(t). By calculating the time difference corresponding to the maximum value of the cross-correlation coefficient, the accurate delay time between the electric-thermal-vibration signals is obtained; the steps are:

[0011] 1) Calculate the cross-correlation function of the input time series x(t) and the output series y(t):

[0012]

[0013] Where T is the acquisition time; ξ is the time shift, C xy (ξ) is the cross-covariance of x(t) and y(t), σ(x), σ(y) are the standard deviations of x(t) and y(t);

[0014] 2) |R xy (ξ)| The ξ0 corresponding to the maximum value is used as the signal time difference point number between the input sequence and the output sequence, and the actual delay time is obtained according to the data collection interval;

[0015] Δt=ξ0×T d

[0016] Where, T d is the data collection interval, Δt is the actual delay time;

[0017] Then, according to the delay time between the electrical, thermal and vibration signals, the lag signal data is shifted forward by ξ0 time points, the data set is realigned, and normalized as follows:

[0018]

[0019] In the formula, x is the value of each transformer parameter, x min 、x max are the minimum and maximum values ​​corresponding to each parameter respectively; It is a normalized value, and its value is between 0 and 1.

[0020] As a further solution of the present invention: in step S2, the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN are combined to establish a transformer data prediction model based on the Transformer-KAN network, and the specific steps include:

[0021] 1) Use the self-attention mechanism in the Transformer encoder to learn the relationship between each time step in the sequence. Multi-head attention weights each time step of the input, and after processing by multiple layers of encoders, generates the output of the last time step, which contains the information of the entire sequence for subsequent predictions;

[0022] 2) The KAN network receives the features of the last time step output by the Transformer encoder, performs nonlinear mapping, captures local complex relationships, and generates the final prediction output;

[0023] The aligned transformer current, load factor, oil temperature, ambient temperature and vibration fundamental frequency amplitude data are used as model inputs, and a composite loss function is used, which combines the Huber loss function and R 2 Loss function, training prediction model; Huber loss function is defined as follows:

[0024]

[0025] In the formula, y is the actual value, is the predicted value, δ is the preset threshold;

[0026] R 2 The loss function is defined as follows:

[0027]

[0028] In the formula, SS res is the residual sum of squares, SS tot is the total sum of squares,

[0029] When the prediction accuracy of the model increases, the Huber loss decreases; when the correlation between the model prediction and the actual value increases, R 2 The closer the score is to 1, the smaller the score is. 2 Score, R 2 The score is subtracted from 1 and then added as a penalty term to the Huber loss function as the final loss function of the prediction model:

[0030]

[0031] As a further solution of the present invention: Step S3 specifically comprises the following steps:

[0032] 1) Use a linear layer to map the output of the Transformer encoder to the original data dimension;

[0033] 2) The reconstruction layer reconstructs the output of the next time step to obtain the reconstructed data; in order to minimize the reconstruction error between the input and output, the mean square error is used as the standard for measuring the reconstruction error, that is, the reconstruction loss, and the calculation formula is as follows:

[0034]

[0035] In the formula, r i is to reconstruct the data, o i is the original data, and N is the number of data points.

[0036] As a further solution of the present invention: Step S4 adopts a joint optimization method to comprehensively predict and reconstruct the two parts of loss, weightedly combine the two parts of loss function, and consider the prediction accuracy and the sensitivity of anomaly detection at the same time; the formula is as follows:

[0037] Loss = αLoss pre +(1-α)Loss rec

[0038] In the formula, Loss pre is the composite loss function of the prediction part, Loss rec is the loss function of the reconstruction part, and the parameter α is the weight of the prediction loss and reconstruction loss in the total loss function;

[0039] After training the prediction-reconstruction joint model according to the loss function Loss, anomaly detection is performed. At each timestamp, the reconstruction module and the prediction module generate reconstruction error and prediction error respectively. If the corresponding reconstruction error or prediction error at a certain timestamp is greater than the anomaly threshold, the timestamp is marked as "abnormal", otherwise it is marked as "normal".

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The delay time between electrical, thermal and vibration signals is accurately quantified through cross-correlation analysis, the mutual relationship between signals is revealed, and the data sets are aligned; the global characteristic modeling under complex multivariate relationships is optimized by combining the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN, which significantly improves the ability to express the complex interactive characteristics of transformer current, oil temperature and vibration signals. The accuracy of transformer status data prediction is improved through the synergy of prediction and anomaly detection; the comprehensive reconstruction error and prediction error effectively capture abnormal fluctuations and changes, and enhance the reliability of anomaly detection; it can provide effective support for intelligent monitoring of the operating status of power transformers. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is an overall flow chart of the transformer electric-thermal-vibration signal prediction and abnormality detection method of the present invention;

[0043] Figure 2 It is a graph of the mutual correlation coefficient between the current, oil temperature and vibration fundamental frequency amplitude of the present invention;

[0044] Figure 3 It is a structural diagram of the transformer data prediction model of the present invention;

[0045] Figure 4 This is a comparison of the current curve prediction effect of the present invention;

[0046] Figure 5 It is an abnormal point identified by the present invention based on the reconstruction error and the prediction error. DETAILED DESCRIPTION

[0047] The present invention will be further described below by way of examples.

[0048] like Figure 1 As shown, a transformer electric-thermal-vibration signal prediction and abnormality detection method includes the following steps:

[0049] S1. Use the cross-correlation analysis method to accurately quantify the delay time between the transformer's electrical, thermal, and vibration signals, recalibrate the data set based on the delay time, align different signals on the time axis, and normalize the data;

[0050] S2. Using the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN, a transformer data prediction model based on the Transformer-KAN network is established, such as Figure 3 As shown; the aligned transformer current, load factor, oil temperature, ambient temperature and vibration fundamental frequency amplitude data are used as model input to predict the transformer operation data;

[0051] S3, the encoder in the shared prediction model, learns the normal data rules and reconstructs the transformer operation signal;

[0052] S4. Utilize the jointly optimized transformer anomaly detection method to detect anomalies from both the reconstruction error and prediction error.

[0053] Furthermore, in step S1, the delay time between the transformer electric-thermal-vibration signals is accurately quantified using a cross-correlation analysis method. Specifically, the transformer current is used as a reference input time series x(t), and the transformer oil temperature and vibration fundamental frequency amplitude are used as output signals y(t). By calculating the time difference corresponding to the maximum value of the cross-correlation coefficient, the accurate delay time between the electric-thermal-vibration signals is obtained. The steps are:

[0054] 1) Calculate the cross-correlation function of the input time series x(t) and the output series y(t):

[0055]

[0056] Where T is the acquisition time; ξ is the time shift, C xy (ξ) is the cross-covariance of x(t) and y(t), σ(x), σ(y) are the standard deviations of x(t) and y(t);

[0057] 2) |R xy (ξ)| The ξ0 corresponding to the maximum value is used as the signal time difference point number between the input sequence and the output sequence, and the actual delay time is obtained according to the data collection interval;

[0058] Δt=ξ0×T d

[0059] Where, T d is the data collection interval, Δt is the actual delay time;

[0060] Then, according to the delay time between the electrical, thermal and vibration signals, the lag signal data is shifted forward by ξ0 time points, the data set is realigned, and normalized as follows:

[0061]

[0062] In the formula, x is the value of each transformer parameter, x min 、x max are the minimum and maximum values ​​corresponding to each parameter respectively; It is a normalized value, and its value is between 0 and 1.

[0063] The current is used as the transformer electrical signal indicator to reflect the transformer load condition and electrical performance; the oil temperature is used as the transformer thermal signal indicator to monitor the transformer temperature and thermal state; the vibration fundamental frequency amplitude is used as the transformer vibration signal indicator to reflect the mechanical movement inside the transformer.

[0064] like Figure 2As shown in the figure, the time delay corresponding to the maximum value of the mutual correlation coefficient between current and oil temperature is 15, that is, when the oil temperature signal is shifted forward by 15 sampling points, the similarity between the two signals reaches the maximum, which is 0.94156. Given that each sampling point in this data set is 5 minutes apart, it can be obtained that the oil temperature signal lags the current by 75 minutes. The maximum value of the mutual correlation coefficient between current and vibration is 0.99857, and the corresponding time delay is 0. It can be almost considered that the vibration signal and the current signal are synchronized in time.

[0065] Furthermore, in step S2, the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN are combined to establish a transformer data prediction model based on the Transformer-KAN network. The specific steps include:

[0066] 1) Use the self-attention mechanism in the Transformer encoder to learn the relationship between each time step in the sequence. Multi-head attention weights each time step of the input, and after processing by multiple layers of encoders, generates the output of the last time step, which contains the information of the entire sequence for subsequent predictions;

[0067] 2) The KAN network receives the features of the last time step output by the Transformer encoder, performs nonlinear mapping, captures local complex relationships, and generates the final prediction output;

[0068] The aligned transformer current, load factor, oil temperature, ambient temperature and vibration fundamental frequency amplitude data are used as model inputs, and a composite loss function is used, which combines the Huber loss function and R 2 Loss function, training prediction model; Huber loss function is defined as follows:

[0069]

[0070] In the formula, y is the actual value, is the predicted value, and δ is the preset threshold. When the prediction error is less than or equal to the threshold, the Huber loss function uses MSE to emphasize the square of the error, making the model more sensitive to small errors. When the prediction error is greater than the threshold, MAE is used to make the model more robust.

[0071] R 2 The loss function is defined as follows:

[0072]

[0073] In the formula, SS res is the residual sum of squares, SS tot is the total sum of squares,

[0074] When the model's prediction accuracy improves, the Huber loss decreases; when the correlation between the model's prediction and the actual value increases, R 2 The score increases as it approaches 1; to minimize the Huber loss and maximize R 2 Score, R 2 The score is subtracted from 1 and then added as a penalty term to the Huber loss function as the composite loss function of the prediction model:

[0075]

[0076] Comparison of current curve prediction effect Figure 4 As shown in the figure. The predicted curve is highly consistent with the actual curve. The prediction model can effectively capture the changes in feature parameters and has high prediction accuracy. In order to verify the superiority of the proposed model, the more mature models are selected for comparison, including Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), Bi-directional Long Short-Term Memory (BiLSTM), Gated Recurrent Unit (GRU) and Transformer model. The comparison results are shown in the following table.

[0077]

[0078]

[0079]

[0080] Furthermore, step S3 specifically includes the following steps:

[0081] 1) Use a linear layer to map the output of the Transformer encoder to the original data dimension;

[0082] 2) The reconstruction layer reconstructs the output of the next time step to obtain the reconstructed data; in order to minimize the reconstruction error between the input and output, the mean square error is used as the standard for measuring the reconstruction error, that is, the reconstruction loss, and the calculation formula is as follows:

[0083]

[0084] In the formula, r i is to reconstruct the data, o i is the original data, and N is the number of data points.

[0085] Furthermore, step S4 adopts a joint optimization method to comprehensively predict and reconstruct the two parts of loss, weightedly combine the two parts of loss function, and consider the prediction accuracy and the sensitivity of anomaly detection at the same time; the formula is as follows:

[0086] Loss = αLoss pre +(1-α)Loss rec

[0087] In the formula, Loss pre is the composite loss function of the prediction model, Loss rec is the loss function of the reconstruction part, and the parameter α is the weight of the prediction loss and reconstruction loss in the total loss function;

[0088] After training the prediction-reconstruction joint model according to the loss function Loss, anomaly detection is performed. At each timestamp, the reconstruction module and the prediction module generate reconstruction error and prediction error respectively. If the corresponding reconstruction error or prediction error at a certain timestamp is greater than the anomaly threshold, the timestamp is marked as "abnormal", otherwise it is marked as "normal".

[0089] like Figure 5 As shown in the figure, taking current as an example, the abnormal points are identified based on the reconstruction error and prediction error. Combined with the actual operation of the transformer, the abnormal points correspond to the moments when the load changes significantly. This effectively shows that the anomaly detection module can accurately capture abnormal fluctuations and accurately identify anomalies.

[0090] After executing the relevant steps, the prediction and anomaly detection of transformer status data can be realized.

Claims

1. A transformer electric-thermal-vibration signal prediction and abnormality detection method, characterized in that: The following steps are involved: S1. Use the cross-correlation analysis method to accurately quantify the delay time between the transformer's electrical, thermal, and vibration signals, recalibrate the data set based on the delay time, align different signals on the time axis, and normalize the data; S2. Using the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN, a transformer data prediction model based on the Transformer-KAN network is established. The aligned transformer current, load rate, oil temperature, ambient temperature and vibration fundamental frequency amplitude data are used as model input to predict the transformer operation data. S3, the encoder in the shared prediction model, learns the normal data rules and reconstructs the transformer operation signal; S4. Utilize the jointly optimized transformer anomaly detection method to detect anomalies from both the reconstruction error and prediction error.

2. A transformer electro-thermal-vibration signal prediction and anomaly detection method according to claim 1, characterized in that: In step S1, the delay time between the transformer electric-thermal-vibration signals is accurately quantified by using the cross-correlation analysis method. Specifically, the transformer current is used as the reference input time series x(t), and the transformer oil temperature and vibration fundamental frequency amplitude are used as the output signal y(t). By calculating the time difference corresponding to the maximum value of the cross-correlation coefficient, the accurate delay time between the electric-thermal-vibration signals is obtained. The steps are: 1) Calculate the cross-correlation function of the input time series x(t) and the output series y(t): Where T is the acquisition time; ξ is the time shift, C xy (ξ) is the cross-covariance of x(t) and y(t), σ(x), σ(y) are the standard deviations of x(t) and y(t); 2) |R xy (ξ)| The ξ0 corresponding to the maximum value is used as the signal time difference point number between the input sequence and the output sequence. According to the data collection interval, the actual delay time is obtained; Δt=ξ0×T d Where, T d is the data collection interval, Δt is the actual delay time; Then, according to the delay time between the electrical, thermal and vibration signals, the lag signal data is shifted forward by ξ0 time points, the data set is realigned, and normalized as follows: In the formula, x is the value of each transformer parameter, x min 、x max are the minimum and maximum values ​​corresponding to each parameter respectively; It is a normalized value, and its value is between 0 and 1.

3. A transformer electric-thermal-vibration signal prediction and abnormality detection method according to claim 1 or 2, characterized in that: In step S2, the self-attention mechanism of the Transformer encoder and the nonlinear feature decomposition capability of KAN are combined to establish a transformer data prediction model based on the Transformer-KAN network. The specific steps include: 1) Use the self-attention mechanism in the Transformer encoder to learn the relationship between each time step in the sequence. Multi-head attention weights each time step of the input, and after processing by multiple layers of encoders, generates the output of the last time step, which contains the information of the entire sequence for subsequent predictions; 2) The KAN network receives the features of the last time step output by the Transformer encoder, performs nonlinear mapping, captures local complex relationships, and generates the final prediction output; The aligned transformer current, load factor, oil temperature, ambient temperature and vibration fundamental frequency amplitude data are used as model inputs, and a composite loss function is used, which combines the Huber loss function and R 2 Loss function, training prediction model; Huber loss function is defined as follows: In the formula, y is the actual value, is the predicted value, δ is the preset threshold; R 2 The loss function is defined as follows: In the formula, SS res is the residual sum of squares, SS tot is the total sum of squares, When the model's prediction accuracy improves, the Huber loss decreases; when the correlation between the model's prediction and the actual value increases, R 2 The closer the score is to 1, the smaller the score is. 2 Score, R 2 The score is subtracted from 1 and then added as a penalty term to the Huber loss function as the final loss function of the prediction model:

4. A transformer electro-thermal-vibration signal prediction and anomaly detection method according to claim 3, characterized in that: Step S3 specifically includes: 1) Use a linear layer to map the output of the Transformer encoder to the original data dimension; 2) The reconstruction layer reconstructs the output of the next time step to obtain the reconstructed data; in order to minimize the reconstruction error between the input and output, the mean square error is used as the standard for measuring the reconstruction error, that is, the reconstruction loss, and the calculation formula is as follows: In the formula, r i is to reconstruct the data, o i is the original data, and N is the number of data points.

5. A transformer electro-thermal-vibration signal prediction and anomaly detection method according to claim 4, characterized in that: Step S4 adopts a joint optimization method to comprehensively predict and reconstruct the two parts of loss, weightedly combine the two parts of loss function, and consider the prediction accuracy and the sensitivity of anomaly detection at the same time; the formula is as follows: Loss=αLoss pre +(1-α)Loss rec In the formula, Loss pre is the composite loss function of the prediction part, Loss rec is the loss function of the reconstruction part, and the parameter α is the weight of the prediction loss and reconstruction loss in the total loss function; After training the prediction-reconstruction joint model according to the loss function Loss, anomaly detection is performed. At each timestamp, the reconstruction module and the prediction module generate reconstruction error and prediction error respectively. If the corresponding reconstruction error or prediction error at a certain timestamp is greater than the anomaly threshold, the timestamp is marked as "abnormal", otherwise it is marked as "normal".

Images