Transformer multi-signal collaborative modeling and anomaly detection method based on dynamic time warping

Abstract

The invention discloses a transformer multi-signal collaborative modeling and anomaly detection method based on dynamic time warping, and aims to improve dynamic time warping and quantify a delay relationship among multiple signals of a transformer by combining with a transformer physical model. According to the method, attention bias is designed based on obtained delay information, a transformer multi-signal feature prediction model based on a Transform model is constructed, theoretical prediction values of feature quantities are calculated according to a physical model of transformer oil temperature and vibration signals, output results of a physical model branch and a data driving branch are dynamically fused through a gating mechanism, and the prediction accuracy of the transformer multi-signal feature prediction model is improved. The multi-semaphore of the transformer is predicted. And reconstructing the multi-signal characteristic quantity of the transformer based on the local characteristics of the transformer data in different time periods, and detecting the abnormity of the transformer in combination with the characteristic distance after dynamic time alignment, the model prediction error and the reconstruction error.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer detection, and specifically relates to a method for collaborative modeling and anomaly detection of multi-signals of a transformer based on dynamic time warping. Background Art

[0002] The proposal and construction of a new power system have put forward higher requirements for the intelligent operation and maintenance as well as overhaul of power transmission and transformation equipment. As an indispensable device in the power grid, a transformer undertakes the key tasks of voltage conversion and electric energy transmission. Once an unplanned outage event occurs in the transformer, it will have a greater impact on social life and the construction of the new power system. Therefore, improving the key technologies for transformer operation and maintenance, real-time monitoring the current operation state of the transformer and accurately identifying abnormal situations, and reducing the number of unplanned outages of equipment are of great significance for maintaining the reliable operation of the transformer and the power grid.

[0003] At present, power operation and maintenance units mainly carry out the state monitoring and diagnosis of transformers based on expert knowledge bases or various industry or enterprise regulations, and regularly inspect and maintain the equipment. However, this regular maintenance method is prone to untimely maintenance, over-maintenance, and resource waste, and it is difficult to reflect the operation state of the transformer in a timely and accurate manner. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method for collaborative modeling and anomaly detection of multi-signals of a transformer based on dynamic time warping, which enhances the ability to distinguish abnormal data by fusing multiple error signals, comprehensively detects abnormal situations of the transformer, reduces the false alarm rate, and improves the accuracy of detection.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A method for collaborative modeling and anomaly detection of multi-signals of a transformer based on dynamic time warping, comprising the following steps:

[0007] S1: Dynamically divide the window length according to the change rate of the electro-thermal-vibration signals of the transformer by using an adaptive sliding window, improve the dynamic time warping algorithm by introducing a delay time constraint term estimated by a physical model, dynamically align the multi-signals of the transformer, and quantify the time delay;

[0008] S2: Based on the dynamic time delay information obtained in step S1, introduce a bias matrix into the attention calculation of the Transformer model to construct a Transformer-KAN multi-signal feature prediction model including delay features;

[0009] S3: Calculate the physical prediction values through the thermodynamic model of the transformer oil temperature and the dynamic model of vibration respectively; use the Transformer-KAN model to extract the data-driven prediction values, and adopt a gating mechanism to dynamically fuse the output results of the physical model branch and the data-driven branch to generate multi-signal prediction values;

[0010] S4: Dynamically adjust the weights of the reconstruction module according to the local fluctuation characteristics of the transformer multi-signals at different time periods, and reconstruct the multi-signal feature quantities by using the weighted reconstruction error;

[0011] S5: Combine the dynamic alignment feature distance, the prediction error obtained based on data-driven, the prediction error obtained based on the physical model, and the multi-signal feature quantity reconstruction error, and calculate the anomaly confidence degree by weighting to detect transformer anomalies.

[0012] Further, in the first step, the dynamic adjustment method of the window length is as follows:

[0013]

[0014] where: W is the window length; β is the adjustment coefficient; is the absolute value of the signal change rate; ε is used to prevent division by zero.

[0015] Further, in the first step, the dynamic time warping algorithm obtains a penalty term by introducing the delay time estimated by the physical model, and constructs a cost function to constrain the deviation between the actual delay and the expected delay of the matching points. The formula is:

[0016] Δt r =t s ×(j―i)

[0017] P(i,j)=|Δt r ―Δt e |

[0018] C(i,j)=|x i ―y j |+λ·P(i,j)

[0019] where: C(i,j) is the cost function; P(i,j) is the penalty term; λ is the penalty coefficient; Δt r is the actual delay time; t s is the sampling interval time; Δt e is the delay time between any two signals in the transformer electro-thermal-vibration signal estimated by the physical model; X = {x1,x2,…x N} and Y = {y1,y2,…y N} are the time series of two signals in the transformer electro-thermal-vibration signal respectively; N is the total number of samplings; i and j are the time points of the two signals respectively.

[0020] Calculate the best matching points between any two signals in the electro-thermal-vibration of the transformer using the dynamic time warping algorithm, convert the difference between the best matching points into the actual time difference, and quantify the dynamic time delay between the signals.

[0021] Furthermore, in the second step, based on the dynamic time delay, construct a bias matrix to obtain the attention weights of the Transformer-KAN multi-signal feature prediction model:

[0022] B = -β·τ

[0023]

[0024] Where: Attn(Q, K, V) is the attention weight; B is the bias matrix; β is the parameter controlling the influence degree of time delay on attention; τ is the dynamic delay time; Q, K, and V are the Query, Key, and Value vectors respectively; d K is the dimension representing the key vector Key.

[0025] Furthermore, in the third step, the thermodynamic model of the transformer oil temperature is:

[0026]

[0027] Where: OT(t) is the oil temperature value corresponding to time t; AT(t) is the ambient temperature at time t; P(t) is the heat energy input at time t; H is the convective heat transfer coefficient; S is the area; C Oil is the equivalent heat capacity of the transformer oil;

[0028] The dynamic model of the transformer vibration is:

[0029]

[0030] Where: v(t) is the vibration amplitude of the transformer box; m is the equivalent mass; δ is the damping coefficient; k is the elastic coefficient.

[0031] Furthermore, in the third step, the multi-signal prediction value is:

[0032] y pre = g·y m +(1 - g)y p

[0033] Where: y pre is the multi-signal prediction value; y m is the data-driven prediction value extracted by the Transformer-KAN multi-signal feature prediction model; y p is the physical prediction value calculated through the thermodynamic model of the transformer oil temperature and the dynamic model of the vibration; g is the gating factor.

[0034] Further, in the fourth step, the method steps for reconstructing multi-signal feature quantities using weighted reconstruction error are as follows:

[0035] S41: Use a Transformer encoder to extract temporal features and obtain a latent representation h of the multi-signal data features of the transformer; reconstruct the input sequence using the Transformer encoder based on h to obtain a reconstruction result;

[0036] S42: For each time step, calculate the local standard deviation within a fixed window; calculate the corresponding weight value according to the weight function:

[0037]

[0038] where: weight(t) is the local weight value; σ t is the local standard deviation; α is a regulation coefficient for balancing the sensitivity of the weight function;

[0039] S43: Use the weighted root mean square error to reconstruct the multi-signal feature quantities; the weighted root mean square error is:

[0040]

[0041] where: r ti is the actual value; r pi is the reconstructed value.

[0042] Further, in the fifth step, the calculation method of the anomaly confidence is as follows:

[0043] S51: Construct positive and negative samples; among them, the positive samples are taken from the sliding window data under the normal operation state of the transformer; the negative samples are samples that deviate from the physical laws artificially added to the normal data;

[0044] S52: Input the positive and negative samples into the Transformer-KAN multi-signal feature prediction model to extract sample features; design a contrast loss function to make the feature distances of the positive samples closer and the feature distances of the negative samples farther. The contrast loss function is:

[0045]

[0046] where: y i,j is the binary label; d i,j is the feature distance; m is a hyperparameter for controlling the minimum distance between negative samples.

[0047] S53: Weight the feature distance obtained from contrast learning, the prediction error obtained from data-driven, the prediction error of the physical model, and the reconstruction error, and calculate the anomaly confidence as:

[0048] S = λ1D + λ2Y m + λ3Y t + λ4R

[0049] Where: D is the feature space distance obtained by contrastive learning calculation; Y m is the prediction error obtained based on data-driven; Y t is the prediction error of the physical model; R is the reconstruction error; λ1, λ2, λ3, λ4 are the weights for adjusting each part.

[0050] The beneficial effects of the present invention are as follows:

[0051] The method for collaborative modeling and anomaly detection of multi-signals of a transformer based on dynamic time warping in the present invention makes full use of the temporal characteristics of signals such as electricity, heat, and vibration of the transformer, combines a dynamic window with physical constraints to accurately capture the dynamic time-varying delay between signals, enabling effective alignment of different signals in the time dimension; designs an attention bias based on delay information on the basis of the Transformer-KAN network, and captures the internal time relationship between each signal through the self-attention mechanism; combines data-driven prediction with prediction based on a physical model, and dynamically fuses through a gating mechanism to achieve more accurate and physically regular prediction of multi-signal quantities of the transformer, providing stronger data support for real-time early warning; the present invention enhances the ability to distinguish abnormal data by fusing multiple error signals, comprehensively detects abnormal conditions of the transformer, reduces the false alarm rate, and improves the accuracy of detection. Description of the Drawings

[0052] In order to make the objectives, technical solutions, and beneficial effects of the present invention clearer, the following drawings are provided by the present invention for illustration:

[0053] Figure 1 is the flowchart of the embodiment of the method for collaborative modeling and anomaly detection of multi-signals of a transformer based on dynamic time warping in the present invention;

[0054] Figure 2 is the dynamic delay change curve graph of oil temperature and vibration fundamental frequency amplitude;

[0055] Figure 3 is the comparison graph of the prediction effect of the oil temperature curve. Detailed Embodiments

[0056] The following further describes the present invention in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not intended to limit the present invention.

[0057] As Figure 1 shown, the method for collaborative modeling and anomaly detection of multi-signals of a transformer based on dynamic time warping in this embodiment includes the following steps.

[0058] S1: Dynamically divide the window length according to the change rate of the transformer's electrical-thermal-vibration signals using an adaptive sliding window. Improve the dynamic time warping algorithm by introducing a delay time constraint term estimated by a physical model to dynamically align the multi-signals of the transformer and quantify the time delay.

[0059] In this embodiment, an adaptive sliding window is used to dynamically divide the window length according to the signal change rate. When the change rate of the signal is high, the corresponding signal fluctuates greatly in a short time, and the window length is reduced to capture the rapid changes of the signal; when the change rate of the signal is low, the corresponding signal changes relatively smoothly, and the window length is extended. In this embodiment, the dynamic adjustment method of the window length is as follows:

[0060]

[0061] Where: W is the window length; β is the adjustment coefficient; is the absolute value of the signal change rate; ε is used to prevent division by zero.

[0062] In this embodiment, the dynamic time warping algorithm (DTW) obtains a penalty term by introducing the delay time estimated by a physical model, constructs a cost function to constrain the deviation between the actual delay and the expected delay of the matching points, and the formula is:

[0063] Δt r = t s ×(j―i)

[0064] P(i,j) = |Δt r ―Δt e |

[0065] C(i,j) = |x i ―y j | + λ·P(i,j)

[0066] Where: C(i,j) is the cost function; P(i,j) is the penalty term; λ is the penalty coefficient; Δt r is the actual delay time; t s is the sampling interval time; Δt e is the delay time between any two signals in the transformer's electrical-thermal-vibration signals estimated by the physical model; X = {x1,x2,…x N} and Y = {y1,y2,…y N} are the time series of two signals in the transformer's electrical-thermal-vibration signals respectively; N is the total number of samples; i and j are the time points of the two signals respectively.

[0067] Use the dynamic time warping algorithm to calculate the best matching points between any two signals in the transformer's electrical-thermal-vibration, convert the difference between the best matching points into the actual time difference, and quantify the dynamic time delay between the signals.

[0068] As shown Figure 2 in the figure, the first sub - figure in the figure is the original signal without DTW processing. It can be seen from the figure that the vibration signal fluctuates more violently than the oil temperature signal, and there is a certain misalignment between the two at the time point. The second sub - figure shows the result after DTW windowing matching and smoothing processing, adjusting the vibration signal to better align with the oil temperature signal as a whole. The third sub - figure is the dynamic delay change between the vibration signal and the oil temperature signal, intuitively reflecting that the dynamic lag characteristics are different in different time periods, that is, the alignment point of the vibration signal relative to the oil temperature signal is constantly changing. This is mainly because the coupling relationship between oil temperature and vibration is not completely strictly fixed and is affected by working conditions changes or the environment.

[0069] S2: Based on the dynamic time delay information obtained in step S1, introduce a bias matrix in the attention calculation of the Transformer model to construct a Transformer - KAN multi - signal feature prediction model containing delay features.

[0070] In this embodiment, the delay features after improved DTW processing and the transformer multi - signal data at the corresponding time steps are used as the input sequence of the model. A bias matrix B is introduced in the attention calculation of the model to construct a transformer multi - signal feature model based on the Transformer - KAN model. Specifically, this embodiment constructs a bias matrix based on the dynamic time delay, and the attention weight of the Transformer - KAN multi - signal feature prediction model is:

[0071] B = -β·τ

[0072]

[0073] where: Attn(Q, K, V) is the attention weight; B is the bias matrix; β is a parameter controlling the influence degree of time delay on attention; τ is the dynamic delay time; Q, K, and V are Query, Key, and Value vectors respectively; d K is the dimension representing the key vector Key.

[0074] S3: Calculate the physical prediction values through the thermodynamic model of transformer oil temperature and the dynamic model of vibration respectively; use the Transformer - KAN model to extract data - driven prediction values, and adopt a gating mechanism to dynamically fuse the output results of the physical model branch and the data - driven branch to generate multi - signal prediction values.

[0075] In this embodiment, the thermodynamic model of transformer oil temperature is:

[0076]

[0077] Where: OT(t) is the oil temperature value corresponding to time t; AT(t) is the ambient temperature at time t; P(t) is the heat energy input at time t; H is the convective heat transfer coefficient; S is the area; C Oil is the equivalent heat capacity of the transformer oil.

[0078] In this embodiment, the dynamic model of the transformer vibration is:

[0079]

[0080] Where: v(t) is the vibration amplitude of the transformer box body; m is the equivalent mass; δ is the damping coefficient; k is the elastic coefficient.

[0081] Solve the differential equations of the thermodynamic model of the transformer oil temperature and the dynamic model of the transformer vibration to obtain the theoretical expected values y of the oil temperature and vibration physical quantities p .

[0082] In this embodiment, a gating mechanism is used to dynamically fuse the output results of the physical model branch and the data-driven branch, and the multi-signal prediction value obtained is:

[0083] y pre = g·y m +(1―g)y p

[0084] Where: y pre is the multi-signal prediction value; y m is the data-driven prediction value extracted by the Transformer-KAN multi-signal feature prediction model; y p is the physical prediction value calculated by the thermodynamic model of the transformer oil temperature and the dynamic model of the vibration; g is the gating factor.

[0085] Figure 3 Specifically, it shows the prediction results of the transformer oil temperature. It can be seen from the figure that the prediction curve and the actual curve basically coincide, and the change trends are consistent. The prediction model can effectively capture the changes of the characteristic parameters, and the prediction accuracy is high. The specific prediction index results of each characteristic quantity of the transformer are shown in Table 1.

[0086] Table 1 Prediction results of each characteristic quantity of the transformer

[0087]

[0088] S4: Dynamically adjust the weights of the reconstruction module according to the local fluctuation characteristics of the transformer multi-signals in different time periods, and reconstruct the multi-signal characteristic quantities by using the weighted reconstruction error.

[0089] According to the fluctuation characteristics of different signals of the transformer at different times, different weights are assigned to the reconstruction losses in the reconstruction module. High-precision reconstruction of data is carried out during stable periods, while data sensitivity is reduced during periods with large fluctuations, so that the reconstruction error of abnormal data deviating from the normal mode is relatively large. In this embodiment, the method steps for reconstructing multi-signal feature quantities using weighted reconstruction error are as follows:

[0090] S41: Use the Transformer encoder to extract temporal features and obtain the latent representation h of the multi-signal data features of the transformer; use the Transformer encoder to reconstruct the input sequence based on h to obtain the reconstruction result.

[0091] S42: For each time step, calculate the local standard deviation within a fixed window; calculate the corresponding weight value according to the weight function:

[0092]

[0093] where: weight(t) is the local weight value; σ t is the local standard deviation; α is the adjustment coefficient for balancing the sensitivity of the weight function;

[0094] S43: Use the weighted root mean square error to reconstruct the multi-signal feature quantities, so that the model achieves a high reconstruction accuracy during stable periods. The weighted root mean square error is:

[0095]

[0096] where: r ti is the actual value; r pi is the reconstructed value.

[0097] S5: Combine the dynamic alignment feature distance, the prediction error obtained based on data-driven, the prediction error obtained based on the physical model, and the multi-signal feature quantity reconstruction error, and calculate the anomaly confidence weighted to detect transformer anomalies.

[0098] In view of the extremely small number of abnormal samples of the transformer, positive and negative samples are constructed, and the contrast loss function is used to learn to distinguish positive and negative samples. Combining the prediction results, reconstruction results, and feature embeddings obtained by contrast learning, the anomaly confidence is calculated to detect anomalies. Specifically, in this embodiment, the calculation method of the anomaly confidence is as follows:

[0099] S51: Construct positive and negative samples; among them, the positive samples are taken from the sliding window data under the normal operation state of the transformer; the negative samples are samples that deviate from the physical law by artificially adding to the normal data;

[0100] S52: Input positive and negative samples into the Transformer-KAN multi-signal feature prediction model to extract sample features; design a contrast loss function to make the feature distances of positive samples closer and those of negative samples farther; the contrast loss function is:

[0101]

[0102] where: y i,j is a binary label; d i,j is the feature distance; m is a hyperparameter that controls the minimum spacing of negative samples.

[0103] S53: Weight the feature distance obtained from contrastive learning, the prediction error obtained from data-driven, the prediction error of the physical model, and the reconstruction error, and calculate the anomaly confidence as:

[0104] S = λ1D + λ2Y m + λ3Y t + λ4R

[0105] where: D is the feature space distance calculated from contrastive learning; Y m is the prediction error obtained from data-driven; Y t is the prediction error of the physical model; R is the reconstruction error; λ1, λ2, λ3, λ4 are the weights for adjusting each part.

[0106] The greater the average distance between the features of the input data and the features of normal samples, the greater the error between the predicted value or reconstructed value and the actual value, the higher the anomaly confidence, and the detection of anomalies.

[0107] This embodiment can realize the prediction and anomaly detection of transformer state data based on the delay relationship of multi-signals of the transformer. First, through the dynamic time alignment algorithm, combined with the temporal attention mechanism, different signals of the transformer are accurately aligned in the time dimension; the data-driven model is fused with the model based on physical laws, and the physical prior knowledge of the transformer is used to make up for the deficiencies of the data-driven model, improving the interpretability of the model; the data-driven prediction, reconstruction error, contrastive learning, and physical knowledge are fused to enhance the prediction, comprehensively and comprehensively evaluate the anomaly possibility of the transformer, providing new ideas for transformer anomaly detection, and effectively guaranteeing the safe operation of the transformer.

[0108] The above-described embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for collaborative modeling and anomaly detection of multiple signals of a transformer based on dynamic time warping, characterized in that: Including the following steps: S1: An adaptive sliding window is adopted to dynamically divide the window length according to the change rate of the transformer's electrical-thermal-vibration signals. By introducing a delay time constraint term estimated by a physical model to improve the dynamic time warping algorithm, the multi-signals of the transformer are dynamically aligned and the time delay is quantified. S2: Based on the dynamic time delay information obtained in step S1, a bias matrix is introduced in the attention calculation of the Transformer model to construct a Transformer-KAN multi-signal feature prediction model including delay features. S3: Calculate the physical prediction values through the thermodynamic model of the transformer oil temperature and the dynamic model of vibration respectively; use the Transformer-KAN model to extract data-driven prediction values, and adopt a gating mechanism to dynamically fuse the output results of the physical model branch and the data-driven branch to generate multi-signal prediction values. S4: Dynamically adjust the weights of the reconstruction module according to the local fluctuation characteristics of the transformer multi-signals at different time periods, and reconstruct the multi-signal feature quantity by using the weighted reconstruction error. S5: Combine the dynamic alignment feature distance, the prediction error obtained based on data-driven, the prediction error obtained based on the physical model, and the multi-signal feature quantity reconstruction error, and calculate the anomaly confidence weighted to detect transformer anomalies.

2. The method for collaborative modeling and anomaly detection of multiple signals of a transformer based on dynamic time warping according to claim 1, characterized in that: In step one, the dynamic adjustment method of the window length is as follows: Where: W is the window length; β is the adjustment coefficient; is the absolute value of the signal change rate; ε is used to prevent division by zero.

3. The method for collaborative modeling and anomaly detection of multiple signals of a transformer based on dynamic time warping according to claim 1, wherein: In step one, the dynamic time warping algorithm obtains a penalty term by introducing the delay time estimated by the physical model, constructs a cost function to constrain the deviation between the actual delay and the expected delay of the matching points, and the formula is: Δt r = t s × (j – i) P(i,j) = |Δt r ―Δt e | C(i,j) = |x i ― y j | + λ·P(i,j) Among them: C(i,j) is the cost function; P(i,j) is the penalty term; λ is the penalty coefficient; Δt r is the actual delay time; t s is the sampling interval time...; Δt e is the delay time between any two signals in the transformer electro-thermal-vibration signals estimated by the physical model; X = {x1, x2,... x N} and Y = {y1, y2,... y N} are the time series of two signals in the transformer electro-thermal-vibration signals respectively; N is the total number of samplings; i and j are the time points of the two signals respectively; Use the dynamic time warping algorithm to calculate the best matching points between any two signals in the transformer's electrical-thermal-vibration, convert the difference between the best matching points into the actual time difference, and quantify the dynamic time delay between the signals.

4. The method for collaborative modeling and anomaly detection of multiple transformer signals based on dynamic time warping according to claim 1, wherein: In step two, based on the dynamic time delay, a bias matrix is constructed to obtain the attention weights of the Transformer-KAN multi-signal feature prediction model: B = -β·τ Among them: Attn(Q, K, V) is the attention weight; B is the bias matrix; β is a parameter that controls the degree of influence of time delay on attention; τ is the dynamic delay time; Q, K, and V are Query, Key, and Value vectors respectively; d K represents the dimension of the key vector Key.

5. The method for collaborative modeling and anomaly detection of multiple signals of a transformer based on dynamic time warping according to claim 1, wherein: In step three, the thermodynamic model of the transformer oil temperature is: Where: OT(t) is the oil temperature value corresponding to time t; AT(t) is the ambient temperature at time t; P(t) is the heat energy input at time t; H is the convective heat transfer coefficient; S is the area; C Oil is the equivalent heat capacity of the transformer oil; The dynamic model of the transformer vibration is: Where: v(t) is the vibration amplitude of the transformer box; m is the equivalent mass; δ is the damping coefficient; k is the elastic coefficient.

6. The method for collaborative modeling and anomaly detection of multiple transformer signals based on dynamic time warping according to claim 1, wherein: In step three, the multi-signal prediction value is: y pre = g·y m +(1―g)y p where: y pre is the multi-signal prediction value; y m is the data-driven prediction value extracted by the Transformer-KAN multi-signal feature prediction model; y p is the physical prediction value calculated by the thermodynamic model of transformer oil temperature and the dynamic model of vibration; g is the gating factor.

7. The method for collaborative modeling and anomaly detection of multiple signals of a transformer based on dynamic time warping according to claim 1, characterized in that: In step four, the method steps of reconstructing the multi-signal feature quantity by using the weighted reconstruction error are: S41: Use the Transformer encoder to extract the temporal features to obtain the latent representation h of the transformer multi-signal data features; reconstruct the input sequence according to h by using the Transformer encoder to obtain the reconstruction result. S42: For each time step, calculate the local standard deviation within a fixed window; calculate the corresponding weight value according to the weight function: where: weight(t) is the local weight value; σ t is the local standard deviation; α is the adjustment coefficient for balancing the sensitivity of the weight function; S43: Reconstruct the multi-signal feature quantity by using the weighted root mean square error; the weighted root mean square error is: where: r ti is the actual value; r pi is the reconstructed value.

8. The method for collaborative modeling and anomaly detection of multiple transformer signals based on dynamic time warping according to claim 1, wherein: In step five, the calculation method of the anomaly confidence is: S51: Construct positive and negative samples; among them, the positive samples are taken from the sliding window data under the normal operation state of the transformer; the negative samples are samples that deviate from the physical laws artificially added to the normal data. S52: Input positive and negative samples into the Transformer-KAN multi-signal feature prediction model to extract sample features; design a contrastive loss function to make the feature distances of positive samples closer and those of negative samples farther apart; the contrastive loss function is as follows: where: y i,j is a binary label; d i,j is the feature distance; m is a hyperparameter that controls the minimum spacing of negative samples; S53: Weight the feature distances obtained from contrastive learning, the prediction errors obtained from data-driven methods, the prediction errors of the physical model, and the reconstruction errors, and calculate the anomaly confidence as: S = λ1D + λ2Y m + λ3Y t + λ4R Where: D is the feature space distance obtained by contrastive learning; Y m is the prediction error obtained based on data-driven; Y t is the prediction error of the physical model; R is the reconstruction error; λ1, λ2, λ3, λ4 are the weights for adjusting each part.

Images