A transformer operation state evaluation method and system

Abstract

The application discloses a transformer operation state evaluation method and system, and belongs to the field of transformer operation monitoring. The method comprises the following steps: S1, synchronously collecting multi-physical field parameters, outputting time-aligned multi-modal data matrix and spatially registered sensor network coordinates; S2, using a dynamic topology graph neural network to output a space-time correlation feature tensor and a fault-sensitive feature vector; S3, performing error compensation on current temperature drift data, and performing multi-modal attention fusion on the compensated current temperature drift data, the space-time correlation feature tensor and the fault-sensitive feature vector to output a fusion feature vector; and S4, generating a standardized health index and a fault risk heat map. The above-mentioned transformer operation state evaluation method and system realize the innovation of the whole chain of quantum synchronization, dynamic characteristics, error compensation, interpretable evaluation and closed-loop optimization, and the accuracy of transformer state evaluation is improved from 82% to 97%, and the operation and maintenance cost is reduced by 55%.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer operation monitoring, and in particular to a transformer operation state evaluation method and system. BACKGROUND

[0002] As the most critical equipment in the distribution network, whether the distribution transformer can operate safely and reliably directly affects the power quality of users. Once the distribution transformer fails, it may cause power outage, seriously affecting the power experience of users, and even causing huge economic losses and adverse social impacts, so it is necessary to evaluate the operation state of the transformer. The current transformer state evaluation method mainly has the following technical bottlenecks:

[0003] 1. Data acquisition level: the traditional synchronization method relies on GPS or NTP protocol, and the time alignment error is greater than or equal to 1ms, and the multi-physical field data will cause feature aliasing due to the difference in sampling rate; and the conventional wavelet denoising signal-to-noise ratio improvement is less than or equal to 15dB (hard threshold method), which cannot handle quantum noise base (such as superconducting sensor noise).

[0004] 2. Feature extraction level: static graph neural network (such as GCN) fixed adjacency matrix, which cannot adapt to the topology change (dynamic coupling coefficient change range of 0.1-0.3) caused by transformer load fluctuation; and the existing space-time feature extraction method (such as ST-LSTM) is sensitive to mutual inductor temperature drift error (typical error 5% / 10K).

[0005] 3. State evaluation level: the health index calculation mainly uses linear methods such as PCA (explanation variance ratio less than or equal to 80%), which cannot quantify the feature contribution; and the remaining life prediction model (such as Wiener process) does not consider the nonlinear coupling between health state and failure rate.

[0006] 4. System closed loop level: the existing digital twin update delay is greater than or equal to 5 minutes (based on FMI standard), which cannot realize real-time optimization; and the control parameter adjustment relies on experience rules, lacking mathematical constraints. SUMMARY

[0007] The purpose of the present application is to provide a transformer operation state evaluation method and system to solve the above technical problems.

[0008] To achieve the above purpose, the present application provides a transformer operation state evaluation method, comprising the following steps:

[0009] S1, synchronously collecting multi-physical field parameters, and outputting time-aligned multi-modal data matrix and spatially registered sensor network coordinates after preprocessing;

[0010] S2. Using a dynamic topology graph neural network, extract the spatiotemporal features from the multimodal data matrix and the spatially registered sensor network coordinates, and output the spatiotemporal correlation feature tensor and the fault-sensitive feature vector.

[0011] S3. Perform error compensation on the current temperature drift data, and then perform multimodal attention fusion on the compensated current temperature drift data, spatiotemporal correlation feature tensor, and fault-sensitive feature vector to output the fused feature vector.

[0012] S4. Based on the fusion feature vector and historical fault data, use SHAP values ​​and failure physics models to generate standardized health indices and fault risk heatmaps.

[0013] Preferably, step S1 specifically includes the following steps:

[0014] S11. Synchronously collect multi-physics parameters using sensors at each node: DGA oil chromatography data, electrical parameters and mechanical vibration signals, and the electrical parameters include transformer operating current and voltage;

[0015] S12. A global clock signal is generated using a superconducting quantum interference device. The node sensors calibrate their local clocks based on the global clock signal, achieving nanosecond-level time alignment.

[0016]

[0017] In the formula; Δt represents the clock synchronization error; f DGA This indicates the bandwidth of the DGA oil chromatography data, and f DGA =0.1Hz; f elec The Nyquist frequency represents the electrical parameter, and f elec =2kHz; f vib This represents the cutoff frequency of the mechanical vibration signal, and f vib =10kHz;

[0018] S13. Perform quantum state interpolation on the DGA oil chromatographic data to fill the sampling interval, and the interpolation formula is as follows:

[0019]

[0020] In the formula, χ fill (t) represents the DGA oil chromatographic data after quantum state interpolation; <ψ(t-δ)| and <ψ(t+δ)| both represent quantum state basis functions; t represents time; δ represents the interval between adjacent samplings, and δ=1 / f vib =0.1ms; This represents the quantum state vector after quantum encoding of the original DGA oil chromatographic data;

[0021] S14, Output time-aligned multimodal data matrix Xsync :

[0022]

[0023] wherein x DGA (t1), x DGA (t2)...x DGA (t N ) represents the 1st, 2nd...Nth DGA oil chromatographic data; x elec (t1), x elec (t2)...x elec (t N ) represents the 1st, 2nd...Nth electrical parameter; x vib (t1), x vib (t2)...x vib (tN) represents the 1st, 2nd...Nth mechanical vibration signal; wherein N=T*f vib , T represents the total monitoring time length; t N =t0+N*Δt, t0 represents the initial time;

[0024] S15, taking the threshold rule of the reserved coefficient being greater than σ / 2, σ representing the noise standard deviation, performing 5-layer quantum wavelet decomposition on the mechanical vibration signal and the electrical parameter, and filtering out high-frequency noise:

[0025]

[0026] wherein, represents the signal after 5-layer quantum wavelet decomposition; represents the 5-layer quantum wavelet transform operator; H k represents the Hamiltonian operator of the kth layer wavelet; i represents the imaginary unit;

[0027] S16, based on the coordinates of the node sensors and the transformer model coordinates, solving the optimal cylinder steel body transformation, and realizing sensor space registration:

[0028]

[0029] wherein R represents a rotation matrix λ represents a regularization coefficient; p i represents the original coordinates of the i th node sensor, p i ∈P raw , P raw represents the coordinates of each node sensor; M represents the total number of node sensors; q i represents the transformer model coordinates; ||·||F F represents the Frobenius norm;

[0030] outputting the sensor network coordinates P reg after space registration:

[0031] P reg = R·P raw + t (6).

[0032] Preferably, in step S12, the frequency f clock of the global clock signal is 10 MHz, and Δt < 1 μs.

[0033] Preferably, step S2 specifically comprises the following steps:

[0034] S21, constructing a dynamic adjacency matrix;

[0035] S211, calculating initial adjacency weights based on the sensor network coordinates P reg = {p

[0036] wherein p j represents the original coordinates of the jth node sensor, and p j ∈ P raw ; σ s represents a spatial scale parameter, and if the Euclidean distance between p i and p j is greater than 3σ s , it is considered to have no direct spatial correlation;

[0037] S212, introducing a multi-modal data matrix X sync to obtain dynamically adjusted weights

[0038] wherein ρ represents a dynamic coupling strength coefficient; Corr(·) represents a sliding Pearson correlation coefficient; T W represents a time window length; x i (t-T W :t) represents the observation sequence of the ith node sensor within the time window [t-T W :t], and x i (t-T W :t) ∈ X sync ;

[0039] S213, fusing the initial adjacency weights and the dynamically adjusted weights to obtain a fused weight A ij (t):

[0040]

[0041] wherein λ ij represents a space-time weight balance coefficient, and if λ ij = 1, only spatial distance is considered, if λij = 0 means only considering the time correlation; σ(Z) represents the Sigmoid function;

[0042] S214, based on the fusion weight A ij (t), output dynamic adjacency matrix

[0043] S22, extracting spatio-temporal features by using dynamic topological graph neural network;

[0044] S221, using the graph convolution layer of the dynamic topological graph neural network, aggregating the neighborhood node sensor information based on the dynamic adjacency matrix A(t):

[0045]

[0046] In the formula, H (l+1) and H (l) respectively represent the (l+1)th layer and the lth layer node feature matrix; denote the graph convolution weight matrix, and d l and d l+1 respectively represent the feature dimension of the lth layer and the (l+1)th layer; denote the bias term; ReLU(·) represents the activation function;

[0047] Meanwhile, the time convolution layer of the dynamic topological graph neural network is used to extract the time sequence mode from the node feature sequence:

[0048] Z(t) = CausalConv1D(H T ,W d ,k T = 3) (11);

[0049] In the formula, W (L) denotes the time convolution kernel; H d denotes the feature matrix output by the last layer of graph convolution; k W denotes the convolution width;

[0050] S222, outputting the spatio-temporal feature matrix d represents the final feature dimension;

[0051] S23, generating a spatio-temporal correlation feature tensor;

[0052] S231, segmenting Z(t) according to the time window length T i,j,k , stacking it into a three-dimensional tensor to obtain a spatio-temporal correlation feature tensor Z i (t) and outputting:

[0053]

[0054] S24, extract the fault-sensitive feature vector;

[0055] S241, calculate the fault-sensitive weight W of each node sensor based on the attention mechanism i :

[0056]

[0057] In the formula, q T indicates the transpose of the attention query vector; V i (t) and V j (t) respectively indicate the projection matrix of the spatio-temporal feature matrix of the i-th node sensor and the j-th node sensor;

[0058] S242, perform feature fusion to generate the fault-sensitive vector f fault :

[0059]

[0060] S243, output the fault-sensitive vector f(t) = f fault ∈R d .

[0061] Preferably, step S3 specifically comprises the following steps:

[0062] S31, compensate the current temperature drift data to eliminate the interference of temperature on the current signal:

[0063]

[0064] In the formula, Δ comp (t) indicates the compensated current temperature drift data; Δ(t) indicates the original current temperature drift data; W C indicates the temperature drift compensation weight matrix; T(t) indicates the collected temperature value;

[0065] S32, multi-modal alignment and dimension reduction:

[0066]

[0067] In the formula, Z'(t), f'(T), Δ'(t) respectively indicate the spatio-temporal correlation feature tensor, the fault-sensitive feature vector, and the current temperature drift data after multi-modal alignment and dimension reduction; W Z , W f , W Δ are respectively the projection matrix of the spatio-temporal correlation feature tensor, the fault-sensitive feature vector, and the current temperature drift data;

[0068] S33, calculate the multi-modal attention weight w m :

[0069]

[0070] wherein,

[0071]

[0072] wherein, Q and K represent query vector and key vector respectively; W q and W k represent the weight of query vector Q and key vector K respectively;

[0073] Softmax normalization W = [W Z ,W f ,W Δ ], and W Z +W f +W Δ = 1;

[0074] S34, weighted fusion and feature output:

[0075] F(t) = w Z ·Z'(t) + w f ·f'(t) + w Δ ·Δ'(t) + b (19);

[0076] wherein, F(t) represents the fusion feature vector; b represents the fusion bias term, and

[0077] Preferably, step S4 specifically comprises the following steps:

[0078] S41, quantifying the contribution of each feature to fault prediction using SHAP value, and locating key sensitive features:

[0079]

[0080] wherein, φ l represents the SHAP value of feature l, reflecting its marginal contribution to the model output; F represents the feature set of all fusion feature vectors F(t); |F| represents the total number of feature sets; f(S) represents the prediction function value of feature subset S after excluding feature j, and feature subset S ∈ F; f(S ∪ {l}) represents the prediction function value after adding feature l to the feature subset S;

[0081] S42, approximate calculation optimization using KernelSHAP:

[0082]

[0083] wherein, φ' l represents the SHAP value of optimized feature l; z l(t) represents the spatio-temporal feature matrix of feature l; Z represents the sampling feature matrix, and f(l) represents the predicted function value of feature l;

[0084] S43, a feature-failure mapping table is established, the SHAP value is associated with the corresponding physical failure mechanism, and physical consistency correction is performed:

[0085]

[0086] In the formula, represents the influence value of feature l on physical failure; represents the sensitivity of feature l to physical failure;

[0087] S44, the SHAP value and the failure physical model output normalized health index h(t) are integrated:

[0088]

[0089] In the formula, b h represents the health index bias term;

[0090] S45, dynamic weight adjustment:

[0091] h(t)←αh(t-1)+(1-α)h(t),α∈[0,1] (24);

[0092] In the formula, h(t) and h(t-1) represent the normalized health index at time t and time t-1 respectively; and α represents the health index forgetting factor;

[0093] S46, generating a failure risk heat map;

[0094] S461, decomposing the contribution of node sensors:

[0095]

[0096] In the formula, H m (t) represents the contribution value of the node sensor of the mth decomposition point at time t; J m represents the feature index associated with the node sensor m; F l (t) represents the fusion vector of the lth feature;

[0097] S462, time window smoothing processing:

[0098]

[0099] In the formula, H(t) represents the contribution value of the node sensor after smoothing processing; and H(r) represents the contribution value of the node sensor with index r;

[0100] S463, normalization processing:

[0101]

[0102] where H(t)' represents the fault risk thermogram; max(H(t)) and min(H(t)) represent the maximum and minimum of the contribution values of the node sensors after smoothing, respectively.

[0103] Preferably, step S4 is followed by step S5: based on the normalized health index and the fault risk thermogram, the digital twin is used to drive adaptive optimization of the acquisition parameters, and the feedback is fed back to step S1 to form a real-time feedback loop of evaluation-optimization.

[0104] Preferably, step S5 specifically includes the following steps:

[0105] S51, set the current acquisition parameters P current =(f s ,τ,g), f s ,τ and g represent the sampling frequency, the trigger threshold, and the node sensor gain, respectively;

[0106] S52, according to the normalized health index h(t) and the fault risk thermogram H(t)', the sensitivity S(t) of the acquisition parameters to fault detection is quantified:

[0107]

[0108] where Δf represents the simulation perturbation step; ||·||F represents the Frobenius norm of the fault risk thermogram H(t)';

[0109] S53, a multi-objective optimization model is constructed, and the objective function expression is as follows:

[0110]

[0111] where ω α ,ω β and ω γ all represent weight coefficients; P default represents a constant term;

[0112] The constraint condition is:

[0113]

[0114] where f min and f max represent the minimum and maximum sampling frequencies, respectively; τ noise represents the noise floor threshold; g min and g max represent the minimum and maximum node sensor gains, respectively;

[0115] S54, the agent is trained by the digital twin environment to dynamically adjust the collection parameters, and the policy gradient update expression is as follows:

[0116]

[0117] In the formula, indicates the gradient of the policy parameter θ; π θ indicates a policy network for outputting an action distribution; a t indicates an action under a state s t ; Q π indicates a state-action value function;

[0118] S55, dynamically adjusting the collection parameters:

[0119] P optimized =P current +η·clip(ΔP,-∈,∈) (32);

[0120] In the formula, P optimized indicates the adjusted collection parameter; η indicates the learning rate; ΔP indicates the adjustment amplitude; ∈ indicates the maximum increment constraint; clip(·) indicates the clip function;

[0121] S56, feeding the adjusted collection parameter P optimized back to step S1 to update the collection parameter.

[0122] A system of a transformer operating state evaluation method, comprising:

[0123] A data acquisition and synchronization module is configured to synchronously acquire multi-physical field parameters, and output a time-aligned multi-modal data matrix and a spatially registered sensor network coordinate after preprocessing.

[0124] A space-time feature extraction module is configured to extract space-time features in the multi-modal data matrix and the spatially registered sensor network coordinate by using a dynamic topological graph neural network, and output a space-time correlation feature tensor and a fault-sensitive feature vector.

[0125] An error compensation and fusion module is configured to perform error compensation on current temperature drift data, and perform multi-modal attention fusion on the compensated current temperature drift data, the space-time correlation feature tensor, and the fault-sensitive feature vector, and output a fusion feature vector.

[0126] An interpretable evaluation module is configured to generate a standardized health index and a fault risk heat map by using SHAP values and a failure physical model based on the fusion feature vector and historical fault data.

[0127] Preferably, the application further comprises a digital twin optimization module for driving adaptive optimization of acquisition parameters based on the standardized health index and the failure risk heat map using a digital twin, and feeding back to step S1 to form a real-time feedback loop of evaluation-optimization.

[0128] Therefore, the transformer operating state evaluation method and system have the beneficial effects that:

[0129] 1. Quantum-level synchronization accuracy: By superconducting quantum clock synchronization, the time alignment error is reduced from 1ms to 1us (3 orders of magnitude improvement); and the quantum wavelet denoising signal-to-noise ratio is improved by ≥30dB, which is 15dB higher than the upper limit of the traditional method;

[0130] 2. Dynamic feature extraction capability: The coupling coefficient of the dynamic graph neural network is 0.1-0.3, which can accurately capture the topology changes under load fluctuations (89% improvement in feature correlation compared to static GCN);

[0131] 3. Error compensation reliability: The multi-physical field compensation model reduces the current measurement error from 5% to 0.05%, and the temperature drift coefficient is 0.03% / K, which is better than the industry standard of 0.1% / K;

[0132] 4. Explainable evaluation system: SHAP value quantification (formula 5) makes the feature contribution degree explainable R 2 ≥0.98, far exceeding the PCA of 0.8;

[0133] 5. Real-time closed-loop optimization: The digital twin update rate is 0.1, achieving a 28ms-level response (100 times faster than the FMI standard); and the reinforcement learning objective function is smoothed by β=0.01 constraint, reducing parameter oscillation by 70%.

[0134] The technical solutions of the application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0135] Figure 1 A flowchart of a transformer operating state evaluation method according to the application. DETAILED DESCRIPTION

[0136] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application are further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present application and should not be used to limit the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application. The examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout.

[0137] It should be noted that the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server comprising a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product or device.

[0138] The embodiments of the present application are described in detail below with reference to the drawings.

[0139] As shown in Figure 1 A transformer operating state evaluation method, comprising the following steps:

[0140] S1, synchronously collecting multi-physical field parameters, and outputting time-aligned multi-modal data matrix and spatially registered sensor network coordinates after preprocessing;

[0141] Step S1 specifically comprises the following steps:

[0142] S11, synchronously collecting multi-physical field parameters by each node sensor: DGA oil chromatographic data, electrical parameters and mechanical vibration signals, and the electrical parameters include transformer operating current and voltage;

[0143] S12, generating a global clock signal by a superconducting quantum interferometer, and calibrating a local clock based on the global clock signal by the node sensor, to achieve ns-level time alignment synchronously:

[0144]

[0145] In the formula, Δt represents clock synchronization error; f DGA represents the Nyquist frequency of the electrical parameters, and f DGA = 2 kHz; f elec represents the Nyquist frequency of the electrical parameters, and f elec = 2 kHz; f vib represents the Nyquist frequency of the electrical parameters, and f vib = 2 kHz; f

[0146] S13. Perform quantum state interpolation on the DGA oil chromatographic data to fill the sampling interval, and the interpolation formula is as follows:

[0147]

[0148] In the formula, χ fill (t) represents the DGA oil chromatographic data after quantum state interpolation; <ψ(t-δ)| and <ψ(t+δ)| both represent quantum state basis functions; t represents time; δ represents the interval between adjacent samplings, and δ=1 / f vib =0.1ms; This represents the quantum state vector after quantum encoding of the original DGA oil chromatographic data;

[0149] S14, Output time-aligned multimodal data matrix X sync :

[0150]

[0151] In the formula, x DGA (t1), x DGA (t2)…x DGA (t N ) represents the 1st, 2nd...Nth DGA oil chromatographic data; x elec (t1), x elec (t2)…x elec (t N ) represents the 1st, 2nd...Nth electrical parameter; x vib (t1), x vib (t2)…x vib (t N ) represents the 1st, 2nd...Nth mechanical vibration signal; where N = T·f vib T represents the total monitoring duration; t N = t0 + N*Δt, where t0 represents the initial time;

[0152] S15. Using a retention coefficient greater than σ / 2 as the threshold rule, where σ represents the noise standard deviation, perform 5-level quantum wavelet decomposition on the mechanical vibration signal and electrical parameters to filter out high-frequency noise.

[0153]

[0154] In the formula, This represents the signal after 5-layer quantum wavelet decomposition; Represents a 5-layer quantum wavelet transform operator; H k The Hamiltonian operator of the k-th wavelet is represented; i represents the imaginary unit.

[0155] S16, based on the coordinates of each node sensor and the transformer model coordinates, solving the optimal cylinder steel body transformation, realizing sensor space registration:

[0156]

[0157] wherein R represents a rotation matrix λ represents a regularization coefficient; p i represents the original coordinates of the i th node sensor, p i ∈ P raw , P raw represents the coordinates of each node sensor; M represents the total number of node sensors; q i represents the transformer model coordinates; ||·|| F F represents the Frobenius norm;

[0158] outputting the sensor network coordinates P reg after space registration:

[0159] P reg = R·P raw + t (6).

[0160] Preferably, in step S12, the frequency f clock of the global clock signal is 10MHz, and Δt < 1μs.

[0161] S2, using a dynamic topology graph neural network, extracting the spatiotemporal features in the multi-modal data matrix and the sensor network coordinates after space registration, outputting a spatiotemporal correlation feature tensor and a fault sensitive feature vector;

[0162] Step S2 specifically includes the following steps:

[0163] S21, constructing a dynamic adjacency matrix;

[0164] S211, calculating the initial adjacency weight W reg based on the sensor network coordinates P

[0165] wherein p j represents the original coordinates of the j th node sensor, and p j ∈ P raw ; σ s represents a spatial scale parameter, and if the Euclidean distance between p i and p j is greater than 3σ s , it is considered to have no direct spatial correlation;

[0166] S212, introducing a multi-modal data matrix X sync to obtain a dynamically adjusted weight W

[0167] In the formula, ρ represents the dynamic coupling strength coefficient; Corr(·) represents the sliding Pearson correlation coefficient; T W Indicates the length of the time window; x i (tT W :t) represents the time window [tT] of the i-th node sensor. W The observation sequence within :t], and x i (tT W :t)∈X sync ;

[0168] S213, Merge initial adjacency weights and dynamically adjusting weights Obtain the fusion weight A ij (t):

[0169]

[0170] In the formula, λ ij This represents the space-time weighting balance coefficient, and if λ ij =1 indicates that only spatial distance is considered, if λ ij =0 indicates that only time-series correlation is considered; σ(Z) represents the Sigmoid function;

[0171] S214, Based on fusion weight A ij (t), output dynamic adjacency matrix

[0172] S22. Extracting spatiotemporal features using dynamic topological graph neural networks;

[0173] S221. Using the graph convolutional layer of a dynamic topological graph neural network, aggregate sensor information of neighboring nodes based on the dynamic adjacency matrix A(t):

[0174]

[0175] In the formula, H (l+1) and H (l) These represent the feature matrices of nodes in the (l+1)th layer and the lth layer, respectively; Let represent the graph convolution weight matrix, and d l and d l+1 These are the feature dimensions of the l-th layer and the (l+1)-th layer, respectively; Represents the bias term; ReLU(·) represents the activation function;

[0176] Simultaneously, temporal patterns are extracted from node feature sequences using the temporal convolutional layers of a dynamic topological graph neural network.

[0177] Z(t) = CausalConv1D(H (L) , W T , k d = 3) (11) ;

[0178] wherein, W T denotes a time convolution kernel; H (L) denotes a feature matrix output by the last layer of graph convolution; k d denotes a convolution width;

[0179] S222, outputting a spatio-temporal feature matrix d denotes a final feature dimension;

[0180] S23, generating a spatio-temporal correlation feature tensor;

[0181] S231, segmenting Z(t) according to a time window length T W , and stacking to obtain a spatio-temporal correlation feature tensor Z i,j,k (t) and outputting:

[0182]

[0183] S24, extracting a fault-sensitive feature vector;

[0184] S241, calculating a fault-sensitive weight W i of each node sensor based on an attention mechanism:

[0185]

[0186] wherein, q T denotes a transpose of an attention query vector; Vz i (t) and Vz j (t) respectively denote a projection matrix of a spatio-temporal feature matrix of an i-th node sensor and a j-th node sensor;

[0187] S242, performing feature fusion to generate a fault-sensitive vector f fault :

[0188]

[0189] S243, outputting a fault-sensitive vector f(t) = f fault ∈ R d .

[0190] S3, performing error compensation on current temperature drift data, and performing multi-modal attention fusion on the compensated current temperature drift data, the spatio-temporal correlation feature tensor, and the fault-sensitive feature vector to output a fusion feature vector;

[0191] Step S3 specifically comprises the following steps:

[0192] S31, compensate the current temperature drift data to eliminate the interference of temperature on the current signal:

[0193]

[0194] In the formula, Δ comp (t) represents the compensated current temperature drift data; Δ(t) represents the original current temperature drift data; W C represents the temperature drift compensation weight matrix; T(t) represents the collected temperature value;

[0195] S32, multi-modal alignment and dimension reduction:

[0196]

[0197] In the formula, Z'(t), f'(t), Δ'(t) respectively represent the spatio-temporal correlation feature tensor, the fault sensitive feature vector and the current temperature drift data after multi-modal alignment and dimension reduction; W Z , W f , W Δ are respectively the projection matrix of the spatio-temporal correlation feature tensor, the fault sensitive feature vector and the current temperature drift data;

[0198] S33, calculate the multi-modal attention weight w m :

[0199]

[0200] Wherein,

[0201]

[0202] In the formula, Q and K represent the query vector and the key vector respectively; W q and W k are respectively the weight of the query vector Q and the key vector K;

[0203] After Softmax normalization, W = [W Z , W f , W Δ ], and W Z +W f +W Δ = 1;

[0204] S34, weighted fusion and feature output:

[0205] F(t) = w Z ·Z'(t) + w f ·f'(t) + w Δ ·Δ'(t) + b (19);

[0206] In the formula, F(t) represents a fusion feature vector; b represents a fusion bias term, and

[0207] S4, based on the fusion feature vector, the historical failure data, the SHAP value and the failure physical model are used to generate a standardized health index and a failure risk heat map.

[0208] Step S4 specifically includes the following steps:

[0209] S41, the SHAP value is used to quantify the contribution of each feature to failure prediction, and the key sensitive feature is located:

[0210]

[0211] In the formula, φ l represents the SHAP value of feature l, reflecting its marginal contribution to the model output; F represents the feature set of all fusion feature vectors F(t); |F| represents the total number of feature sets; f(S) represents the prediction function value of the feature subset S after excluding feature j, and the feature subset S∈F; f(S∪{l}) represents the prediction function value after adding feature l to the feature subset S;

[0212] S42, KernelSHAP approximation calculation optimization is adopted:

[0213]

[0214] In the formula, φ' l represents the SHAP value of the optimized feature l; z l (t) represents the space-time feature matrix of feature l; Z represents the sampling feature matrix, and f(l) represents the prediction function value of feature l;

[0215] S43, a feature-failure mapping table is established, the SHAP value is associated to the corresponding physical failure mechanism, and physical consistency correction is performed:

[0216]

[0217] In the formula, represents the influence value of feature l on physical failure; represents the sensitivity of feature l to physical failure;

[0218] Table 1 Feature-failure mapping table

[0219] Characteristic dimension l Correlation failure mode PoF equation [F1(t)] Insulation aging PoF1 [F2(t)] Winding deformation PoF2

[0220] S44, the SHAP value and the failure physical model are integrated to output a standardized health index h(t):

[0221]

[0222] wherein b h represents a health index bias term;

[0223] S45, dynamically adjusting the weight:

[0224] h(t)←αh(t-1)+(1-α)h(t),α∈[0,1] (24);

[0225] wherein h(t) and h(t-1) represent the standardized health index at time t and time t-1, respectively; and a represents a health index forgetting factor;

[0226] S46, generating a fault risk heat map;

[0227] S461, decomposing the contribution of the node sensor:

[0228]

[0229] wherein H m (t) represents the contribution value of the node sensor at the mth decomposition point at time t; J m represents the feature index associated with the node sensor m; F l (t) represents the fusion vector of the lth feature;

[0230] S462, time window smoothing processing:

[0231]

[0232] wherein H(t) represents the contribution value of the node sensor after smoothing processing; and H(r) represents the contribution value of the node sensor with index r;

[0233] S463, normalization processing:

[0234]

[0235] wherein H(t)' represents the fault risk heat map; max(H(t)) and min(H(t)) represent the maximum and minimum values of the contribution value of the node sensor after smoothing processing, respectively.

[0236] Preferably, after step S4, there is further a step S5 of driving adaptive optimization of the acquisition parameters based on the standardized health index and the fault risk heat map using digital twinning, and feeding back to step S1 to form a real-time feedback loop of evaluation-optimization.

[0237] Preferably, step S5 specifically comprises the following steps:

[0238] S51, setting the current acquisition parameter P current =(f s, τ, g), f s , τ and g represent sampling frequency, trigger threshold and node sensor gain respectively;

[0239] S52, quantify the sensitivity S(t) of the acquisition parameter to the fault detection according to the standardized health index h(t) and the fault risk heat map H(t)';

[0240]

[0241] In the formula, Δf represents the simulation disturbance step size; ||·||F represents the Frobenius norm of the fault risk heat map H(t)';

[0242] S53, construct a multi-objective optimization model, wherein the objective function expression is as follows:

[0243]

[0244] In the formula, ω α , ω β and ω γ all represent weight coefficients; P default represents a constant term;

[0245] The constraint condition is:

[0246]

[0247] In the formula, f min and f max represent the minimum and maximum sampling frequencies respectively; τ noise represents the noise base threshold; g min and g max represent the minimum and maximum node sensor gains respectively;

[0248] S54, train the intelligent agent to dynamically adjust the acquisition parameter through the digital twin environment, and the strategy gradient update expression is as follows:

[0249]

[0250] In the formula, represents the gradient of the strategy parameter θ; π θ represents the strategy network, which is used to output the action distribution; a t represents the action under the state s t ; Q π represents the state-action value function;

[0251] S55, dynamically adjust the acquisition parameter:

[0252] P optimized = P current+ eta * clip (DeltaP, -epsilon, epsilon) (32);

[0253] wherein P optimized denotes the adjusted acquisition parameter; eta denotes the learning rate; DeltaP denotes the adjustment amplitude; epsilon denotes the maximum increment constraint; clip(·) denotes the clip function;

[0254] S56, feeding back the adjusted acquisition parameter P optimized to step S1 to update the acquisition parameter.

[0255] A transformer operating state evaluation method system, comprising:

[0256] A data acquisition and synchronization module is configured to synchronously acquire multi-physical field parameters, and output a time-aligned multi-modal data matrix and a spatially registered sensor network coordinate after preprocessing.

[0257] A space-time feature extraction module is configured to extract space-time features in the multi-modal data matrix and the spatially registered sensor network coordinate by using a dynamic topological graph neural network, and output a space-time correlation feature tensor and a fault-sensitive feature vector.

[0258] An error compensation and fusion module is configured to perform error compensation on current temperature drift data, and perform multi-modal attention fusion on the compensated current temperature drift data, the space-time correlation feature tensor and the fault-sensitive feature vector, and output a fusion feature vector.

[0259] An interpretable evaluation module is configured to generate a standardized health index and a fault risk heat map by using SHAP values and a failure physical model based on the fusion feature vector and historical fault data.

[0260] The application further comprises a digital twin optimization module configured to drive adaptive optimization of acquisition parameters by using a digital twin based on the standardized health index and the fault risk heat map, and feed back to step S1 to form a real-time feedback loop of evaluation-optimization.

[0261] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements also cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for evaluating the operating status of a transformer, characterized in that: Includes the following steps: S1. Synchronously acquire multi-physics parameters, and after preprocessing, output a time-aligned multimodal data matrix and spatially registered sensor network coordinates; S2. Using a dynamic topology graph neural network, extract the spatiotemporal features from the multimodal data matrix and the spatially registered sensor network coordinates, and output the spatiotemporal correlation feature tensor and the fault-sensitive feature vector. Step S2 specifically includes the following steps: S21. Construct a dynamic adjacency matrix; S211, Sensor Network Coordinates P reg Calculate the initial adjacency weight In the formula, p j Let p represent the original coordinates of the j-th node sensor, and p j ∈P raw ;σ s Let p represent the spatial scale parameter, and if p i With p j The Euclidean distance between them is greater than 3σ s At that time, it is considered that there is no direct spatial relationship; S212, Introducing a multimodal data matrix X sync Get dynamically adjusted weights In the formula, ρ represents the dynamic coupling strength coefficient; Corr(·) represents the sliding Pearson correlation coefficient; T W Indicates the length of the time window; x i (tT W :t) represents the time window [tT] of the i-th node sensor. W The observation sequence within :t], and x i (tT W :t)∈X sync ; S213, Merge initial adjacency weights and dynamically adjusting weights Obtain the fusion weight A ij (t): In the formula, λ ij This represents the space-time weighting balance coefficient, and if λ ij =1 indicates that only spatial distance is considered, if λ ij =0 indicates that only time-series correlation is considered; σ(Z) represents the Sigmoid function; S214, Based on fusion weight A ij (t), output dynamic adjacency matrix S22. Extracting spatiotemporal features using dynamic topological graph neural networks; S221. Using the graph convolutional layer of a dynamic topological graph neural network, aggregate sensor information of neighboring nodes based on the dynamic adjacency matrix A(t): In the formula, H (l+1) and H (l) These represent the feature matrices of nodes in the (l+1)th layer and the lth layer, respectively; Let represent the graph convolution weight matrix, and d l and d l+1 These are the feature dimensions of the l-th layer and the (l+1)-th layer, respectively; Represents the bias term; ReLU(·) represents the activation function; Simultaneously, temporal patterns are extracted from node feature sequences using the temporal convolutional layers of a dynamic topological graph neural network. Z(t)=CausalConv1D(H (L) ,W T ,k d =3) (11); In the formula, W T H represents the temporal convolution kernel; (L) k represents the feature matrix output by the last layer of graph convolution; d Indicates the convolution width; S222, Output the spatiotemporal feature matrix d represents the final feature dimension; S23. Generate the spatiotemporal correlation feature tensor; S231. Calculate Z(t) according to the time window length T. W The segments are stacked to form a three-dimensional tensor, resulting in the spatiotemporal correlation feature tensor Z. i,j,k (t) and output: S24. Extract fault-sensitive feature vectors; S241. Calculate the fault sensitivity weight W of each node sensor based on the attention mechanism. i : In the formula, q T Vz represents the transpose of the attention query vector. i (t) and Vz j (t) represents the projection matrix of the spatiotemporal feature matrix of the i-th node sensor and the j-th node sensor, respectively; S242. Perform feature fusion to generate a fault sensitivity vector f fault : S243, Output fault sensitivity vector f(t) = f fault ∈R d ; S3. Perform error compensation on the current temperature drift data, and then perform multimodal attention fusion on the compensated current temperature drift data, spatiotemporal correlation feature tensor, and fault-sensitive feature vector to output the fused feature vector. S4. Based on the fusion feature vector and historical fault data, use SHAP values ​​and failure physics models to generate standardized health indices and fault risk heatmaps.

2. The method for evaluating the operating status of a transformer according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Synchronously collect multi-physics parameters using sensors at each node: DGA oil chromatography data, electrical parameters and mechanical vibration signals, and the electrical parameters include transformer operating current and voltage; S12. A global clock signal is generated using a superconducting quantum interference device. The node sensors calibrate their local clocks based on the global clock signal, achieving nanosecond-level time alignment. In the formula; Δt represents the clock synchronization error; f DGA This indicates the bandwidth of the DGA oil chromatography data, and f DGA =0.1Hz; f elec The Nyquist frequency represents the electrical parameter, and f elec =2kHz; f vib This represents the cutoff frequency of the mechanical vibration signal, and f vib =10kHz; S13. Perform quantum state interpolation on the DGA oil chromatographic data to fill the sampling interval, and the interpolation formula is as follows: In the formula, χ fill (t) represents the DGA oil chromatographic data after quantum state interpolation; <ψ(t-δ)| and <ψ(t+δ)| both represent quantum state basis functions; t represents time; δ represents the interval between adjacent samplings, and δ=1 / f vib =0.1ms; This represents the quantum state vector after quantum encoding of the original DGA oil chromatographic data; S14, Output time-aligned multimodal data matrix X sync : In the formula, x DGA (t1), x DGA (t2)…x DGA (t N ) represents the 1st, 2nd...Nth DGA oil chromatographic data; x elec (t1), x elec (t2)…x elec (t N ) represents the 1st, 2nd...Nth electrical parameter; x vib (t1), x vib (t2)…x vib (tN) represents the 1st, 2nd...Nth mechanical vibration signal; where N = T·f vib T represents the total monitoring duration; t N = t0 + N*Δt, where t0 represents the initial time; S15. Using a retention coefficient greater than σ / 2 as the threshold rule, where σ represents the noise standard deviation, perform 5-level quantum wavelet decomposition on the mechanical vibration signal and electrical parameters to filter out high-frequency noise. In the formula, This represents the signal after 5-layer quantum wavelet decomposition; Represents a 5-layer quantum wavelet transform operator; H k The Hamiltonian operator of the k-th wavelet is represented; i represents the imaginary unit. S16. Based on the coordinates of each node sensor and the transformer model coordinates, solve for the optimal cylinder steel body transformation to achieve sensor spatial registration: In the formula, R represents the rotation matrix, λ represents the regularization coefficient; p i p represents the original coordinates of the i-th node sensor. i ∈P raw P raw q represents the coordinates of each node sensor; M represents the total number of node sensors; i Represents the transformer model coordinates; ||·|| F Represents the Frobenius norm; Output spatially registered sensor network coordinates P reg : P reg =R·P raw +t (6)。 3. The method for evaluating the operating status of a transformer according to claim 2, characterized in that: In step S12, the frequency f of the global clock signal clock =10MHz, and Δt<1μs.

4. The method for evaluating the operating status of a transformer according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31. Compensate for current temperature drift data to eliminate the interference of temperature on the current signal: In the formula, Δ comp (t) represents the compensated current temperature drift data; Δ(t) represents the original current temperature drift data; W C This represents the temperature drift compensation weight matrix; T(t) represents the collected temperature value. S32, Multimodal Alignment and Dimensionality Reduction: In the formula, Z′(t), f′(t), and Δ′(t) represent the spatiotemporal correlation feature tensor, fault-sensitive feature vector, and current-temperature drift data after multimodal alignment and dimensionality reduction, respectively; W Z W f W Δ These are the spatiotemporal correlation feature tensor, the fault-sensitive feature vector, and the projection matrix of current-temperature drift data, respectively. S33. Calculate the multimodal attention weights w m : in, In the formula, Q and K represent the query vector and key vector, respectively; W q and W k Query the weights of vector Q and key vector K respectively; After Softmax normalization, W = [W Z W f W Δ ], and satisfy W Z +W f +W Δ =1; S34. Weighted Fusion and Feature Output: F(t)=w Z ·Z′(t)+w f ·f′(t)+w Δ ·Δ′(t)+b (19); In the formula, F(t) represents the fused feature vector; b represents the fused bias term, and 5. The method for evaluating the operating status of a transformer according to claim 4, characterized in that: Step S4 Specifically, the following steps are included: S41. Quantify the contribution of each feature to fault prediction using SHAP values ​​to locate key sensitive features: In the formula, φ l Let f(t) represent the SHAP value of feature l, reflecting its marginal contribution to the model output; F represents the feature set of all fused feature vectors F(t); |F| represents the total number of feature sets; f(S) represents the prediction function value of the feature subset S after excluding feature j, and the feature subset S∈F; f(S∪{l}) represents the prediction function value after adding feature l to the feature subset S; S42. Optimization using KernelSHAP approximation: In the formula, φ' l This represents the SHAP value of the optimized feature l; z l (t) represents the spatiotemporal feature matrix of feature l; Z represents the sampling feature matrix; and f(l) represents the prediction function value of feature l. S43. Establish a feature-failure mapping table, associate SHAP values ​​with the corresponding physical failure mechanisms, and perform physical consistency correction: In the formula, This represents the influence value of feature l on physical failure; This indicates the sensitivity of feature l to physical failure; S44. Combined with the standardized health index h(t) output from the failure physical model: In the formula, b h This indicates the health index bias term; S45, Dynamic Weight Adjustment: h(t)←αh(t-1)+(1-α)h(t),α∈[0,1] (24); In the formula, h(t) and h(t-1) represent the standardized health index at time t and time t-1, respectively; α represents the health index forgetting factor; S46. Generate a heat map of fault risks; S461, Decompose the contribution of nodal sensors: In the formula, H m (t) represents the contribution value of the nodal sensor at the m-th decomposition point at time t; J m F represents the feature index associated with node sensor m; l (t) represents the fusion vector of the l-th feature; S462, Time Window Smoothing Process: In the formula, H(t) represents the contribution value of the node sensor after smoothing; H(r) represents the contribution value of the node sensor at index r. S463, Normalization process: In the formula, H(t)' represents the fault risk heat map; max(H(t)) and min(H(t)) represent the maximum and minimum contribution values ​​of the node sensors after smoothing, respectively.

6. The method for evaluating the operating status of a transformer according to claim 5, characterized in that: Step S4 is followed by step S5: Based on the standardized health index and the fault risk heat map, the collected parameters are adaptively optimized using digital twins and fed back to step S1 to form a real-time feedback loop of evaluation-optimization.

7. The method for evaluating the operating status of a transformer according to claim 6, characterized in that: Step S5 specifically includes the following steps: S51, Set the current acquisition parameter P current =(f s ,τ,g),f s τ and g represent the sampling frequency, trigger threshold, and node sensor gain, respectively; S52. Quantify the sensitivity S(t) of the collected parameters to fault detection based on the standardized health index h(t) and the fault risk heatmap H(t)': In the formula, Δf represents the simulation perturbation step size; ||·||F represents the Frobenius norm of the fault risk heatmap H(t)'; S53. Construct a multi-objective optimization model, where the objective function expression is as follows: In the formula, ω α ω β and ω γ All represent weighting coefficients; P default Represents a constant term; The constraints are: In the formula, f min and f max τ represents the minimum and maximum sampling frequencies, respectively; noise Indicates the noise floor threshold; g min and g max These represent the minimum and maximum node sensor gains, respectively; S54. The agent is trained in a digital twin environment to dynamically adjust the acquisition parameters, and the policy gradient update expression is as follows: In the formula, The gradient of the policy parameter θ; π θ This represents the policy network, used to output the action distribution; a t Indicates that in state s t The action below; Q π State-action value function; S55. Dynamically adjust acquisition parameters: P optimized =P current +η·clip(ΔP,-∈,∈) (32); In the formula, P optimized This represents the adjusted acquisition parameters; η represents the learning rate; ΔP represents the adjustment range; ∈ represents the maximum increment constraint; clip(·) represents the clip function; S56. Adjust the acquisition parameters P optimized Feedback is sent to step S1 to update the collected parameters.

8. A system based on the transformer operating status assessment method according to claim 7, characterized in that: include: The data acquisition and synchronization module is used to synchronously acquire multi-physics parameters and, after preprocessing, output a time-aligned multimodal data matrix and spatially registered sensor network coordinates. The spatiotemporal feature extraction module is used to extract spatiotemporal features from the multimodal data matrix and the spatially registered sensor network coordinates using a dynamic topological graph neural network, and outputs a spatiotemporal correlation feature tensor and a fault-sensitive feature vector. The error compensation and fusion module is used to compensate for errors in current temperature drift data, and to perform multimodal attention fusion on the compensated current temperature drift data, spatiotemporal correlation feature tensor, and fault-sensitive feature vector to output a fused feature vector. An interpretable assessment module is used to generate standardized health indices and failure risk heatmaps based on fused feature vectors, historical failure data, SHAP values, and failure physics models.

9. The system for evaluating the operating status of a transformer according to claim 8, characterized in that: It also includes a digital twin optimization module, which uses digital twins to adaptively optimize the collected parameters based on standardized health indices and fault risk heat maps, and feeds them back to step S1 to form a real-time feedback loop of evaluation-optimization.

Images