Method for extracting and fusing multi-dimensional features of substation secondary circuit

Abstract

This invention belongs to the field of power system automation operation and maintenance technology, and relates to a method for multi-dimensional feature extraction and fusion of substation secondary circuits. The method includes: acquiring text logs, configuration files, and electrical digital sequences from the substation secondary circuit generation dataset; extracting semantic features, time-series waveform features, and topological adjacency matrices in parallel; processing semantic features and time-series waveform features through a cross-attention mechanism to generate a preliminary fusion vector; converting the preliminary fusion vector into a topological node embedding vector using a graph network operator; feeding the topological node embedding vector back to the cross-attention mechanism update loop using a topological invariant orthogonal constraint causal operator; calculating the topological constraint matrix and truncating the cross-attention weights; correcting the feature projection geometric scale; and achieving feature space alignment. This invention solves mapping distortion, eliminates the semantic gap, improves feature fidelity and topological alignment accuracy, and ensures the geometric convergence of the multimodal feature matrix within a unified mapping space.

Description

Technical Field

[0001] This invention belongs to the field of power system automation operation and maintenance technology, and relates to a method for multi-dimensional feature extraction and fusion of substation secondary circuits. Background Technology

[0002] Currently, utilizing intelligent measurement and control and multi-source data analysis for safety prevention and condition diagnosis of primary and secondary systems in substations has become an important means of automated operation and maintenance in power systems. However, existing technologies still have certain limitations in addressing the dynamic adjustment of substation topology as operating modes change, and in ensuring the coordinated operation of primary and secondary models throughout the entire lifecycle.

[0003] First, there is the substation topology identification technology based on static parsing of configuration files. For example, Chinese invention patent CN109857885B discloses an automatic generation method for main wiring diagrams of intelligent substations based on SCD files. This method identifies the virtual connection relationships and topological connection relationships between devices by parsing the substation configuration description file and constructing a pre-sorted traversal tree. This scheme mainly relies on static configuration text for topology analysis, which is difficult to accurately reflect the changes in the transient switching of switches and the actual physical state of equipment during operation. At the same time, there may be non-standard formats and missing configuration information in the SCD file. Existing methods fail to transform these uncertainties into dynamic constraints when updating the topology matrix. Therefore, when the substation operation mode changes, the identification results may show topology omissions or connection errors that are inconsistent with the actual physical network structure.

[0004] Second, there is the secondary equipment error prevention verification technology based on graph neural networks and semantic knowledge graphs. For example, Chinese invention patent CN119051250A discloses an intelligent error prevention method for secondary equipment models based on relation-aware graph neural networks. This method constructs a power equipment database into a semantic knowledge graph and dynamically updates the association weights between devices through a neighborhood relation perception module and a graph attention network. This technology is mainly used for semantic recognition and open-loop prediction verification of business instructions, which can improve the feature extraction accuracy of complex secondary equipment association relationships. However, the equipment association score obtained by semantic reasoning has not yet formed a closed-loop association with the real-time physical monitoring data of the primary and secondary sides of the substation, such as real-time steady-state electrical quantities. When the execution deviation of the preceding operation occurs, the existing method is difficult to reflect the actual impact of the deviation on the subsequent control path in a timely manner, and therefore it is not easy to identify potential misoperation risks under extreme concurrent operating conditions.

[0005] Third, there is the substation post-fault diagnosis technology based on multimodal information fusion. For example, Chinese invention patent application CN120197018A discloses a substation fault diagnosis method, device and system based on multimodal fusion. This method comprehensively utilizes alarm information, fault waveform data and SCADA measurement data to extract and fuse multidimensional features of different types of fault information, and predicts the fault results accordingly. Existing diagnostic models of this kind usually treat fault diagnosis as a relatively independent post-fault analysis process, and rarely consider the continuous impact of the status of graph data during the long-term dynamic changes of the operation mode. In the whole life cycle operation and maintenance optimization process, if there is no mandatory linkage mechanism to feed back the post-fault diagnosis results such as protection failure to operate to the pre-fault error prevention logic, the prediction results are prone to cumulative deviation with the operation process, which may eventually exceed the safety boundary required by the virtual and real fusion of the whole station, making the output results difficult to use in actual operation.

[0006] Therefore, how to adapt to the dynamic changes in primary and secondary topologies during the intelligent operation and maintenance of substations, and how to establish a two-way linkage between pre-emptive error prevention and post-event diagnosis so that relevant analysis and judgment always conform to the safety boundaries of the actual physical system, remains a problem that existing technologies need to solve. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art, solve the technical problems of feature space mismatch in the joint transformation mapping process of heterogeneous operating data and the lack of topological constraints for accurate alignment of cross-modal features, and provide a method for multi-dimensional feature extraction and fusion of substation secondary circuits.

[0008] To achieve the above-mentioned objectives, this invention provides a method for multi-dimensional feature extraction and fusion of secondary circuits in substations, comprising the following steps: Step S1: Extract multi-dimensional features of heterogeneous data and generate a topological adjacency matrix: Obtain the heterogeneous operation dataset generated by the secondary circuit of the substation, which includes equipment change text logs, virtual terminal configuration files and electrical quantity digital sequences; call the heterogeneous operator to process the heterogeneous operation dataset, extract semantic feature vectors and time-series waveform feature vectors, and parse the virtual terminal configuration files to generate a topological adjacency matrix. Step S2: Use the cross-attention mechanism to generate a preliminary multimodal fusion vector: Call the multi-head cross-attention mechanism component to process the semantic feature vector and the temporal waveform feature vector, and generate a preliminary multimodal fusion feature vector; Step S3, topological embedding feedback constraints and orthogonal truncation to achieve feature alignment: The multimodal preliminary fusion feature vector is converted into topological nodes. The graph attention network operator is called in combination with the topological adjacency matrix to perform topological graph embedding operation and calculate the topological node embedding vector. The orthogonal constraint causal operator of topological invariant is configured inside the multi-head cross-attention mechanism component. The topological node embedding vector is fed back as a static topological constraint invariant to the parameter update loop of the multi-head cross-attention mechanism component. The conjugate dot product of the topological node embedding vector and the query matrix inside the multi-head cross-attention mechanism component is calculated to obtain the topological constraint matrix. The orthogonal projection truncation of the cross-attention weights is performed using the topological constraint matrix to achieve multimodal data feature alignment.

[0009] Before extracting the time-series waveform feature vector by calling heterogeneous operators in parallel, the present invention further includes the following filtering steps: Step S11, calculate the absolute value of the first-order difference of the electrical quantity digital sequence in real time. When the absolute value of the first-order difference is greater than 3.5 times the steady-state mean, automatically switch to the blocking state and import the 3rd level discrete wavelet decomposition operator; Step S12, truncate the non-stationary pulse components with energy density exceeding 0.70, and use inverse wavelet transform to reconstruct a pure sequence to replace the original electrical quantity digital sequence, thereby blocking the penetration of pulse noise from the source.

[0010] The present invention further includes the following topological perturbation regularization step after generating the topological adjacency matrix by parsing the virtual terminal configuration file: Step S13, applying a random hidden edge removal operator with an edge removal probability of 0.15 to the topological adjacency matrix to construct a virtual topological matrix containing perturbation features; and using the virtual topological matrix as a regularization constraint term to input the graph attention network operator to carry out joint graph embedding calculation, thereby broadening the decision boundary of the multimodal preliminary fusion feature vector in the multidimensional feature space and eliminating the spatial mapping drift caused by loop fault samples.

[0011] The specific steps for extracting the semantic feature vector in step S1 of the present invention are as follows: Step S14, call the text mining operator with the substation secondary circuit knowledge lexicon, perform word segmentation and stop word filtering on the equipment change text log, and map to generate a semantic encoding vector containing the equipment change action, time and the relationship between virtual and physical terminals, as the semantic feature vector.

[0012] The specific steps for mapping and generating the preliminary multimodal fusion feature vector in step S2 of the present invention are as follows: Step S21, the semantic feature vector and the temporal waveform feature vector are mapped to a unified low-dimensional feature space through linear transformation matrices respectively, and the corresponding multimodal key vector and multimodal value vector are calculated and generated; and the multi-head cross-attention operation logic is called to perform joint probability weighting on the multimodal key vector and multimodal value vector to generate the preliminary multimodal fusion feature vector.

[0013] The specific steps for performing orthogonal projection truncation on the cross-attention weights in step S3 of the present invention are as follows: Step S31, calculate the orthogonal complement space projection matrix of the topological constraint matrix in the orthogonal basis space, and perform matrix dot product with the original cross-attention weight matrix inside the multi-head cross-attention mechanism component using the orthogonal complement space projection matrix to block the weight update path that does not conform to the physical topological constraints of the quadratic loop and lock the topological invariance between modes.

[0014] The specific steps of calling the graph attention network operator to perform the topological graph embedding operation in step S3 of the present invention are as follows: Step S32, attach the multimodal preliminary fused feature vector to each node of the topological adjacency matrix, calculate the attention allocation coefficient between adjacent secondary loop nodes using the multi-head graph attention mechanism, perform aggregation update on the feature vector of adjacent nodes according to the attention allocation coefficient, and calculate and output the topological node embedding vector containing the local topological association structure.

[0015] After calculating and outputting the topology node embedding vector in step S3, the present invention further includes the following adaptive fusion weight allocation and final fusion steps: calculating normalized information entropy values ​​for the aligned semantic feature vector, the aligned temporal waveform feature vector, and the topology node embedding vector respectively, quantifying the information uncertainty under the current operating conditions, inversely allocating adaptive fusion weights based on the normalized information entropy values, and using the adaptive fusion weights to linearly weight and combine each aligned feature vector to generate the final multidimensional fusion feature vector.

[0016] After generating the final multidimensional fused feature vector, the present invention further includes the following state anomaly discrimination step: inputting the final multidimensional fused feature vector into the secondary circuit state classification operator for pattern recognition, and outputting defect circuit identification parameters and state anomaly probability indicators that characterize the current operating conditions of the substation's secondary circuit.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: 1. By using the topological node embedding vector and the query matrix to form a topological constraint matrix, and by truncating the cross attention weight through orthogonal projection, feature components that do not conform to the topological connection relationship of the quadratic loop can be excluded. This reduces the feature distortion caused by the different spatial scales of the semantic feature vector and the temporal waveform feature vector. As a result, the fused features can better preserve the original data and its topological relationship, making the configuration defects that are easily covered by feature space offset more clearly expressed, and providing a more reliable feature basis for subsequent state anomaly judgment.

[0018] 2. By embedding the topology node vector as a constraint and feeding it back to the multi-head cross-attention mechanism component, the variation range of the cross-attention weight can be limited according to the actual topological relationship of the secondary loop. When the electrical quantity digital sequence exhibits nonlinear distortion or abnormal fluctuations, this constraint can reduce the transmission of interference between different modal features, prevent the weight update from deviating from the normal range, and make the alignment process of cross-modal features more stable. This can improve the mapping accuracy of different types of operating data in a unified feature space, and also help improve the adaptability and pattern recognition efficiency of the multimodal data fusion model.

[0019] 3. By applying a random hidden object removal operator to the topological adjacency matrix and constructing a virtual topological matrix, a moderate topological perturbation constraint can be provided for the graph attention network operator, reducing the impact of a small number of defective samples on the feature distribution and decision boundary. This processing can reduce the spatial mapping drift caused by the uneven distribution of rare loop fault samples, enabling the model to maintain a relatively stable recognition ability under small sample conditions and reducing overfitting to limited training samples. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the multimodal feature extraction and alignment process of the present invention; Figure 2 This is a schematic diagram of the substation field equipment operation and maintenance diagnostic system of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings.

[0022] Example 1: This embodiment discloses a method for multi-dimensional feature extraction and fusion of secondary circuits in substations, which runs on a data processing system and includes the following steps: Step S1: Extract multi-dimensional features of heterogeneous data and generate a topological adjacency matrix: Obtain the heterogeneous operation dataset generated by the secondary circuit of the substation, which includes equipment change text logs, virtual terminal configuration files and electrical quantity digital sequences; call the heterogeneous operator to process the heterogeneous operation dataset, extract semantic feature vectors and time-series waveform feature vectors, and parse the virtual terminal configuration files to generate a topological adjacency matrix. Step S2: Use the cross-attention mechanism to generate a preliminary multimodal fusion vector: Call the multi-head cross-attention mechanism component to process the semantic feature vector and the temporal waveform feature vector, and generate a preliminary multimodal fusion feature vector; Step S3, topological embedding feedback constraints and orthogonal truncation to achieve feature alignment: The multimodal preliminary fusion feature vector is converted into topological nodes. The graph attention network operator is called in combination with the topological adjacency matrix to perform topological graph embedding operation and calculate the topological node embedding vector. The orthogonal constraint causal operator of topological invariant is configured inside the multi-head cross-attention mechanism component. The topological node embedding vector is fed back as a static topological constraint invariant to the parameter update loop of the multi-head cross-attention mechanism component. The conjugate dot product of the topological node embedding vector and the query matrix inside the multi-head cross-attention mechanism component is calculated to obtain the topological constraint matrix. The orthogonal projection truncation of the cross-attention weights is performed using the topological constraint matrix to achieve multimodal data feature alignment.

[0023] Before extracting the time-series waveform feature vector by calling heterogeneous operators in parallel, the following filtering steps are also included in this embodiment: Step S11, calculate the absolute value of the first-order difference of the electrical quantity digital sequence in real time. When the absolute value of the first-order difference is greater than 3.5 times the steady-state mean, automatically switch to the blocking state and import the 3rd level discrete wavelet decomposition operator; Step S12, truncate the non-stationary pulse components with energy density exceeding 0.70, and use inverse wavelet transform to reconstruct a pure sequence to replace the original electrical quantity digital sequence, thereby blocking the penetration of pulse noise from the source.

[0024] This embodiment, after generating the topological adjacency matrix by parsing the virtual terminal configuration file, also includes the following topological perturbation regularization step: Step S13, applying a random hidden edge removal operator with an edge removal probability of 0.15 to the topological adjacency matrix to construct a virtual topological matrix containing perturbation features; and using the virtual topological matrix as a regularization constraint term to input into the graph attention network operator to perform joint graph embedding calculation, thereby broadening the decision boundary of the multimodal preliminary fusion feature vector in the multidimensional feature space and eliminating the spatial mapping drift caused by loop fault samples.

[0025] The specific steps for extracting the semantic feature vector in step S1 of this embodiment are as follows: Step S14, call the text mining operator with the substation secondary circuit knowledge lexicon, perform word segmentation and stop word filtering on the equipment change text log, and map to generate a semantic encoding vector containing the equipment change action, time, and virtual and physical terminal association relationship, as the semantic feature vector.

[0026] The specific steps for mapping and generating the preliminary multimodal fusion feature vector in step S2 of this embodiment are as follows: Step S21, the semantic feature vector and the time-series waveform feature vector are mapped to a unified low-dimensional feature space through linear transformation matrices, and the corresponding multimodal key vector and multimodal value vector are calculated and generated; and the multi-head cross-attention operation logic is called to perform joint probability weighting on the multimodal key vector and multimodal value vector to generate the preliminary multimodal fusion feature vector.

[0027] The specific steps for performing orthogonal projection truncation on the cross-attention weights in step S3 of this embodiment are as follows: Step S31, calculate the orthogonal complement space projection matrix of the topological constraint matrix in the orthogonal basis space, and perform matrix dot product with the original cross-attention weight matrix inside the multi-head cross-attention mechanism component using the orthogonal complement space projection matrix to block the weight update path that does not conform to the physical topological constraints of the quadratic loop and lock the topological invariance between modes.

[0028] The specific steps for calling the graph attention network operator to perform topological graph embedding operation in step S3 of this embodiment are as follows: Step S32, attach the multimodal preliminary fused feature vector to each node of the topological adjacency matrix, calculate the attention allocation coefficient between adjacent secondary loop nodes using the multi-head graph attention mechanism, perform aggregation update on the feature vector of adjacent nodes based on the attention allocation coefficient, and calculate and output the topological node embedding vector containing the local topological association structure.

[0029] After calculating and outputting the topology node embedding vector in step S3, this embodiment further includes the following adaptive fusion weight allocation and final fusion steps: calculate the normalized information entropy value for the aligned semantic feature vector, the aligned time-series waveform feature vector, and the topology node embedding vector respectively, quantify the information uncertainty under the current operating conditions, allocate the adaptive fusion weight in reverse according to the normalized information entropy value, and use the adaptive fusion weight to linearly weight and combine each aligned feature vector to generate the final multidimensional fusion feature vector.

[0030] After generating the final multidimensional fusion feature vector, this embodiment also includes the following state anomaly discrimination step: inputting the final multidimensional fusion feature vector into the secondary circuit state classification operator for pattern recognition, and outputting defect circuit identification parameters and state anomaly probability indicators that characterize the current operating conditions of the substation's secondary circuit.

[0031] Example 2: In this embodiment, during the continuous operation of a smart substation secondary circuit monitoring scenario, when frequent circuit breaker operations trigger high-frequency transient electromagnetic interference, the data acquisition unit acquires a heterogeneous operational dataset generated by the substation secondary circuit. This heterogeneous operational dataset includes equipment change text logs, virtual terminal configuration files, and electrical quantity digital sequences. A time window sliding alignment operator is set at the feature extraction front end. This operator monitors transient abrupt changes in the electrical quantity digital sequences and slides a 5-second time window forward and backward based on these transient abrupt changes, aligning and binding the equipment change text logs and virtual terminal configuration files that fall within the time window. The feature extraction unit concurrently calls heterogeneous operators to process the heterogeneous operational dataset. Specifically, a text mining operator with a substation secondary circuit knowledge lexicon performs word segmentation and stop word filtering on the equipment change text logs, mapping equipment change actions, time, and virtual / physical terminal relationships into semantic encoding vectors. A bidirectional encoder representation model then extracts contextual features and outputs semantic feature vectors. One-dimensional convolutional neural networks utilize multiple layers of convolutional kernels to process high-frequency timing fluctuations in digital electrical quantity sequences, outputting timing waveform feature vectors. Simultaneously, it parses each secondary loop node and its virtual terminal connection relationship in the virtual terminal configuration file to generate a topological adjacency matrix. .

[0032] The control fusion unit sets up a multi-head cross-attention mechanism component to integrate semantic feature vectors. With time-series waveform feature vector Each feature is mapped to a unified low-dimensional feature space through a linear transformation matrix, forming a query matrix, a multimodal key vector, and a multimodal value vector. Cross-attention weights are then calculated in eight independent attention subspaces. The calculation rule for the cross-attention weight matrix is ​​expressed as follows: ,in, Indicates the first The modality and the first Cross-attention weights between modalities The query matrix representing the first modality features. The transpose of the key matrix representing the second modality features, with superscript... This represents the matrix transpose operation. The scaling operator is represented; the multi-head cross-attention mechanism component uses cross-attention weights to perform joint probability weighting on the multimodal value vectors, generating a preliminary multimodal fusion feature vector. The graph network configuration unit will initially fuse feature vectors from multiple modalities. Dependent on topological adjacency matrix The initial node feature vectors of each node are used to calculate the attention allocation coefficients between adjacent secondary loop nodes using a multi-head graph attention mechanism. Then, the feature vectors of adjacent nodes are aggregated and updated based on the attention allocation coefficients, and the topological node embedding vector containing the local topological association structure is output. .

[0033] Preliminary fusion of feature vectors in multimodal models Calculation and topology node embedding vector The outputs are configured with orthogonal constraint causal operators with topological invariants. These operators are implemented in the hardware processing system using a field-programmable gate array (FPGA) hardware pipeline or a dedicated tensor computation kernel. Internally, they include a topological feature cache register, a cross-spatial projection multiplier, an orthogonal complement space acquisition logic unit, and a weight mask truncation circuit. The topological feature cache register reads the topological node embedding vector from the previous sampling period. The cross-space projection multiplier performs dimension alignment and vector inner product operations with the query matrix of the current channel. The orthogonal complement space acquisition logic unit generates a topological constraint matrix based on the conjugate dot product result and calculates the orthogonal complement space projection matrix of this topological constraint matrix in the orthogonal basis space. Then, the orthogonal complement space projection matrix is ​​multiplied by the original cross-attention weight matrix inside the multi-head cross-attention mechanism component. The weight mask truncation circuit outputs a low-order binary mask array and sends it to the cross-attention multiplier to block weight update paths that do not conform to the physical topological constraints of the quadratic loop; the topological node embedding vector This serves as a static topological constraint invariant that is fed back to the parameter update loop of the multi-head cross-attention mechanism component. After the cross-attention weights are truncated by orthogonal projection, the output is a multimodal preliminary fusion feature vector aligned with topological constraints. Topological invariant orthogonal constraint causal operators also embed topological nodes into vectors. The spatial norm is incorporated as a correction component into the scaling operator. The denominator term is used to limit the value boundaries of each element in the cross-attention weight matrix.

[0034] An adaptive gated filtering unit is set at the input of a one-dimensional convolutional neural network. This unit calculates the absolute value of the first-order difference of the electrical quantity digital sequence in real time. When the absolute value of the first-order difference is greater than 3.5 times the steady-state mean, it switches to a blocking state, and the electrical quantity digital sequence is fed into a level 3 discrete wavelet decomposition operator to truncate high-frequency non-stationary pulse components with energy density exceeding 0.70. Then, the pure sequence is reconstructed through inverse wavelet transform, and the pure sequence replaces the original electrical quantity digital sequence and is input into the one-dimensional convolutional neural network. The topological adjacency matrix is ​​generated. Then, the graph network configuration unit is configured to move to the topological adjacency matrix according to an edge culling probability of 0.15. A random hidden surface removal operator is applied to construct a virtual topology matrix containing perturbation features. This virtual topology matrix is ​​then used as a regularization constraint term to input the graph attention network operator for joint graph embedding computation, thereby broadening the multimodal preliminary fusion feature vector. The decision boundary in the multidimensional feature space; under the operating conditions of this embodiment, when the multimodal feature vectors aligned by topological constraints are initially fused... When the state information entropy value is greater than the safety constraint limit of 0.85 and lasts for 3 consecutive sampling periods, the above-mentioned topology disturbance regularization process is initiated. Each parameter is determined based on the physical simulation experiment and statistical results under the typical operating conditions of the secondary circuit of the intelligent substation. Among them, the first-order difference absolute value threshold is taken as 3.5 times the steady-state mean. When it is lower than this multiple, it is easy to judge the steady-state small fluctuations as transient pulses. When it is higher than this multiple, high-frequency pulses may pass through the filtering unit. The energy density threshold of 0.70 is determined based on the boundary position of the energy spectrum distribution of the pulse noise and the effective signal. When it is higher than 0.70, it may cause the loss of high-frequency effective features. The edge blanking probability of 0.15 is used to achieve a balance between introducing topology disturbances and maintaining the connectivity of the backbone topology. The safety constraint limit of 0.85 is determined based on the 3 sigma statistical boundary of the uncertainty index of the system's maximum information entropy. When the state information entropy value exceeds this limit, it indicates that the heterogeneous data loss rate has reached a critical state.

[0035] During normal operation, the value arbitration unit calculates the aligned semantic feature vector, the aligned temporal waveform feature vector, and the topological node embedding vector, respectively. The normalized information entropy value is used to inversely allocate adaptive fusion weights, and the aligned feature vectors are linearly weighted and combined; the allocation rule of the adaptive fusion weights is expressed as follows: ,in, Indicates the first Adaptive fusion weights for each feature modality Indicates the first The normalized information entropy value of each feature mode. Indicates the first participating in the summation The normalized information entropy values ​​of each feature mode; the final multidimensional fused feature vector is represented as: ,in, Indicates the first The node feature vectors after topology alignment of each feature mode; the substation secondary circuit state anomaly detection module will finally fuse the multi-dimensional feature vectors. The secondary circuit state classification operator is input for pattern recognition. The pattern recognition result is used as the defect circuit identification parameter, and the corresponding classification probability is used as the state anomaly probability index. Thus, the current operating condition of the substation's secondary circuit is output. Under high-frequency transient electromagnetic interference conditions, the topological adjacency relationship forms boundary constraints on the high-frequency background noise and timing jitter components in the multimodal preliminary fusion feature vector, so that the cross-modal features are aligned under the secondary circuit topology and the system operating state converges to the preset circuit state monitoring boundary.

[0036] Example 3: In this embodiment, within a 110kV smart substation secondary circuit test environment constructed using a digital-physical hybrid simulation system, the data acquisition unit retrieves field records and generates a heterogeneous operational dataset, which serves as the unified input source for each group of experiments. This heterogeneous operational dataset includes equipment change logs carrying unstructured text information, virtual terminal configuration files carrying semi-structured topology information, and electrical quantity digital sequences carrying structured time-series waveform information. To simulate electromagnetic interference in industrial settings, Gaussian white noise with a signal-to-noise ratio of 20dB is superimposed on the electrical quantity digital sequences, and a 50Hz power frequency is mixed in. Interference harmonics; when determining the window length of the time window sliding alignment operator, the data processing system selects parameters based on the timestamp resolution of the data stream and the buffer overflow probability; an excessively long time window will increase the data processing load, while an excessively short time window will disrupt the transient evolution process reflected by the equipment change text log and the electrical quantity digital sequence; for this circuit breaker tripping transient condition, the sliding alignment time window length is set to 5s, so that the equipment change text log, virtual terminal configuration file and electrical quantity digital sequence falling within this time range are aligned according to a unified time reference before being input into the feature extraction unit.

[0037] Four test groups were set up under the above input conditions. The first group used complete multimodal feature extraction, topology graph embedding, and feedback constraint structure; the second group removed the semantic feature vectors corresponding to device change text logs. The third group removes the topology adjacency matrix generated by the virtual terminal configuration file. The fourth group retains all modal inputs but disconnects the loop of orthogonal projection truncation of cross-attention weights by the topological invariant orthogonal constraint causal operator. All groups use the same electrical quantity digital sequence and test conditions to eliminate the influence of input differences on the comparison results. During the test, the adaptive gating filter unit continuously calculates the absolute value of the first-order difference of the electrical quantity digital sequence. When the absolute value of the first-order difference exceeds 3.5 times the steady-state mean, the filter unit imports the electrical quantity digital sequence into a level 3 discrete wavelet decomposition operator, truncating non-stationary pulse components with energy densities exceeding 0.70, and then reconstructs the pure sequence through inverse wavelet transform. The feature extraction unit generates a time-series waveform feature vector based on the pure sequence. and semantic feature vector Common input multi-head cross-attention mechanism component; graph attention network operator combined with topological adjacency matrix A topological node embedding vector is generated, and the orthogonal constraint causal operator of the topological invariant feeds this vector back to the update process of the cross attention weights. The cross attention weights are obtained by scaling the dot product of the query matrix of the first modality feature and the key matrix of the second modality feature, and the spatial norm of the topological node embedding vector is used as the denominator correction component of the scaling operator, thereby restricting the boundary of the element values ​​of the weight matrix.

[0038] After the above processing, the first group using the complete structure achieved a defect loop identification accuracy of 98.4% and a feature space alignment geometric distortion factor of 0.12; the second group had its semantic feature vectors removed. Subsequently, equipment change information could not participate in cross-modal association, the defect loop identification accuracy dropped to 74.3%, and the feature space alignment geometric distortion factor increased to 0.45; the third group removed the topological adjacency matrix. Subsequently, the cross-attention weights lost the constraint of the secondary loop connection relationship, and the defect loop identification accuracy dropped to 68.1%, while the feature space alignment geometric distortion factor increased to 0.62. In the fourth group, after disconnecting the orthogonal projection truncated loops, although semantic features, temporal waveform features, and topological graph embedding results were still retained, the topological node embedding vector could not be applied to the update path of the cross-attention weights. The defect loop identification accuracy was 81.5%, and the feature space alignment geometric distortion factor was 0.33. The results of the four groups indicate that the semantic feature vector... Timing waveform feature vector and topological adjacency matrix When participating in the calculation together, the orthogonal constraint causal operator of topological invariants can use the physical connection relationship of the quadratic loop to limit the feature projection range. The identification accuracy of the complete structure is higher than that of the three comparative test groups, and the feature space alignment geometric distortion factor is lower than that of the three comparative test groups.

[0039] To verify the applicability of the first-order differential absolute value trigger threshold, gradient experiments were conducted using thresholds of 1.5, 3.5, and 5.5. When the trigger threshold was set to 1.5, even minor fluctuations under normal steady-state conditions triggered the filtering unit, and the electrical quantity digital sequence was frequently introduced into the discrete wavelet decomposition and inverse wavelet transform link, resulting in a reduction in local fine-grained temporal deformation features. The defect loop identification accuracy dropped to 83.2%, and the data processing delay increased to 45.8 ms. When the trigger threshold was set to 5.5, transient electromagnetic interference pulses with amplitudes reaching 4.0 times the steady-state mean did not trigger the filter. Electrical quantity digital sequences containing high-frequency pulses directly entered the one-dimensional convolutional neural network, causing a shift in the spatial scale of the temporal waveform features. The defect loop identification accuracy dropped to 71.6%. Under a 305V voltage level, when the trigger threshold was set to 3.5, the filtering unit was able to block high-frequency pulses while retaining effective temporal waveform features, and the defect loop identification accuracy reached 98.4%.

[0040] When assessing the impact of information uncertainty on the fusion process, three levels of working conditions—low, medium, and high—were constructed by adjusting the test data loss rate. In the low information uncertainty condition, the normalized information entropy value of each feature vector remained at 0.35. The value arbitration unit reversibly allocated adaptive fusion weights based on the normalized information entropy value and linearly weighted the aligned semantic feature vector, temporal waveform feature vector, and topological node embedding vector. The resulting pattern recognition convergence time for the multidimensional fused feature vector was 10.4 ms. As the data loss rate increased, the normalized information entropy value rose to 0.65. The value arbitration unit then redistributed the adaptive fusion weights for each modality, and the pattern recognition convergence time became 12.1 ms.

[0041] When the data loss rate continues to increase, and the state information entropy value of the multidimensional fused feature vector eventually exceeds the safety constraint limit of 0.85 and persists for 3 sampling periods, the graph network configuration unit moves towards the topological adjacency matrix. A random hidden-line removal operator with an edge removal probability of 0.15 is applied to construct a virtual topology matrix containing perturbation features. This virtual topology matrix is ​​then used as a regularization constraint term to input the graph attention network operator for joint graph embedding computation. Under the combined effects of strong noise and high data loss rate, after initiating this topology perturbation regularization process, the pattern recognition convergence time remains at 14.2 ms. Without initiating this process, the feature vector diverges, and the convergence time exceeds 120.0 ms. This demonstrates that by allocating adaptive fusion weights based on the normalized information entropy value and introducing a virtual topology matrix when the state information entropy value reaches the trigger condition, the final multidimensional fused feature vector can maintain a convergent state under different information uncertainty conditions.

[0042] Example 4: In this embodiment, when the substation secondary circuit operates continuously for more than 180 days, the circuit breaker contact loss causes the high-frequency transient electromagnetic interference intensity to exceed 5.0 times the steady-state average, and the expansion of the heterogeneous network causes a lag in the update of the virtual terminal configuration file, the data acquisition unit continuously acquires the equipment change text log, the virtual terminal configuration file, and the electrical quantity digital sequence to form a heterogeneous operation dataset. The feature extraction unit monitors the electrical quantity digital sequence, and when the time-series voltage amplitude deviates from the steady-state baseline by 15%, it uses the corresponding transition inflection point as the time reference, and extracts an analysis time domain interval spanning 5 seconds forward and backward. The equipment change text log and electrical quantity digital sequence falling within this interval are aligned. The graph network configuration unit reads the current virtual terminal configuration file and generates an initial topology adjacency matrix according to the virtual terminal connection relationship. In this system, each intelligent electronic device corresponds to a topological node, and the static connections in the virtual terminal loops correspond to directed topological edges, thereby converting the secondary loop connection relationship into matrix data that can be processed by the graph attention network operator.

[0043] The control fusion unit calls the multi-head cross-attention mechanism component to convert the semantic feature vectors. With time-series waveform feature vector Input eight independent attention subspaces and compute the initial fusion feature vector of the multimodal modes. The graph network configuration unit will Dependent on topological adjacency matrix For each node, perform topological graph embedding operations using graph attention network operators, and output topological node embedding vectors. To avoid the current sampling period With not yet generated To prevent circular dependencies, the system employs time discretization with a delay of one sampling period. During the forward computation of the current sampling period, it reads the computationally calculated and stored data from the cache unit from the previous sampling period. It is used as a static topological constraint invariant input to the multi-head cross-attention mechanism component, so that a single forward computation maintains a unidirectional acyclic data flow.

[0044] Within the multi-head cross-attention mechanism component, a topologically invariant orthogonal constrained causal operator is configured. Before performing the conjugate dot product operation, a learnable cross-modal space aligned linear transformation matrix is ​​first applied. Mapped to the same low-dimensional feature space as the query matrix, then dimension alignment and normalization are performed. The topological constraint matrix is ​​obtained by performing a complex-domain conjugate transpose dot product with the query matrix. Then, the orthogonal complement projection matrix of the topological constraint matrix in the orthogonal basis space is calculated. This projection matrix is ​​then multiplied by the original cross-attention weight matrix, setting weights that do not conform to the physical topological connection relationship of a quadratic loop to zero. This yields a preliminary multimodal fusion feature vector aligned with topological constraints. The previous sampling period Simultaneously, as a static topological constraint invariant, the parameter update loop is fed back to the multi-head cross-attention mechanism component to limit the cross-attention weight update path in subsequent sampling periods.

[0045] Topological invariant orthogonal constraint causal operator will also The spatial norm is added as a correction component to the denominator of the scaling operator; to avoid the denominator becoming too small or the weight values ​​diverging due to changes in the characteristic amplitude, it is first calculated... The second norm of the electrical quantity is calculated and dimensionlessly normalized using a smoothing correction constant, limiting its dynamic range to between 0.1 and 2.0. This dynamic range is then added to the square root scalar of the original channel dimension to form a composite scaling factor. When nonlinear distortion occurs in the electrical quantity acquisition, causing a sudden change in the norm, a global maximum norm truncation threshold is used to limit the upper limit of the denominator, keeping the cross-attention weights within a computable range. An adaptive gated filtering unit is set at the input of the one-dimensional convolutional neural network to calculate the absolute value of the first-order difference of the electrical quantity digital sequence in real time. When the absolute value of the first-order difference is greater than 3.5 times the steady-state mean, the electrical quantity digital sequence is imported into a level 3 discrete wavelet decomposition operator to truncate high-frequency non-stationary pulse components with energy densities exceeding 0.70. The pure sequence is then reconstructed through inverse wavelet transform and used to replace the original electrical quantity digital sequence as input to the one-dimensional convolutional neural network.

[0046] The value arbitration unit calculates the aligned semantic feature vector, the aligned temporal waveform feature vector, and the topological node embedding vector, respectively. The normalized information entropy value; when the state information entropy value corresponding to the final fusion process is greater than the safety constraint limit of 0.85 and lasts for 3 sampling periods, the graph network configuration unit moves to the topological adjacency matrix. A random hidden-face removal operator with an edge removal probability of 0.15 is applied to construct a virtual topology matrix containing perturbation features. This virtual topology matrix is ​​then used as a regularization constraint term to input the graph attention network operator for joint graph embedding computation. When the state information entropy value is no greater than 0.85, the value arbitration unit inversely allocates adaptive fusion weights based on the normalized information entropy value of each feature mode, and performs linear weighted combination on each aligned feature vector to generate the final multidimensional fusion feature vector. .

[0047] During topology perturbation regularization, the system follows Calculate the dynamic reconstruction probability, where, This indicates the intensity of topology reconfiguration demand within the current sliding time window. This represents the proportional adjustment coefficient, which is set to 0.15. Represents the initial fusion feature vector of multimodal modes The statistical variance within the current sliding time window The smoothing term, denoted as 0.01, is used to prevent the denominator from being zero. The dynamic reconstruction probability is used to adjust the regularization penalty strength of the virtual topology matrix in the loss function within the graph attention network. The virtual topology matrix remains a regularization constraint term for joint graph embedding computation and is not replaced by the dynamic reconstruction probability. This regularization penalty adjusts the graph embedding weights of the graph attention network operators in the discrete training steps through gradient backpropagation, and applies to the feature topology alignment stage. The adaptive fusion weights linearly weight each feature vector after topology alignment, with topology alignment performed first and feature fusion performed later. The substation secondary circuit state anomaly detection module receives the final multi-dimensional fused feature vector. The data is then input into the secondary loop state classification operator for pattern recognition, and the defect loop identification parameters and state anomaly probability index are output. Under the conditions of high-frequency transient electromagnetic interference and lag in updating virtual terminal configuration files, the topology constraint matrix restricts the update range of cross-attention weights, so that semantic features, time-series waveform features and topology node features are aligned according to the secondary loop connection relationship, and the system operating state converges to the preset loop state monitoring boundary.

[0048] Example 5: This embodiment combines Figures 1 to 2 This section describes a method for multi-dimensional feature extraction and fusion of secondary circuits in substations, such as... Figure 1 As shown, step S1 extracts multi-dimensional features from heterogeneous data and generates a topological adjacency matrix. Then, the process flow proceeds to step S2, where a cross-attention mechanism is used to produce a preliminary multimodal fusion vector. Finally, the process flow continues to step S3 to achieve topological embedding feedback constraints and orthogonal truncation to achieve feature alignment.

[0049] like Figure 2As shown, the substation field equipment operating environment includes intelligent electronic devices and circuit breakers and circuits. Intelligent electronic devices contain virtual terminal configuration files and equipment change text logs, while circuit breakers and circuits contain electrical quantity digital sequences and high-frequency transient electromagnetic interference. The operating and interference data generated by the substation field equipment operating environment are transmitted to the data acquisition and filtering hardware pipeline via a field data bus. This hardware pipeline includes a data acquisition unit and an adaptive gated filtering unit. The data acquisition unit acquires heterogeneous operating datasets and performs static reference threshold calibration. The adaptive gated filtering unit performs first-order difference absolute value calculation, discrete wavelet decomposition operator, and clean sequence reconstruction. The data processed by the data acquisition and filtering hardware pipeline is then transmitted through an internal high-speed pipeline. The data is transmitted to a dedicated computing kernel in the hardware processing system. This kernel includes a feature extraction unit, a control fusion unit, and a graph network configuration unit. The feature extraction unit includes a time window sliding alignment operator, a text mining operator, a one-dimensional convolutional neural network, and a topology parsing operator. The control fusion unit includes a multi-head cross-attention mechanism component, a topological invariant orthogonal constraint causal operator, and a cross-attention multiplier. The graph network configuration unit includes a graph attention network operator and a random hidden surface removal operator. The multi-dimensional feature matrix output by the dedicated computing kernel is transmitted to an operation and maintenance diagnostic output server. This server includes a value arbitration unit, a secondary loop state classification operator, defect loop identification parameters, and an abnormal state probability index.

[0050] Example 6: In this embodiment, when a newly commissioned substation enters the secondary circuit field deployment and feature space initial parameter calibration stage, the data acquisition unit continuously acquires a 60-second raw electrical quantity digital sequence in a static environment where no external excitation signal is applied to the circuit. This sequence is then input into an adaptive gated filtering unit for background noise baseline calibration. The adaptive gated filtering unit calculates the time-domain standard deviation of the raw electrical quantity digital sequence, thereby determining a static reference threshold to reduce the impact of electromagnetic environment differences at different sites on subsequent feature extraction and multimodal data feature alignment. The calculation rule for the static reference threshold is expressed as follows: ,in, Indicates the static reference threshold. This represents the proportional adjustment coefficient, which is fixed at 3.0. This represents the time-domain standard deviation of the original electrical quantity digital sequence during a 60-second static acquisition period. The system will... The data is written into the internal register of the adaptive gated filter unit to initialize the measurement baseline of the steady-state mean and the input amplitude normalization parameter. After entering the online operation stage, the adaptive gated filter unit calculates the first-order difference absolute value of the electrical quantity digital sequence in real time, and uses whether the first-order difference absolute value is greater than 3.5 times the steady-state mean as the blocking condition, and imports the electrical quantity digital sequence that meets the condition into the level 3 discrete wavelet decomposition operator.

[0051] After writing the static baseline threshold, the graph network configuration unit imports the virtual terminal configuration file of the newly commissioned substation and generates a topology adjacency matrix based on the virtual terminal connection relationship. The feature extraction unit processes the pre-loaded device change text logs into semantic feature vectors. And extract timing waveform feature vectors from the electrical quantity digital sequence. Then use the static reference threshold The amplitude scales of the two types of input features are normalized; the multi-head cross-attention mechanism component will... and Mapping to a unified low-dimensional feature space generates a preliminary multimodal fusion feature vector. The diagram network configuration unit will... Dependent on topological adjacency matrix Each node is used to perform topological graph embedding operations through graph attention network operators, and the output is a topological node embedding vector. .

[0052] The control fusion unit invokes the orthogonal constraint causal operator of topological invariants to embed the topological nodes into the vector. As a static topological constraint invariant, it is fed back to the parameter update loop of the multi-head cross-attention mechanism component, and utilizes The topological constraint matrix is ​​generated by the conjugate dot product of the query matrix and the topological constraint matrix. When the field data bus transmission generates a 125μs delay, the system uses the hysteresis residual corresponding to this delay as the correction amount between the feature projection results of the current sampling period and the previous sampling period. This correction is used to update the orthogonal complement projection matrix of the topological constraint matrix in the orthogonal basis space. Then, the updated orthogonal complement projection matrix is ​​multiplied by the original cross-attention weight matrix to block the weight update path that does not conform to the physical topological constraints of the secondary loop. After initial parameter calibration and orthogonal projection truncation, the projection range of the cross-modal features converges to the boundary defined by the static connection relationship of the secondary loop. The system operation state enters the preset loop state monitoring boundary where the defect loop identification accuracy is greater than or equal to 95% and the pattern recognition convergence time is stable within 15ms.

[0053] Example 7: In this embodiment, during online monitoring of the secondary circuit in a substation, when the equipment change text log records the activation / deactivation of protection devices, switching of soft pressure plates, or adjustment of virtual terminal connection relationships, and simultaneously when the electrical quantity digital sequence exhibits corresponding timing fluctuations, the data acquisition unit acquires the equipment change text log, virtual terminal configuration file, and electrical quantity digital sequence to form a heterogeneous operation dataset. The adaptive gating filtering unit calculates the absolute value of the first-order difference of the electrical quantity digital sequence in real time. When this value is greater than 3.5 times the steady-state mean, the electrical quantity digital sequence is imported into a level 3 discrete wavelet decomposition operator to truncate non-stationary pulse components with energy densities exceeding 0.70, and then reconstructs the pure sequence through inverse wavelet transform. The feature extraction unit calls the heterogeneous operator to process the heterogeneous operation dataset, obtaining semantic feature vectors respectively. and time-series waveform eigenvectors It also parses the virtual terminal configuration file to generate a topological adjacency matrix. The multi-head cross-attention mechanism component will integrate semantic feature vectors. With time-series waveform feature vector Mapped to a unified low-dimensional feature space, a preliminary multimodal fusion feature vector is generated by weighting the multimodal key vector and the multimodal value vector using joint probability weights. ;Graph attention network operators will Dependent on topological adjacency matrix For each node, the node features are aggregated and updated based on the attention allocation coefficients between adjacent secondary loop nodes to obtain the topological node embedding vector. .

[0054] Topological invariant orthogonal constraint causal operator embeds topological nodes into vectors As a static topological constraint invariant, it is fed back to the parameter update loop of the multi-head cross-attention mechanism component, and utilizes The topological constraint matrix is ​​generated by the conjugate dot product of the query matrix and the topological constraint matrix. The system calculates the orthogonal complement projection matrix of the topological constraint matrix in the orthogonal basis space, and performs a matrix dot product of this projection matrix with the original cross-attention weight matrix to truncate weight update paths that do not conform to the physical topological connection relationship of the quadratic loop, so that the semantic features, temporal waveform features and topological node features are aligned in the same feature space. The value arbitration unit calculates the aligned semantic feature vector, the aligned temporal waveform feature vector and the topological node embedding vector respectively. The normalized information entropy value is used to inversely allocate adaptive fusion weights, and the aligned feature vectors are linearly weighted to generate the final multidimensional fusion feature vector. When the uncertainty of the running data reaches the trigger condition for topology perturbation regularization, the graph network configuration unit moves to the topology adjacency matrix according to an edge culling probability of 0.15. A random hidden surface removal operator is applied to construct a virtual topology matrix, which is then used as a regularization constraint term to input the graph attention network operator for joint graph embedding computation.

[0055] The secondary loop state classification operator receives the final multidimensional fused feature vector. The system performs pattern recognition on each candidate operating state and outputs the normalized probability corresponding to each candidate state. The system uses the candidate state with the highest probability and its corresponding secondary loop node as the defect loop identification parameter, and uses the normalized probability corresponding to the candidate state as the state anomaly probability index, thereby determining the location of the secondary loop corresponding to the abnormal state and its probability of occurrence. The state anomaly discrimination result is updated synchronously with the final multi-dimensional fusion feature vector of the current sampling period, so that equipment change information, electrical quantity fluctuations and virtual terminal topology relationships can jointly participate in the secondary loop operating state discrimination.

Claims

1. A method for multi-dimensional feature extraction and fusion of secondary circuits in substations, characterized in that, Includes the following steps: Step S1: Obtain the heterogeneous operation dataset generated by the substation secondary circuit, which includes equipment change text logs, virtual terminal configuration files, and electrical quantity digital sequences; The heterogeneous operator is invoked to process the heterogeneous running dataset, extract semantic feature vectors and time-series waveform feature vectors, and parse the virtual terminal configuration file to generate a topological adjacency matrix; Step S2: Call the multi-head cross-attention mechanism component to process the semantic feature vector and the temporal waveform feature vector, and produce a preliminary multimodal fusion feature vector; Step S3: Convert the multimodal preliminary fusion feature vector into topological nodes, and combine the topological adjacency matrix with the graph attention network operator to perform topological graph embedding operation to calculate the topological node embedding vector; configure the topological invariant orthogonal constraint causal operator inside the multi-head cross-attention mechanism component, and feed the topological node embedding vector as a static topological constraint invariant back to the parameter update loop of the multi-head cross-attention mechanism component; calculate the conjugate dot product of the topological node embedding vector and the query matrix inside the multi-head cross-attention mechanism component to obtain the topological constraint matrix, and use the topological constraint matrix to perform orthogonal projection truncation on the cross-attention weights to achieve multimodal data feature alignment.

2. The method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 1, characterized in that, Before extracting the time-series waveform feature vector by calling heterogeneous operators in parallel, the following filtering steps are also included: Step S11, calculate the absolute value of the first-order difference of the electrical quantity digital sequence in real time. When the absolute value of the first-order difference is greater than 3.5 times the steady-state mean, automatically switch to the blocking state and import the 3rd level discrete wavelet decomposition operator; Step S12, truncate the non-stationary pulse components with energy density exceeding 0.70, and use the inverse wavelet transform to reconstruct a pure sequence to replace the original electrical quantity digital sequence, thereby blocking the penetration of pulse noise from the source.

3. The method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 1, characterized in that, After parsing the virtual terminal configuration file to generate the topological adjacency matrix, the following topological perturbation regularization steps are also included: Step S13, apply a random hidden edge removal operator with an edge removal probability of 0.15 to the topological adjacency matrix to construct a virtual topological matrix containing perturbation features; and use the virtual topological matrix as a regularization constraint term to input the graph attention network operator to carry out joint graph embedding calculation, broaden the decision boundary of the multimodal preliminary fusion feature vector in the multidimensional feature space, and eliminate the spatial mapping drift caused by loop fault samples.

4. The method for multi-dimensional feature extraction and fusion of substation secondary circuits according to claim 1, characterized in that, The specific steps for extracting the semantic feature vector in step S1 are as follows: Step S14, call the text mining operator with the substation secondary circuit knowledge lexicon, perform word segmentation and stop word filtering on the equipment change text log, and map to generate a semantic encoding vector containing the equipment change action, time, and virtual and physical terminal association relationship, as the semantic feature vector.

5. The method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 1, characterized in that, The specific steps for mapping and generating the preliminary multimodal fusion feature vector in step S2 are as follows: Step S21, the semantic feature vector and the temporal waveform feature vector are mapped to a unified low-dimensional feature space through linear transformation matrices, and the corresponding multimodal key vector and multimodal value vector are calculated and generated; and the multi-head cross-attention operation logic is called to perform joint probability weighting on the multimodal key vector and multimodal value vector to generate the preliminary multimodal fusion feature vector.

6. The method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 1, characterized in that, The specific steps for performing orthogonal projection truncation on the cross-attention weights in step S3 are as follows: Step S31, calculate the orthogonal complement space projection matrix of the topological constraint matrix in the orthogonal basis space, and perform matrix dot product with the original cross-attention weight matrix inside the multi-head cross-attention mechanism component using the orthogonal complement space projection matrix to block the weight update path that does not conform to the physical topological constraints of the quadratic loop and lock the topological invariance between modes.

7. The method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 1, characterized in that, The specific steps for calling the graph attention network operator to perform the topological graph embedding operation in step S3 are as follows: Step S32, attach the multimodal preliminary fused feature vector to each node of the topological adjacency matrix, calculate the attention allocation coefficient between adjacent secondary loop nodes using the multi-head graph attention mechanism, perform aggregation update on the feature vector of adjacent nodes based on the attention allocation coefficient, and calculate and output the topological node embedding vector containing the local topological association structure.

8. The method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 1, characterized in that, After calculating the topology node embedding vector in step S3, the following adaptive fusion weight allocation and final fusion steps are also included: normalized information entropy values ​​are calculated for the aligned semantic feature vector, aligned temporal waveform feature vector and topology node embedding vector respectively, the information uncertainty under the current operating condition is quantified, adaptive fusion weights are allocated in reverse according to the normalized information entropy values, and the aligned feature vectors are linearly weighted and combined using the adaptive fusion weights to generate the final multidimensional fusion feature vector.

9. A method for multi-dimensional feature extraction and fusion of secondary circuits in a substation according to claim 8, characterized in that, After generating the final multidimensional fusion feature vector, the following state anomaly discrimination steps are also included: inputting the final multidimensional fusion feature vector into the secondary circuit state classification operator for pattern recognition, and outputting defect circuit identification parameters and state anomaly probability indicators that characterize the current operating conditions of the substation's secondary circuit.

Images