Multi-modal depth perception and adaptive segmentation full-spectrum primitive intelligent identification method, system and device for primary main wiring diagram of plant station, and medium

Abstract

The invention discloses a multi-modal depth perception and adaptive segmentation full-spectrum primitive intelligent identification method, system and device for a primary main wiring diagram of a plant station, and a medium, and belongs to the technical field of intelligent identification of engineering drawings of power system plant stations, and the method comprises the steps: fusing the dual-source data of a vector diagram and a grating diagram, and constructing a standard primitive library; secondly, primitive detection is realized by adopting a recognition and segmentation collaborative multi-stage network, and the contour precision is improved through model distillation; secondly, iteratively optimizing the recognition capability of the model on the difficult sample by using an active learning mechanism; and finally, finishing primitive classification and result optimization in combination with deep metric learning and electrical topology constraints. According to the invention, through multi-modal data fusion and multi-scale feature adaptive fusion, the recognition capability of a small target and a fine primitive is improved; by means of a double-branch heterogeneous architecture and a structural distillation mechanism, pixel-level high-precision edge segmentation is realized, and the identification problem of dense adhesion primitives is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system station engineering drawing intelligent recognition, in particular to a full-spectrum graph element intelligent recognition method, system, device and medium for multi-modal deep perception and adaptive segmentation of station primary main wiring diagram. BACKGROUND

[0002] In the existing intelligent processing technology of station primary main wiring diagram, a processing flow based on single-modal image is usually adopted, which can realize the basic functions of symbol detection, character recognition and topology reconstruction, but there are three major core shortcomings: first, the adaptation ability to symbol style diversity caused by different design templates, ages and drawing standards is poor, the model generalization is weak, and the false detection and missed detection rates are high; second, under the conditions of character and symbol adhesion, overlap and complex noise interference, the reliability of conventional OCR and element separation method is insufficient, which leads to semantic association failure; third, topology reconstruction is heavily dependent on heuristic rules, and the processing of line segment intersection and broken line is unstable, and there is lack of automatic verification and correction mechanism for fusing multi-modal evidence, which finally leads to that the robustness, accuracy and automation degree of the existing scheme are difficult to meet the needs of digital transformation of power system under real complex engineering blueprint scene. SUMMARY

[0003] In view of the above problems, the present application provides a full-spectrum graph element intelligent recognition method, system, device and medium for multi-modal deep perception and adaptive segmentation of station primary main wiring diagram.

[0004] Therefore, the technical problem solved by the present application is to solve the common problems of topology information loss and pixel detail recognition deficiency in the existing power engineering drawing recognition technology caused by relying on a single data source, especially through the design of multi-modal data fusion and multi-path detection network architecture.

[0005] To solve the above technical problems, the present application provides the following technical scheme: a full-spectrum graph element intelligent recognition method for multi-modal deep perception and adaptive segmentation of station primary main wiring diagram, comprising, The vector diagram and raster diagram of the power plant station blueprint are collected as double-source data, and the double-source data is fused in a unified coordinate system through multi-modal feature alignment; based on the fused double-source data, a graph element template is generated through image normalization processing and deep feature extraction embedding, and a standard graph element library is constructed using a vectorized symbol description method; the graph element is input into a multi-stage detection and segmentation network, a double-model recognition and segmentation is adopted, and the contour information is fused through model distillation to output the graph element recognition result; based on the confidence and error of the detection and segmentation result, uncertain samples are screened, and an active learning strategy is used to add the uncertain samples to the training set to iteratively optimize the model parameters of the multi-stage detection and segmentation network; the similarity between the feature vector of the detection and segmentation result and the template in the standard graph element library is calculated using deep metric learning to complete the graph element classification, and the classification result is optimized based on the graph theory constraint of the electrical topological relationship.

[0006] As a preferred scheme of the full-spectrum graph element intelligent recognition method for multi-modal deep perception and adaptive segmentation of power plant station primary main wiring diagram, wherein: the double-source data is fused in a unified coordinate system through multi-modal feature alignment includes, The vector diagram and raster diagram of the power plant station are collected as double-source data, and the double-source data is fused through multi-modal data collaborative preprocessing.

[0007] A standard graph element library is constructed based on the fused data, and a multi-stage detection and segmentation network is used to cooperatively perform recognition and segmentation tasks.

[0008] As a preferred scheme of the full-spectrum graph element intelligent recognition method for multi-modal deep perception and adaptive segmentation of power plant station primary main wiring diagram, wherein: the double-source data is fused in a unified coordinate system through multi-modal feature alignment includes, From the preprocessed double-source data, a typical power equipment graph element is extracted as a basis.

[0009] Through standardization processing and feature learning, a graph element template is generated.

[0010] Combined with the vectorized symbol description and incremental update mechanism, a standardized graph element template library is constructed.

[0011] As a preferred scheme of the full-spectrum graph element intelligent recognition method for multi-modal deep perception and adaptive segmentation of power plant station primary main wiring diagram, wherein: the double-source data is fused in a unified coordinate system through multi-modal feature alignment includes, The double-branch heterogeneous architecture is used to perform sparse detection and dense segmentation in parallel.

[0012] The sparse detection branch learns fine-grained contour information from the dense segmentation branch through structural distillation to output the detection and segmentation results of the graph element.

[0013] As a preferred scheme of the full-spectrum graph element intelligent identification method for the multi-modal deep perception and adaptive segmentation of the primary main wiring diagram of the substation, wherein the utilization of the double-branch heterogeneous architecture to perform sparse detection and dense segmentation in parallel includes, A multi-scale feature adaptive fusion network is constructed to extract fusion features from shallow details to deep semantics.

[0014] A recognition and segmentation double-model collaborative architecture is utilized to process sparse large targets and dense small targets respectively and output positioning and contour information in parallel.

[0015] The sparse detection branch learns fine-grained contour information from the dense segmentation branch through structural distillation to output the detection and segmentation results of the graph element. Fine-grained contour knowledge learned by the segmentation model is transferred to the recognition model through a model distillation strategy.

[0016] The double-branch outputs are fused, and after post-processing optimization, the graph element detection frame and pixel-level segmentation results are generated.

[0017] The beneficial effects of the preferred technical scheme are that through multi-scale feature fusion and a double-model collaborative architecture, the rapid positioning requirements of sparse large targets and the fine contour extraction requirements of dense small targets are effectively unified; further, with the aid of a structural distillation mechanism, the detection branch can fully learn the boundary detail features extracted by the segmentation branch, thereby improving the contour recognition accuracy of complex adherent graph elements and small targets while maintaining detection efficiency, and finally through feature fusion and post-processing optimization, the detection frame and pixel-level segmentation results are synergistically improved in accuracy and integrity.

[0018] As a preferred scheme of the full-spectrum graph element intelligent identification method for the multi-modal deep perception and adaptive segmentation of the primary main wiring diagram of the substation, wherein the based on the confidence and error of the detection and segmentation results, uncertain samples are screened, and a high-uncertainty sample is added to the training set using an active learning strategy includes, Based on the confidence evaluation of the recognition model recognition result, difficult samples with uncertainty are automatically screened out.

[0019] An active learning strategy is adopted to preferentially send the uncertainty into the training process.

[0020] The beneficial effects of this preferred technical solution are that by establishing an automatic screening mechanism for difficult samples based on confidence assessment, it is possible to accurately locate the uncertainty region in the model identification; combined with the active learning strategy, these representative samples are preferentially included in the training cycle, thus realizing the precise allocation of model optimization resources.

[0021] As a preferred embodiment of the multimodal depth perception and adaptive segmentation full-spectrum primitive intelligent recognition method for primary wiring diagrams of power plants described in this invention, the step of using depth metric learning to calculate the similarity between the feature vectors of the detection and segmentation results and templates in the standard primitive library to complete primitive classification includes... By using deep metric learning, the feature similarity between the recognition result and the template in the standard primitive library is calculated, thus completing the preliminary classification of primitives.

[0022] Based on the electrical connections and topological relationships of the power blueprint, graph theory constraint rules are constructed.

[0023] The graph matching algorithm is used to verify and optimize the preliminary classification results, and the identification results that do not conform to electrical logic are eliminated.

[0024] The beneficial effects of this preferred technical solution are that it achieves accurate comparison and preliminary classification of primitive features through deep metric learning, thereby improving the ability to distinguish between similar primitives; then, it introduces graph theory constraint rules based on electrical topology and uses graph matching algorithms to logically verify the recognition results, successfully integrating knowledge from the power system domain into the recognition process.

[0025] This invention provides a full-spectrum primitive intelligent recognition system for multimodal depth perception and adaptive segmentation of primary wiring diagrams of power plants.

[0026] To address the aforementioned technical problems, this invention provides the following technical solution: a full-spectrum primitive intelligent recognition system for multimodal depth perception and adaptive segmentation of primary wiring diagrams in power plants, comprising: a data acquisition and multimodal preprocessing module, a standard primitive library construction and management module, a multi-stage detection and segmentation network module, a difficult sample mining and active learning module, and a feature matching and result post-processing module.

[0027] The data acquisition and multimodal preprocessing module collects vector and raster images of power plant blueprints to form dual-source data, and fuses the dual-source data in a unified coordinate system through multimodal feature alignment.

[0028] The standard primitive library construction and management module, based on the fused dual-source data, generates primitive templates through image normalization processing and deep feature extraction and embedding, and constructs the standard primitive library using a vectorized symbol description method.

[0029] The multi-stage detection and segmentation network module inputs the graph element to the multi-stage detection and segmentation network, adopts a recognition and segmentation double model to cooperate, and outputs a graph element recognition result by fusing contour information through model distillation.

[0030] The difficult sample mining and active learning module screens uncertain samples based on the confidence and error of the detection and segmentation result, and adds the uncertain samples to the training set by using an active learning strategy to iteratively optimize the model parameters of the multi-stage detection and segmentation network.

[0031] The feature matching and result post-processing module calculates the similarity of the feature vectors of the detection and segmentation result and templates in a standard graph element library by using deep metric learning to complete graph element classification, and applies graph theory constraints based on electrical topological relations to optimize the classification result.

[0032] The application provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram.

[0033] The application provides a computer readable storage medium, which stores a computer program, and the computer program realizes the steps of the full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram when executed by a processor.

[0034] The application has the beneficial effects that: the application improves the recognition ability of small targets and fine graph elements through multi-modal data fusion and multi-scale feature adaptive fusion; the application realizes high-precision edge segmentation at the pixel level by means of a double-branch heterogeneous architecture and a structure distillation mechanism, and effectively solves the recognition problem of dense and adhered graph elements; and the application enhances the distinguishing ability of easily confused graph elements by combining an active learning strategy and a post-processing strategy based on topological rules, reduces false detection and missed detection, and improves processing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0036] Figure 1 A full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram is provided.

[0037] Figure 2A whole framework diagram of a full-spectrum graph element intelligent identification system for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0039] Embodiment 1, reference Figure 1 For an embodiment of the present application, the embodiment provides a full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram, comprising: S1, collecting vector diagrams and raster diagrams of power station blueprints to form double-source data, and fusing the double-source data in a unified coordinate system through multi-modal feature alignment.

[0040] S2, based on the fused double-source data, generating a graph element template through image normalization processing and deep feature extraction embedding, and constructing a standard graph element library using a vectorization symbol description method.

[0041] S3, inputting the graph element into a multi-stage detection and segmentation network, using a recognition and segmentation double-model collaboration, and outputting a graph element recognition result through model distillation fusion contour information.

[0042] S4, based on the confidence and error of the detection and segmentation result, screening uncertain samples, and using an active learning strategy to add the uncertain samples to the training set to iteratively optimize the model parameters of the multi-stage detection and segmentation network.

[0043] S5, using deep metric learning to calculate the similarity of the feature vectors of the detection and segmentation result and the templates in the standard graph element library to complete graph element classification, and based on the electrical topological relationship, applying graph theory constraints to optimize the classification result.

[0044] The present application fuses the double-source data of vector diagrams and raster diagrams, constructs a standard graph element library with scale and rotation invariance, and realizes collaborative optimization of recognition and segmentation tasks based on a multi-stage detection and segmentation network. Further, with the help of model distillation and active learning mechanism, the adaptive ability of the model to difficult samples is continuously improved. Finally, through deep metric learning and graph theory topological constraints, the graph element classification result is ensured to have both visual accuracy and electrical logical rationality, forming a complete technical solution that can accurately identify various graph elements in the primary main wiring diagram of a power station.

[0045] Embodiment 2, which is an embodiment of the present application, provides a full-spectrum graph element intelligent identification method based on the previous embodiment and oriented to a plant station primary main wiring diagram, comprising: The fusion of the dual-source data in the unified coordinate system through the multi-modal feature alignment in S1 comprises steps A1-A2: A1, collect the vector diagram and the raster diagram of the power plant station as dual-source data, and fuse them through multi-modal data collaborative preprocessing.

[0046] A2, construct a standard graph element library based on the fused data, and collaboratively perform the identification and segmentation tasks by using a multi-stage detection and segmentation network.

[0047] In the embodiments of the present application, the collaborative identification and segmentation branch in A2 collaboratively performs the identification and segmentation tasks, which is specifically realized by constructing a heterogeneous neural network architecture integrating a sparse detection branch and a dense segmentation branch. The sparse detection branch completes the coarse positioning and classification of the graph elements based on a cascaded detector, and the dense segmentation branch realizes the pixel-level contour extraction through a deformable convolution and a point rendering module. The two branches interact through shared multi-scale feature maps, and the fine-grained boundary knowledge learned by the segmentation branch is transferred to the detection branch through a structure distillation mechanism. Finally, the identification results with positioning accuracy and contour accuracy are output through non-maximum suppression and weight fusion.

[0048] In an alternative embodiment, the collaborative identification and segmentation task can be realized by constructing a double-branch heterogeneous architecture within a single deep learning network. Specifically, the network contains a sparse detection branch and a dense segmentation branch, and the two branches share the multi-scale feature maps generated by the feature extraction backbone network. The sparse detection branch uses a candidate box-based detector for positioning and rough classification of the graph elements, while the dense segmentation branch uses a pixel-level classification network for fine contour prediction of the graph elements. The two branches realize collaborative processing through information interaction at the feature level and fusion at the result level.

[0049] In another alternative embodiment, the collaborative identification and segmentation task can also be realized by deploying two independent and interactive special models. Specifically, a target detection model is used to perform the identification task, outputting the bounding box and class of the graph element, and a semantic segmentation model is used to perform the segmentation task, outputting the pixel-level mask of the graph element. The collaboration of the two models is realized through an asynchronous information fusion module, which guides the prediction of the segmentation model with the positioning information output by the detection model as the region of interest, and feeds back the fine contour information output by the segmentation model to the detection model for optimization and verification of the bounding box, thereby forming a collaborative closed loop.

[0050] In the embodiments of the present application, the image normalization processing and the deep feature extraction in S2, i.e. morphological normalization and deep features, specifically include: using a combination of morphological opening and closing operations to filter noise and smooth the contour of the original graph element, and then enhancing the texture features through adaptive histogram equalization; subsequently, inputting the normalized image into a pre-trained ResNet-50 network to extract a 2048-dimensional feature vector from the last convolutional layer thereof; and finally embedding the high-dimensional features into a 256-dimensional vector space through a feature dimension reduction module to form a graph element deep feature representation with rotation invariance and scale robustness.

[0051] In an alternative embodiment, the image normalization processing and the deep feature extraction can be realized by combining image preprocessing based on morphological operations with a pre-trained convolutional neural network (CNN). Specifically, the image normalization processing includes removing noise points and burrs in the graph element using morphological operations such as opening and closing operations, and normalizing the size of the graph element; the deep feature extraction is to input the normalized graph element image into a CNN model (such as ResNet, VGG) pre-trained on a large image dataset, and extract a high-dimensional feature vector from the fully connected layer or the last convolutional layer thereof as the deep feature representation of the graph element.

[0052] In another alternative embodiment, the image normalization processing and the deep feature extraction can also be realized by combining contour purification based on image segmentation with visual Transformer (ViT) feature encoding. Specifically, the image normalization processing first separates the graph element body from the complex background accurately through a semantic segmentation model, and extracts the minimum bounding rectangle region thereof to achieve geometric and illumination normalization; the deep feature extraction is to input the purified graph element region into a visual Transformer model, use its powerful global context perception ability, and generate more discriminative graph element deep features by calculating the embedding vector output by the class label.

[0053] Further, the generation of the graph element template in S2 through image normalization processing and deep feature extraction embedding, and the construction of the standard graph element library using vectorized symbolic description method include steps B1-B3: B1. Extracting typical power equipment graph elements as the basis from the preprocessed dual-source data.

[0054] B2. Generating a graph element template through standardization processing and feature learning.

[0055] B3. Constructing a standardized graph element template library in combination with vectorized symbolic description and incremental update mechanism.

[0056] In the embodiments of the present application, the generation of the template of the graphic element in B2 is achieved by image normalization processing and depth feature extraction, that is, a standardized graphic element template is generated. The specific implementation is as follows: first, the original graphic element is subjected to image normalization processing based on morphological opening operation and closing operation to eliminate noise and smooth the contour; then, the normalized graphic element image is input into a pre-trained ResNet-50 convolutional neural network on the ImageNet dataset, and a fixed-dimension depth feature vector is extracted from the last global pooling layer thereof; finally, the feature vector and the vector contour fitted by the Bezier curve are stored together to form a standardized graphic element template that combines pixel-level appearance information and geometric shape information.

[0057] In an alternative embodiment, the generation of the template of the graphic element can be achieved by geometric normalization based on a segmentation mask and a Vision Transformer feature extractor. Specifically, the accurate pixel-level mask of the graphic element is obtained by using a semantic segmentation model, and the graphic element is subjected to affine transformation according to the mask to achieve geometric normalization of its orientation and scale. Then, the normalized graphic element image is input into a pre-trained Vision Transformer model, and the output vector of the [CLS] marker is extracted as the depth feature descriptor of the graphic element, thereby generating the template.

[0058] In another alternative embodiment, the generation of the template of the graphic element can also be achieved by multi-view image enhancement and depth metric learning feature embedding. Specifically, the original graphic element is first subjected to random rotation, scaling and mirror transformation to generate multiple enhanced views. Then, a pre-trained Siamese network is used as a feature encoder, and Triplet Loss is introduced in the training process to make the features of the same class of graphic elements gather in the embedding space. Finally, the mean of the feature vectors of the multiple enhanced views is used as the stable and discriminative template feature of the graphic element.

[0059] In the embodiments of the present application, the fusion of the contour information in S3 by model distillation is achieved by transferring the fine-grained contour information learned by the dense segmentation branch to the sparse detection branch through structural distillation. The specific implementation is as follows: in the training stage, the parameters of the dense segmentation branch are fixed as a teacher model to guide the learning of the sparse detection branch. The same batch of image data is input into the double-branch network, the high-response feature map generated by the segmentation branch in the boundary region is used as fine-grained contour knowledge, and a designed distillation loss function (such as a feature imitation loss based on KL divergence) is used to force the corresponding layer features of the detection branch to approach them, so that the detection branch learns the more accurate boundary perception ability implicitly while completing the positioning and classification tasks. Finally, only the optimized detection branch needs to be used in the inference stage to output the results that combine classification confidence and contour accuracy.

[0060] In an alternative embodiment, the fusion of contour information through model distillation can be attention mask based feature imitation; specifically, a boundary attention mask is generated using the pixel-level prediction map output by the segmentation branch, which is used to weight the activation values of the boundary regions in the feature space, and then through a mean square error loss function, the feature map of the detection branch is forced to align with the weighted teacher feature map in the boundary region, so as to realize the directional transfer of contour knowledge.

[0061] In another alternative embodiment, the fusion of contour information through model distillation can also be the transfer of structural information through relationship distillation; specifically, not only the output features of the teacher model are imitated, but further the relationship structure within the features is imitated, for example, the mutual relationship matrix between different spatial position points in the segmentation branch feature map is calculated as structural knowledge, and a relationship distillation loss is designed to force the detection branch to learn the same relationship pattern, thereby transferring the understanding ability of the segmentation model for the overall structure of the graph element to the detection model.

[0062] Further, the output of the graph element recognition result in S3 by adopting the recognition and segmentation dual-model cooperation and fusing the contour information through model distillation includes steps C1-C2: C1, performing sparse detection and dense segmentation in parallel by using a dual-branch heterogeneous architecture.

[0063] C2, learning fine-grained contour information from the dense segmentation branch by the sparse detection branch through structural distillation to output the detection and segmentation results of the graph element.

[0064] Further, C1 performing sparse detection and dense segmentation in parallel by using a dual-branch heterogeneous architecture includes steps C11-C12: C11, constructing a multi-scale feature adaptive fusion network to extract fusion features from shallow details to deep semantics.

[0065] C12, using a recognition and segmentation dual-model cooperative architecture to process sparse large targets and dense small targets respectively and output positioning and contour information in parallel.

[0066] Further, C2 learning fine-grained contour information from the dense segmentation branch by the sparse detection branch through structural distillation to output the detection and segmentation results of the graph element includes steps C21-C22: C21, transferring the fine-grained contour knowledge learned by the segmentation model to the recognition model through a model distillation strategy.

[0067] C22, fusing the outputs of the dual branches, and generating the graph element detection frame and the pixel-level segmentation result after post-processing optimization.

[0068] Further, based on the confidence and error of the detection and segmentation results, uncertain samples are screened, and high-uncertainty samples are added to the training set using an active learning strategy including steps D1-D2: D1, based on the confidence evaluation of the recognition model recognition result, automatically screen out uncertain difficult samples.

[0069] D2, using an active learning strategy, preferentially sending uncertain samples into the training process.

[0070] Further, the similarity between the feature vector of the detection and segmentation result in S5 and the template in the standard graph element library is calculated using deep metric learning to complete graph element classification, including steps F1-F3: F1, calculate the feature similarity between the recognition result and the template in the standard graph element library through deep metric learning, and complete the preliminary classification of the graph element.

[0071] Specifically, input each graph element recognition result (image patch in the detection box or binary mask after segmentation) into a pre-trained feature encoding network (e.g., the same CNN or Transformer network as used when building the graph element library), and map it to a high-dimensional feature vector For , calculate the cosine similarity or Euclidean distance reciprocal between all standard feature vectors of each class of graph element template in the standard graph element library .

[0072] According to the nearest neighbor principle or the highest average similarity principle, assign a preliminary class label and the corresponding confidence score . .

[0073] F2, based on the electrical connection and topological relationship of the power blueprint, construct graph theory constraint rules.

[0074] Based on electrical engineering knowledge, abstract the connection rules of the power primary wiring diagram into graph theory constraints. Specifically: Define different categories of graph elements as different types of graph nodes (e.g., transformer nodes, circuit breaker nodes, bus nodes).

[0075] Define electrical connection lines (or direct connection relationships) as edges in the graph.

[0076] Formalize a series of electrical logic rules, such as: Connectivity constraint: a "disconnector" node must be directly connected to a "circuit breaker" node and a "bus" node.

[0077] Degree constraint: the degree (number of connections) of a "transmission line" node is usually 1 or 2.

[0078] Adjacency constraint: a "transformer" node should not be directly adjacent to another "transformer" node.

[0079] Further, from the constructed graph theory constraint rules, all the relevant constraint rules of the preliminary classification of V are retrieved . .

[0080] For each rule , check whether the local neighborhood of V (the nodes and edges directly connected to it) in the identified result graph satisfies . .

[0081] Each time a rule is satisfied , score 1; otherwise, score 0. Alternatively, for non-binary rules (such as requiring the number of connections to be within a certain range), a satisfaction degree (between 0 and 1) can be calculated.

[0082] The local rule compliance rate of V is calculated as: where k is the total number of rules related to this category.

[0083] Evaluate the importance of node V in the entire blueprint topology. Analyze the global connectivity.

[0084] For example, measure the importance of V in the network by calculating graph theory indicators such as node betweenness centrality.

[0085] Normalize the calculated centrality value to between 0 and 1. A graph element will have a higher value if it is on a critical electrical path.

[0086] Finally, the topology consistency score of node VV is the weighted combination of the above local and global factors: where is an adjustable weight parameter used to control the relative importance of local rules and global structure in scoring.

[0087] F3, use graph matching algorithm to verify and optimize the preliminary classification results, and eliminate the identification results that do not meet the electrical logic.

[0088] ​​All F1 primary classified primitives are taken as nodes, and connecting edges are established according to their spatial adjacency relationship on the blueprint (for example, by analyzing the relative position of the end points of the connecting lines and the primitive symbols), so as to construct a to-be-verified "recognition result graph" .

[0089] The is matched with the graph theory constraint rules constructed in F2. Each node and its local neighborhood in the are traversed to check whether it meets the predefined rules.

[0090] For the recognition result with serious constraint violation (for example, a "current transformer" directly connects two "busbars", which does not conform to the electrical logic) and low F1 stage confidence , the system marks it as a false positive and removes it.

[0091] For the result with moderate constraint violation but still acceptable, a reclassification process is started. For example, a primitive initially classified as a "circuit breaker" will be forced to recalculate the similarity in the "disconnector" category of the standard primitive library if its connection relationship is more consistent with the topology pattern of a "disconnector". If the new similarity is higher than a certain threshold, the category will be corrected to "disconnector".

[0092] The final classification confidence is the weighted fusion result of the initial visual similarity confidence and the topology consistency score , that is , represents the balance factor.

[0093] Embodiment 3, refer to Figure 2 , an embodiment of the present application, which provides a full-spectrum primitive intelligent recognition system for power plant primary wiring diagram multi-modal deep perception and adaptive segmentation, comprising: a data acquisition and multi-modal preprocessing module, a standard primitive library construction and management module, a multi-stage detection and segmentation network module, a difficult sample mining and active learning module, and a feature matching and result post-processing module.

[0094] The data acquisition and multi-modal preprocessing module acquires the vector graph and raster graph of the power plant blueprint to form dual-source data, and fuses the dual-source data in a unified coordinate system through multi-modal feature alignment.

[0095] The standard primitive library construction and management module generates primitive templates through image normalization processing and deep feature extraction embedding based on the fused dual-source data, and constructs a standard primitive library using a vectorized symbol description method.

[0096] The multi-stage detection and segmentation network module inputs the graph element into the multi-stage detection and segmentation network, adopts a dual model of recognition and segmentation in cooperation, and outputs the graph element recognition result by fusing contour information through model distillation.

[0097] The difficult sample mining and active learning module screens uncertain samples based on the confidence and error of the detection and segmentation result, and adds the uncertain samples to the training set by using an active learning strategy to iteratively optimize the model parameters of the multi-stage detection and segmentation network.

[0098] The feature matching and result post-processing module calculates the similarity of the feature vector of the detection and segmentation result and the template in the standard graph element library by using deep metric learning to complete the graph element classification, and optimizes the classification result based on the graph theory constraint of the electrical topological relationship.

[0099] The embodiment also provides an electronic device suitable for a full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram as proposed in the above embodiment.

[0100] The embodiment also provides a storage medium having a computer program stored thereon, and the program is executed by a processor to realize the full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram as proposed in the above embodiment.

[0101] The storage medium proposed in the embodiment and the full-spectrum graph element intelligent identification method for multi-modal deep perception and adaptive segmentation of a power station primary main wiring diagram proposed in the above embodiment belong to the same inventive concept, and the technical details not described in the embodiment can be referred to the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.

[0102] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and necessary universal hardware, and of course can also be implemented by hardware, but in many cases the former is a better implementation. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH, a hard disk, or an optical disc, and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A full-spectrum diagram element intelligent identification method for substation-oriented primary wiring diagram multi-modal deep perception and adaptive segmentation, characterized in that: include, The vector and raster images of the power plant blueprints constitute dual-source data, which are then fused in a unified coordinate system through multimodal feature alignment. Based on the fused dual-source data, primitive templates are generated through image normalization processing and deep feature extraction and embedding, and a standard primitive library is constructed using a vectorized symbol description method. The primitives are input into a multi-stage detection and segmentation network, which employs a dual-model approach of recognition and segmentation, and fuses contour information through model distillation to output primitive recognition results. Based on the confidence and error of the detection and segmentation results, uncertain samples are screened, and an active learning strategy is used to add uncertain samples to the training set to iteratively optimize the model parameters of the multi-stage detection and segmentation network. Deep metric learning is used to calculate the similarity between the feature vectors of the detection and segmentation results and the templates in the standard primitive library to complete primitive classification. Graph theory constraints are applied based on electrical topology relationships to optimize the classification results.

2. The full-spectrum graph neural intelligent identification method for substation-oriented primary wiring diagram multi-modal deep perception and adaptive segmentation of claim 1, wherein: The method of fusing dual-source data in a unified coordinate system through multimodal feature alignment includes, Vector and raster images of power plants were collected as dual-source data and fused through multimodal data collaborative preprocessing. A standard primitive library is built based on fused data, and a multi-stage detection and segmentation network is used to collaboratively perform recognition and segmentation tasks.

3. The full-spectrum graph neural intelligent identification method for substation-oriented primary single-line diagram multi-modal deep perception and adaptive segmentation of claim 2, wherein: The process of generating primitive templates through image normalization processing and deep feature extraction and embedding, and constructing a standard primitive library using a vectorized symbol description method includes: Typical power equipment elements are extracted from the preprocessed dual-source data as the basis; Primitive templates are generated through standardization and feature learning; By combining vectorized symbol description with incremental update mechanism, a standardized primitive template library is constructed.

4. The full-spectrum graph neural intelligence identification method for substation-oriented primary distribution diagram multi-modal deep perception and adaptive segmentation of claim 3, wherein: The method employs a dual-model approach combining recognition and segmentation, and fuses contour information through model distillation to output primitive recognition results, including... Utilize a dual-branch heterogeneous architecture to perform sparse detection and dense segmentation in parallel; Structural distillation enables sparse detection branches to learn fine-grained contour information from dense segmentation branches, thereby outputting primitive detection and segmentation results.

5. The full-spectrum graph neural intelligence identification method for substation-oriented primary distribution diagram multi-modal deep perception and adaptive segmentation of claim 4, wherein: The parallel execution of sparse detection and dense segmentation using a dual-branch heterogeneous architecture includes, Construct a multi-scale feature adaptive fusion network to extract fused features from shallow details to deep semantics; By utilizing a dual-model collaborative architecture for recognition and segmentation, sparse large targets and dense small targets are processed separately, and localization and contour information are output in parallel. The step of enabling sparse detection branches to learn fine-grained contour information from dense segmentation branches through structural distillation to output primitive detection and segmentation results includes, The fine-grained contour knowledge learned by the segmentation model is transferred to the recognition model through the model distillation strategy. The dual-branch outputs are merged and post-processed for optimization to generate primitive detection boxes and pixel-level segmentation results.

6. The full-spectrum graph neural intelligence identification method for substation-oriented primary single-line diagram multi-modal deep perception and adaptive segmentation of claim 4, wherein: Based on the confidence and error of the detection and segmentation results, uncertain samples are screened, and an active learning strategy is used to add high-uncertainty samples to the training set. Based on the confidence assessment of the recognition model's recognition results, difficult samples with uncertainty are automatically filtered out. An active learning strategy is adopted to prioritize introducing uncertainty into the training process.

7. The full-spectrum graph neural intelligence identification method for substation-oriented primary single-line diagram multi-modal deep perception and adaptive segmentation of claim 4, wherein: The similarity between the feature vector of the detection and segmentation result and the template in the standard graph element library is calculated by deep metric learning to complete the classification of the graph element, including, The similarity between the feature vector of the detection and segmentation result and the template in the standard graph element library is calculated by deep metric learning to complete the classification of the graph element, including, Based on the electrical connection and topological relationship of the power blueprint, graph theory constraint rules are constructed; The preliminary classification result is verified and optimized by using a graph matching algorithm, and the recognition result that does not conform to the electrical logic is eliminated.

8. A full-spectrum graph element intelligent identification system for substation primary main wiring diagram oriented multi-modal deep perception and adaptive segmentation, applying a full-spectrum graph element intelligent identification method for substation primary main wiring diagram oriented multi-modal deep perception and adaptive segmentation according to any one of claims 1-7, characterized in that, It includes: The data acquisition and multi-modal preprocessing module, the standard graph element library construction and management module, the multi-stage detection and segmentation network module, the difficult sample mining and active learning module, and the feature matching and result post-processing module; The data acquisition and multi-modal preprocessing module acquires the vector graph and raster graph of the power plant station blueprint to form double-source data, and fuses the double-source data in a unified coordinate system through multi-modal feature alignment; The standard graph element library construction and management module generates a graph element template through image normalization processing and deep feature extraction embedding based on the fused double-source data, and constructs a standard graph element library using a vectorized symbol description method; The multi-stage detection and segmentation network module inputs the graph element into the multi-stage detection and segmentation network, adopts a dual-model collaborative recognition and segmentation method, and outputs the graph element recognition result by fusing contour information through model distillation; The difficult sample mining and active learning module selects uncertain samples based on the confidence and error of the detection and segmentation result, and adds the uncertain samples to the training set using an active learning strategy to iteratively optimize the model parameters of the multi-stage detection and segmentation network; The feature matching and result post-processing module calculates the similarity between the feature vector of the detection and segmentation result and the template in the standard graph element library by deep metric learning to complete the classification of the graph element, and optimizes the classification result based on the graph theory constraint of the electrical topological relationship. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the full-spectrum graph element intelligent recognition method for power plant station primary wiring diagram multi-modal deep perception and adaptive segmentation according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the full-spectrum graph element intelligent recognition method for power plant station primary wiring diagram multi-modal deep perception and adaptive segmentation according to any one of claims 1-7.

Images