Hydropower station structural plane analysis method and system based on prompt learning and medium

Abstract

The invention discloses a hydropower station structural plane analysis method and system based on prompt learning and a medium. Relates to the technical field of engineering geology. The method comprises the following steps: acquiring multi-modal basic data of a structural surface of a hydropower station, and preprocessing the multi-modal basic data; inputting the preprocessed multi-modal basic data into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model performs cross-modal mapping of multi-modal basic data by sharing an embedding space, and introduces a prompt learning mechanism to perform similarity analysis to obtain a hydropower station structural surface classification result and a similarity result; according to the scheme, a multi-modal AI framework and domain knowledge driving method is fused, and cross-modal feature fusion of multi-modal data such as texts, images and spaces is realized through a multi-modal large model and a prompt learning technology; and meanwhile, engineering specifications are coded into priori knowledge of the model by utilizing a prompt learning template, so that the integrating degree of an output result and an industrial standard is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering geology, and in particular to a method, system and medium for analyzing the structural surface of a hydropower station based on prompt learning. Background Art

[0002] In recent years, the application of artificial intelligence technology in the field of geological engineering has gradually deepened, but there is still a significant gap between its technical path and the needs of engineering practice. At the AI ​​technology level, existing research focuses on the single-modal data processing paradigm. However, due to the lack of a cross-modal alignment mechanism, it is difficult to integrate text descriptions and spatial topological information, resulting in fragmented feature expression. In addition, traditional deep learning models (such as deep neural networks (DNNs)) generally face the dilemma of interpretability. Their black box characteristics are out of touch with the semantics of geological engineering specifications, making it difficult to convert the output results into verifiable technical reports. Insufficient dynamic adaptability is also a core bottleneck. Existing clustering algorithms (such as K-means) rely on fixed threshold parameters and have a high misjudgment rate in areas with strong spatial variability, which seriously restricts the ability to generalize across engineering scenarios.

[0003] Traditional geological workflows are highly empirically dependent and inefficient. For example, analyzing hydropower station structural surfaces requires manual integration of heterogeneous data from multiple sources (including exploration adit text records, 2D maps, and 3D coordinates), and visual comparison to label occurrence parameters and determine connectivity. Empirical studies have shown that manually processing thousands of structural surfaces takes months, and classification inconsistencies are high due to subjective judgment. While commercial geotechnical software supports interface simulation and visualization, its lack of intelligent analysis modules means that parameter setting and result correction still rely on manual intervention. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the traditional geological workflow is highly experience-dependent and inefficient. The purpose of the present invention is to provide a hydropower station structural surface analysis method, system and medium based on prompt learning, integrating a multimodal AI architecture with a domain knowledge-driven method, and realizing cross-modal feature fusion of multimodal data such as text, image and space through a multimodal large model and prompt learning technology; at the same time, using prompt learning templates, the engineering specifications are encoded as prior knowledge of the model, thereby enhancing the conformity of the output results with industry standards; in addition, by introducing an artificial feedback iteration mechanism to construct a closed-loop verification system, the transformation of laboratory results into engineering practice is effectively promoted.

[0005] The present invention is achieved through the following technical solutions: This solution provides a hydropower station structural surface analysis method based on prompt learning, including: Collecting multimodal basic data of the hydropower station structure surface and preprocessing the multimodal basic data; the multimodal basic data includes: spatial data, occurrence data, width data and geological feature data; The pre-processed multimodal basic data is input into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to perform cross-modal mapping of the multimodal basic data and introduces a prompt learning mechanism for similarity analysis; Output the classification results and similarity results of the hydropower station structure surface.

[0006] A further optimization scheme is that the multimodal basic data is preprocessed, including the following methods: De-noise the multimodal basic data first; Perform integrity checks based on the denoised multimodal basic data and fill in missing data; Normalize the filled multimodal basic data.

[0007] A further optimization scheme is that the structural surface analysis model uses a shared embedding space to perform cross-modal mapping of multimodal basic data; including the following methods: Multimodal basic data is fused based on a multimodal Transformer structure; the input embedding of the multimodal Transformer structure is: ; ; ; ; Among them, Concat() represents the concatenation operation; it means connecting multiple vectors (or matrices) end to end along a specific dimension to form a larger vector (or matrix); E represents a concatenated integrated embedding vector, E P Represents the position vector, MLP P () represents a multilayer perceptron for processing position data, which converts the input position data pi into an embedding vector ; pi represents the i-th position data of the input; E I Represents an image vector; CNN() represents a convolutional neural network used to transform the input image Convert to embedding vector ;I k Represents the kth image data input; E A Represents attribute vector, MLP a () is a multi-layer perceptron used to transform attribute data (such as text description, etc.) into an embedding vector , a j Represents the jth attribute data of the input, including text, category labels or other non-image, non-position features.

[0008] Modeling the interaction between different modal data based on the multi-head attention mechanism: ; Among them, Attention() represents the attention calculation function, which is used to calculate the attention weight between the query Q, key K and value V, and generate a weighted value vector as output; softmax() represents the normalized exponential function, which is used to convert the dot product of the query Q and key K into the attention weight to ensure that the sum of the weights is 1; T represents the transpose of the matrix; d k Indicates the dimension of the key vector; scaling factor This prevents excessive dot product results from causing the vanishing gradient problem in softmax. Attention(Q, K, V) calculates the similarity between Q and K, generates attention weights, and then sums these weighted sums for V, enabling inter-modal interaction. This mechanism dynamically captures the correlation between data from different modalities.

[0009] A further optimization scheme is that the multimodal Transformer structure is trained with data structure A, and the data structure A is: ; in, represents the alignment loss, which is used to measure the distance between similar samples in the embedding space; Represent the embedding vectors of sample i and sample j respectively; represents the Euclidean distance between the embedding vectors of sample i and sample j, and the calculation formula is: ; Represents the indicator function, which outputs 1 when the labels of sample i and sample j are the same, and 0 otherwise.

[0010] A further optimization scheme is to introduce a prompt learning mechanism to perform similarity analysis; including the following methods: Constructing a prompt learning template; the prompt learning template includes a hydropower station structural surface data input model, a similarity calculation template, a similarity comprehensive calculation template and a judgment template; A prompt learning template is introduced into the structural surface analysis model, and the prompt learning template is used to guide the structural surface analysis model to determine whether any two structural surfaces of a hydropower station are the same based on multimodal basic data after cross-modal mapping.

[0011] A further optimization solution is that the similarity calculation template includes: Spatial similarity calculation template: ; Occurrence similarity calculation template: ; Extended length template: ; in, represents the spatial similarity between hydropower station structure surface i and hydropower station structure surface j; xi, yi, zi represents the spatial coordinates of the hydropower station structure surface i; xj, yj, zj represents the spatial coordinates of the hydropower station structure surface j; αi represents the inclination of the hydropower station structural surface i; αj represents the inclination of the hydropower station structure surface j; βi represents the inclination angle of the hydropower station structure surface i; βj represents the inclination angle of the hydropower station structure surface j; Represents the set of exposed widths of structural surface i in n flat holes.

[0012] A further optimization solution is that the similarity comprehensive calculation template includes: The similarity S(i, j) between hydropower station structural surface i and hydropower station structural surface j is: ; in, Represents the weighting coefficients of spatial similarity, occurrence similarity, width and extension length.

[0013] A further optimization solution is that the judgment template includes: Preset similarity threshold threshold similarity ; when When , it is considered that the hydropower station structural surface i and the hydropower station structural surface j are the same structural surface.

[0014] This solution also provides a hydropower station structural surface analysis system based on prompt learning, which is used to implement the above-mentioned hydropower station structural surface analysis method based on prompt learning. The system includes: Acquisition module, used to collect multi-dimensional basic data of the hydropower station structure surface; A preprocessing module, used for preprocessing multidimensional basic data; An analysis module is used to input the pre-processed multi-dimensional basic data into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to effectively map the multi-dimensional basic data and introduces a prompt learning mechanism for similarity analysis; The output module is used to output the classification results and similarity results of the hydropower station structure surface.

[0015] The present solution also provides a computer-readable medium having a computer program stored thereon, and the computer program is executed by a processor to implement the above-mentioned method for analyzing the structural surface of a hydropower station based on prompt learning.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. The present invention provides a method, system, and medium for analyzing the structural surface of a hydropower station based on prompt learning. The method integrates a multimodal AI architecture with a domain knowledge-driven approach, and achieves cross-modal feature fusion of multimodal data such as text, images, and space through a multimodal large model and prompt learning technology. At the same time, the prompt learning template is used to encode engineering specifications as prior knowledge of the model, thereby enhancing the conformity of the output results with industry standards.

[0017] 2. The proposed method, system, and medium for hydropower station structural surface analysis based on prompt learning address the complex data types and diverse sources in hydropower station structural surface analysis. This method uses a shared embedding space to effectively map multidimensional basic data, integrating the geometric parameters (occurrence, spatial coordinates), physical properties, and engineering constraints of the structural surface to improve the model's accuracy and generalization capabilities in structural surface feature extraction and relationship inference tasks. 3. The present invention provides a method, system, and medium for analyzing the structural surface of a hydropower station based on prompt learning. A prompt learning mechanism is introduced to assist in determining the connectivity of the structural surfaces of a hydropower station. The prompt learning mechanism automatically infers and learns the connectivity between structural surfaces by combining structural surface data (such as spatial coordinates, occurrence, width, etc.) with prior knowledge (such as structural surface width, extension length, etc.). It can not only process complex structural surface data, but also perform intelligent reasoning between multiple structural surfaces, helping to achieve automated structural surface merging and connectivity analysis, thereby greatly improving the efficiency and accuracy of hydropower station structural surface analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings: Figure 1 A flow chart of the hydropower station structural surface analysis method based on prompt learning; Figure 2 Schematic diagram of the hydropower station structural surface analysis system based on prompt learning. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0020] To address the high experience-dependence and low efficiency of traditional geological workflows, this embodiment provides the following: Example 1 This embodiment provides a hydropower station structural surface analysis method based on prompt learning, such as Figure 1 Shown, including: Step 1: Collect multimodal basic data of the hydropower station structure surface and pre-process the multimodal basic data; the multimodal basic data includes: spatial data, occurrence data, width data and geological feature data; During the implementation of the hydropower station structural surface intelligent analysis system, data acquisition and preprocessing are the basis for ensuring the accuracy and reliability of the system. The main goal of this step is to collect relevant hydropower station structural surface data and clean, normalize, and convert it for subsequent modeling and analysis. First, the system needs to collect multimodal data related to the structural surface from the hydropower station survey data, mainly including: Spatial data: spatial coordinate information of each hydropower station structure surface ( x , y , z ), data is collected through high-precision measuring instruments such as total stations, laser scanners, drones and other equipment to ensure data accuracy.

[0021] Developed data: Developed information of each hydropower station's structural surface, including dip α and dip angle β. These data are usually obtained through on-site geological surveys and measurement equipment, using directional instruments or automated equipment for collection.

[0022] Width data: The width data of each structural surface is obtained through measurement tools or numerical simulation during on-site surveys. Different measurement tools and methods may affect the accuracy of the data, so unified standards are required for collection.

[0023] Geological characteristic data: including the physical properties of the structural surface (such as strength, friction angle, etc.), crack data, etc.; these data help to further analyze the stability and possible extension characteristics of the structural surface.

[0024] The collected raw data often contains noise or invalid data. Data cleaning and screening are key steps to ensure the reliability of subsequent analysis results. Therefore, the multimodal basic data is preprocessed. The specific methods include: S11, first denoise the multimodal basic data; detect and delete data points that obviously do not conform to physical laws or measurement errors, such as measurement errors, invalid coordinate values ​​or unreasonable occurrence data (for example, inclination angles exceeding the normal range).

[0025] S12, perform integrity check based on the denoised multimodal basic data and fill in missing data; ensure that the spatial coordinates, occurrence and width information of each structural surface data are complete, and missing data need to be marked or filled; then filter and deduplicate the data to avoid interference from duplicate data, especially duplicate records of the same structural surface, and screen out independent structural surface data through spatial clustering or similarity measurement methods.

[0026] S13, normalize the filled multimodal basic data. In order to facilitate the subsequent machine learning model training and inference, the collected data needs to be normalized and converted; standardize data units: unify data units to ensure that all spatial coordinates, angles, widths and other data use the same measurement units (for example: meters, degrees, etc.).

[0027] Coordinate normalization: Normalize spatial coordinates as needed so that the model can process data of different scales and improve training efficiency and accuracy.

[0028] Calculation of structural surface width and extension length: Based on the collected structural surface width information, the structural surface extension length is derived according to the GB50287 standard. Combining the structural surface width and extension length, the structural surface grade information (such as Grade I, Grade II, etc.) is generated, providing labeled data for subsequent model training.

[0029] Step 2: Input the pre-processed multimodal basic data into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to perform cross-modal mapping of the multimodal basic data, and introduces a prompt learning mechanism for similarity analysis; In step 2, the structural surface analysis model uses a shared embedding space to perform cross-modal mapping of the multimodal basic data; In the multimodal data fusion process, this solution adopts a multimodal large model framework to achieve effective mapping of each modal data through a shared embedding space. Specifically, each modal input data is processed by its own network and effectively aligned in the shared space to ensure that different modal data can interact and fuse in a unified space. Specific implementation methods include: Multimodal basic data is fused based on a multimodal Transformer structure; the input embedding of the multimodal Transformer structure is: ; ; ; ; Among them, Concat() represents the concatenation operation; it means connecting multiple vectors (or matrices) end to end along a specific dimension to form a larger vector (or matrix); E represents a concatenated integrated embedding vector, E P Represents the position vector, MLP P () represents a multilayer perceptron for processing position data, which converts the input position data pi into an embedding vector ;pi represents the i-th position data of the input; E I Represents an image vector; CNN() represents a convolutional neural network used to transform the input image Convert to embedding vector ;I k Represents the kth image data input; E A Represents attribute vector, MLP a () is a multi-layer perceptron used to transform attribute data (such as text description, etc.) into an embedding vector , a j Represents the jth attribute data of the input, including text, category labels, or other non-image, non-position features. The interaction relationship between different modal data is modeled based on the multi-head attention mechanism: ; Among them, Attention() represents the attention calculation function, which is used to calculate the attention weight between the query Q, key K and value V, and generate a weighted value vector as output; softmax() represents the normalized exponential function, which is used to convert the dot product of the query Q and key K into the attention weight to ensure that the sum of the weights is 1; T represents the transpose of the matrix; d k Indicates the dimension of the key vector; scaling factor This prevents excessive dot product results from causing the vanishing gradient problem in softmax. Attention(Q, K, V) calculates the similarity between Q and K, generates attention weights, and then sums these weighted sums for V, enabling inter-modal interaction. This mechanism dynamically captures the correlation between data from different modalities.

[0030] The multimodal Transformer structure is trained with data structure A, which is: ; in, represents the alignment loss, which is used to measure the distance between similar samples in the embedding space; Represent the embedding vectors of sample i and sample j respectively; represents the Euclidean distance between the embedding vectors of sample i and sample j, and the calculation formula is: ; Represents the indicator function, which outputs 1 when the labels of sample i and sample j are the same, and 0 otherwise.

[0031] During the multimodal data fusion phase, training is performed using data structure A. The goal is to enhance the model's representational capabilities by aligning the feature vectors of data from different modalities. During fine-tuning, mixed-precision training is used to maintain training stability and significantly improve training efficiency.

[0032] The complexity of hydropower station structural surface data makes it difficult to analyze and identify the same structural surface using traditional methods. In particular, it is not intuitive to determine whether the structural surfaces between different adit holes are the same structural surface. In this context, the introduction of a prompt learning template can provide the model with an intelligent reasoning method for structural surfaces. The introduction of the prompt learning mechanism for similarity analysis includes the following methods: S21, constructing a prompt learning template; the prompt learning template includes a hydropower station structural surface data input model, a similarity calculation template, a similarity comprehensive calculation template and a judgment template; the hydropower station structural surface data input model specifically includes basic information of the hydropower station structural surface: the spatial position of the hydropower station structural surface i: , occurrence: ,width:( ), extended length: ; ...the spatial position of structural surface j: , occurrence: ,width: , extended length: .

[0033] S22, introducing a prompt learning template into the structural surface analysis model, wherein the prompt learning template is used to guide the structural surface analysis model to determine whether any two structural surfaces of the hydropower station are the same based on the multimodal basic data after cross-modal mapping.

[0034] The similarity calculation template includes: Spatial similarity calculation template: ; Occurrence similarity calculation template: ; Extended length template: ; in, represents the spatial similarity between hydropower station structure surface i and hydropower station structure surface j; xi, yi, zi represents the spatial coordinates of the hydropower station structure surface i; xj, yj, zj represents the spatial coordinates of the hydropower station structure surface j; αi represents the inclination of the hydropower station structural surface i;αj represents the inclination of the hydropower station structure surface j; βi represents the inclination angle of the hydropower station structure surface i; βj represents the inclination angle of the hydropower station structure surface j; Represents the set of exposed widths of structural surface i in n flat holes.

[0035] According to GB50287 specification and engineering geological analysis requirements, function Used to expose the width of the structural surface Mapped to its possible extension length .function is a piecewise function based on the specification, which converts the width of the structural surface into Classify and correspond to the extension length range. The specific form is as follows: ; Since the exposure width of the same structural surface in different horizontal holes may be inconsistent, it is necessary to comprehensively estimate the overall extension length by combining the widths of multiple exposure points. In this case, the function needs to be expanded to obtain: ; The maximum value is taken because a larger width corresponds to a longer potential extension length.

[0036] When the spatial similarity and attitude similarity of the hydropower station structural surface i and the hydropower station structural surface j meet the preset spatial similarity threshold and the preset attitude similarity threshold, it is considered that the two structural surfaces may be the same in space. At the same time, if the extension lengths of the hydropower station structural surface i and the hydropower station structural surface j are close, it is considered that the two structural surfaces may be the same in space.

[0037] The similarity comprehensive calculation template includes: The similarity S(i, j) between hydropower station structural surface i and hydropower station structural surface j is: ; in, represent the weighting coefficients of spatial similarity, occurrence similarity, width and extension length respectively; L i represents the extension length of the hydropower station structure surface i; L j Represents the extension length of the hydropower station structure surface j.

[0038] The judgment template includes: Preset similarity threshold threshold similarity ; when When , it is considered that the hydropower station structural surface i and the hydropower station structural surface j are the same structural surface.

[0039] This embodiment integrates a multimodal AI architecture with a domain knowledge-driven approach. Using a large multimodal model and prompt learning technology, it achieves cross-modal feature fusion of multimodal data, including text, images, and spatial data. Furthermore, using prompt learning templates, engineering specifications are encoded into the model's prior knowledge, thereby enhancing the alignment of output results with industry standards. To address the complex data types and diverse sources in hydropower station structural surface analysis, a shared embedding space is used to effectively map multidimensional basic data. This integrates the geometric parameters (disposition, spatial coordinates), physical properties, and engineering constraints of the structural surface, improving the model's accuracy and generalization in structural surface feature extraction and relationship inference. This solution also introduces a prompt learning mechanism to assist in determining the connectivity of hydropower station structural surfaces. This prompt learning mechanism automatically infers and learns the connectivity between structural surfaces by combining structural surface data (such as spatial coordinates, disposition, and width) with prior knowledge (such as structural surface width and extension length). The core concept of the prompt learning mechanism is to use the model to infer the connectivity between different structural surfaces based on given structural surface data. Specifically, the model first determines whether two structural surfaces are likely to belong to the same surface based on their occurrence information. It then uses their spatial position and surface width to deduce their possible extension lengths. Finally, based on the relationship between extension lengths and spatial distances, it determines whether two structural surfaces are connected. This model not only processes complex surface data but also performs intelligent reasoning across multiple surfaces, enabling automated surface merging and connectivity analysis, significantly improving the efficiency and accuracy of hydropower station surface analysis.

[0040] Example 2 This embodiment provides a hydropower station structural surface analysis system based on prompt learning, such as Figure 2 As shown, for implementing the method for analyzing the structural surface of a hydropower station based on prompt learning described in Example 1, the system includes: Acquisition module, used to collect multi-dimensional basic data of the hydropower station structure surface; A preprocessing module, used for preprocessing multidimensional basic data; An analysis module is used to input the pre-processed multi-dimensional basic data into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to effectively map the multi-dimensional basic data and introduces a prompt learning mechanism for similarity analysis; The output module is used to output the classification results and similarity results of the hydropower station structure surface.

[0041] Example 3 This embodiment provides a computer-readable medium having a computer program stored thereon. The computer program is executed by a processor to implement the method for analyzing the structural surface of a hydropower station based on prompt learning as described in Example 1. Specifically, the following steps are performed: Step 1: Collect multimodal basic data of the hydropower station structure surface and pre-process the multimodal basic data; the multimodal basic data includes: spatial data, occurrence data, width data and geological feature data; Step 2: Input the pre-processed multimodal basic data into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to perform cross-modal mapping of the multimodal basic data and introduces a prompt learning mechanism for similarity analysis; Step 3: Output the classification results and similarity results of the hydropower station structure surface.

[0042] After completing the structure surface data analysis based on prompt learning, the next step is to integrate the analysis method into the intelligent analysis system and verify the accuracy and practicality of the system. Specifically, the accuracy and consistency of the analysis results output by the model can be verified through manual review or reference to existing standard data sets. Professional geologists or experts can be used for evaluation to ensure high consistency between the system results and the manual analysis results.

[0043] Performance evaluation: The system was deployed in a typical hydropower project for iterative closed-loop verification, using the following performance indicators for evaluation: Accuracy Acc: ; Recall R: ; False positive rate FPR: ; Average prediction time ; Among them, TP represents true positive examples, that is, samples correctly predicted by the model as positive; FP represents false positive examples, that is, samples incorrectly predicted by the model as positive; TN represents true negative examples, that is, samples correctly predicted by the model as negative; FN represents false negative examples, that is, samples incorrectly predicted by the model as negative.

[0044] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hydropower station structural surface analysis method based on prompt learning, characterized in that: include: Collect multimodal basic data of the hydropower station structure surface and preprocess the multimodal basic data; The multimodal basic data includes: spatial data, occurrence data, width data and geological feature data; The pre-processed multimodal basic data is input into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to perform cross-modal mapping of the multimodal basic data and introduces a prompt learning mechanism for similarity analysis; Output the classification results and similarity results of the hydropower station structure surface.

2. The hydropower station structural surface analysis method based on prompt learning according to claim 1 is characterized in that: The preprocessing of the multimodal basic data includes the following methods: De-noise the multimodal basic data first; Perform integrity checks based on the denoised multimodal basic data and fill in missing data; Normalize the filled multimodal basic data.

3. The hydropower station structural surface analysis method based on prompt learning according to claim 1 is characterized in that: The structural surface analysis model uses a shared embedding space to perform cross-modal mapping of multimodal basic data; including method: Multimodal basic data is fused based on a multimodal Transformer structure; the input embedding of the multimodal Transformer structure is: ； ； ； ； Among them, Concat() represents the concatenation operation; E represents a concatenated comprehensive embedding vector, E P Represents the position vector, MLP P () represents a multi-layer perceptron for processing position data, which converts the input position data pi into an embedding vector ; pi represents the i-th position data of the input; E I Represents an image vector; CNN() represents a convolutional neural network used to transform the input image Convert to embedding vector ;I k Represents the kth image data input; E A Represents attribute vector, MLP a () represents a multi-layer perceptron, which is used to transform attribute data Convert to embedding vector , a j Represents the jth attribute data of the input, including text, category labels or other non-image, non-position features. Modeling the interaction between different modal data based on the multi-head attention mechanism: ； Among them, Attention() represents the attention calculation function, which is used to calculate the attention weight between the query Q, key K and value V, and generate a weighted value vector as output; softmax() represents the normalized exponential function, which is used to convert the dot product of the query Q and key K into the attention weight to ensure that the sum of the weights is 1; T represents the transpose of the matrix; d k Indicates the dimension of the key vector.

4. The hydropower station structural surface analysis method based on prompt learning according to claim 3 is characterized in that: The multimodal Transformer structure is trained with data structure A, which is: ； in, represents the alignment loss, which is used to measure the distance between similar samples in the embedding space; Represent the embedding vectors of sample i and sample j respectively; represents the Euclidean distance between the embedding vectors of sample i and sample j.

5. The method for analyzing the structural surface of a hydropower station based on prompt learning according to claim 1, characterized in that: The introduction of the prompt learning mechanism to perform similarity analysis; Includes methods: Constructing a prompt learning template; the prompt learning template includes a hydropower station structural surface data input model, a similarity calculation template, a similarity comprehensive calculation template and a judgment template; A prompt learning template is introduced into the structural surface analysis model, and the prompt learning template is used to guide the structural surface analysis model to determine whether any two structural surfaces of a hydropower station are the same based on multimodal basic data after cross-modal mapping.

6. The method for analyzing the structural surface of a hydropower station based on prompt learning according to claim 5 is characterized in that: The similarity calculation template includes: Spatial similarity calculation template: ; Occurrence similarity calculation template: ; Extended length template: ; in, represents the spatial similarity between hydropower station structure surface i and hydropower station structure surface j; xi, yi, zi represents the spatial coordinates of the hydropower station structure surface i; xj,yj,zj represents the spatial coordinates of the hydropower station structure surface j; αi represents the inclination of the hydropower station structural surface i; αj represents the inclination of the hydropower station structure surface j; βi represents the inclination angle of the hydropower station structure surface i; βj represents the inclination angle of the hydropower station structure surface j; Represents the set of exposed widths of structural surface i in n flat holes.

7. The method for analyzing the structural surface of a hydropower station based on prompt learning according to claim 6 is characterized in that: The similarity comprehensive calculation template includes: The similarity S(i, j) between hydropower station structural surface i and hydropower station structural surface j is: ； in, Represents the weighting coefficients of spatial similarity, occurrence similarity, width and extension length.

8. The method for analyzing the structural surface of a hydropower station based on prompt learning according to claim 6, characterized in that: The judgment template includes: Preset similarity threshold threshold similarity ; when When , it is considered that the hydropower station structural surface i and the hydropower station structural surface j are the same structural surface.

9. The hydropower station structural surface analysis system based on prompt learning is characterized by: The system for implementing the method for analyzing the structural surface of a hydropower station based on prompt learning according to any one of claims 1 to 8 comprises: Acquisition module, used to collect multi-dimensional basic data of the hydropower station structure surface; A preprocessing module, used for preprocessing multidimensional basic data; An analysis module is used to input the pre-processed multi-dimensional basic data into the constructed structural surface analysis model for classification and similarity analysis; the structural surface analysis model uses a shared embedding space to effectively map the multi-dimensional basic data and introduces a prompt learning mechanism for similarity analysis; The output module is used to output the classification results and similarity results of the hydropower station structure surface.

10. A computer-readable medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the hydropower station structural surface analysis method based on prompt learning as described in any one of claims 1 to 8.

Images