Research and development-knowledge property right disconnection automatic detection method and system based on multi-source heterogeneous data space-time correlation

Abstract

The application relates to the technical field of data processing, in particular to a research and development-knowledge property right disconnection automatic detection method and system based on multi-source heterogeneous data space-time correlation, which effectively solves the cross-system barriers of multi-source heterogeneous data in time anchor points and technical semantics by constructing a unified data collection and entity alignment mechanism, and realizes automatic correlation and standardized grouping of research and development projects and knowledge property right data. On this basis, a three-dimensional comprehensive scoring model is innovatively introduced to quantitatively evaluate the project-patent correlation strength from three dimensions of time sequence consistency, personnel consistency and technical correlation, accurately identify the hidden disconnection risks caused by the misalignment of project cycle and patent examination cycle, the drift of technical route or the disconnection of inventors and researchers. At the same time, the output prediction model is combined to regress and estimate the expected number of patents of the project, and the actual correlated output is compared to realize the abnormal detection of the research and development input conversion efficiency.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to an automatic detection method and system for R&D-intellectual property disconnect based on the spatiotemporal correlation of multi-source heterogeneous data. Background Technology

[0002] With the deepening of the Science and Technology Innovation Board and the registration-based IPO system reform, regulatory agencies are increasingly stringent in their review of companies' "hard technology" attributes. The inherent logical correlation between R&D investment and intellectual property output has become a core focus in IPO inquiries. Currently, during the listing preparation process, companies typically rely on manual review of project completion reports, patent certificates, and R&D expense subsidiary ledgers to verify the correspondence between intellectual property and R&D projects. However, companies' R&D progress data is scattered across project management systems such as PLM (Product Lifecycle Management) and ERP (Enterprise Resource Planning), while patent information is stored in intellectual property management platforms and external patent databases. These two systems not only exhibit significant structural heterogeneity (i.e., the coexistence of structured project nodes and unstructured technical texts), but also lack unified time anchors (such as project initiation date, completion date, and priority date) and cross-system technical semantic mapping mechanisms.

[0003] Existing detection methods are mostly post-event sampling or comparisons based on simple time sorting. They cannot effectively address the hidden risks caused by misalignment between project cycles and patent examination cycles, drifting of technical routes, or separation of R&D personnel from inventors. This can easily lead to companies overlooking major flaws before filing their applications, resulting in high costs for on-the-spot remediation. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide an automatic detection method and system for R&D-intellectual property disconnect based on the spatiotemporal correlation of multi-source heterogeneous data, so as to solve the problems in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data in this application includes:

[0007] The system acquires R&D project data for multiple R&D projects and intellectual property data for multiple intellectual property rights from the target company. The R&D project data includes project initiation time, completion time, project description text, list of R&D personnel, and investment amount. The intellectual property data includes application date, list of inventors, abstract text, and IPC classification.

[0008] The R&D project data and the intellectual property data are matched based on semantic similarity and relationship graph to obtain multiple project data groups, wherein each project data group contains the R&D project data of one R&D project and the associated intellectual property data;

[0009] Calculate the overall correlation score of multiple project data groups, and perform correlation decoupling detection based on the overall correlation score, wherein the overall correlation score includes the generation of time series consistency score, personnel consistency score, and technology relevance score;

[0010] The system predicts the intellectual property output of each R&D project based on a pre-built output prediction model, and detects output disconnect based on the predicted intellectual property output and the associated intellectual property data.

[0011] In one embodiment of this application, the R&D project data and the intellectual property data are matched based on semantic similarity and relationship graphs to obtain multiple project data groups, including:

[0012] Based on the fine-tuned language model, the R&D project data and the intellectual property data are vectorized to obtain the first semantic feature vector and the second semantic feature vector.

[0013] Calculate the cosine similarity between the first semantic feature vector and the second semantic feature vector;

[0014] For each first semantic feature vector, a second semantic feature vector with a cosine similarity greater than or equal to a preset threshold is used as the associated second semantic feature vector; and the associated intellectual property data of each R&D project is determined based on the association relationship of the semantic feature vectors.

[0015] Based on a pre-built relational graph model, intellectual property data that is not associated with R&D projects is matched with R&D projects, and the intellectual property data that matches the R&D projects is used as associated intellectual property data.

[0016] Project data groups are constructed based on the R&D project data and related intellectual property data of each R&D project.

[0017] In one embodiment of this application, the method for constructing the relational graph model includes:

[0018] S1, define node features and relation edges, wherein the node features include project text semantic vectors and intellectual property text semantic vectors, and the relation edges include personnel intersection, time window, IPC classification and text similarity;

[0019] S2, obtain a training sample set, wherein the training sample set includes multiple training samples, and each training sample includes node feature samples and correlation labels;

[0020] S3, Substitute the training samples into the relational graph convolutional network, and propagate them in the relational graph convolutional network to obtain the project node embedding vector and the intellectual property node embedding vector;

[0021] S4, based on the predicted correlation between the project node embedding vector and the intellectual property node embedding vector mapped by the scoring layer, wherein the scoring layer is a multilayer perceptron or a bilinear transform function;

[0022] S5, calculate the loss between the predicted correlation degree and the correlation degree label based on the pre-constructed loss function, and adjust the internal parameters of the relationship graph convolutional network and the scoring layer based on the backpropagation of the loss;

[0023] S6. Repeat steps S3-S5 until training is complete, resulting in a relation graph model that includes a relation graph convolutional network and a scoring layer.

[0024] In one embodiment of this application, matching intellectual property data not associated with R&D projects with R&D projects based on a pre-built relationship graph model includes:

[0025] A graph input vector is constructed based on the second semantic feature vector of intellectual property data that is not associated with R&D projects and the first semantic feature vector of multiple R&D project data, respectively.

[0026] The graph input vector is input into a pre-constructed relational graph model to obtain a correlation score.

[0027] For the same intellectual property data, if there is a target R&D project data with the highest relevance score that is greater than or equal to a preset relevance threshold, the intellectual property data is determined to match the target R&D project data.

[0028] In one embodiment of this application, calculating the composite score of multiple project data groups includes:

[0029] Based on the R&D project data and intellectual property data in the aforementioned project data group, a time-series consistency score is calculated between the R&D project and each intellectual property. Personnel consistency score Technology relevance score The time-series consistency score Personnel consistency score Technology relevance score The mathematical expressions are as follows:

[0030]

[0031]

[0032]

[0033] In the formula, The application date of the intellectual property right. The project's completion date, This is a gathering of project R&D personnel. It is a collection of inventors of intellectual property rights. For project text semantic vectors, For the semantic vector of the intellectual property abstract text;

[0034] Based on the time consistency score Personnel consistency score Technology relevance score Calculate the overall correlation degree The comprehensive correlation degree The mathematical expression is:

[0035]

[0036] In the formula, As the first weight, As the second weight, It is the third weight;

[0037] Comprehensive correlation of multiple intellectual property rights A weighted summation is performed to obtain the overall relevance score of the R&D project, which includes the overall relevance of multiple intellectual property rights. The weights are based on the specifications in the management system.

[0038] In one embodiment of this application, the detection of association disconnect based on the comprehensive association score includes:

[0039] The correlation comprehensive score is compared with a preset comprehensive score threshold. If the correlation comprehensive score is greater than or equal to the comprehensive score threshold, it is determined that there is no disconnection of intellectual property rights in the R&D project; otherwise, it is determined that there is a disconnection of intellectual property rights in the R&D project.

[0040] In one embodiment of this application, the method for constructing the output prediction model includes:

[0041] Obtain the investment amount, number of R&D personnel, time span, IPC classification number, and number of intellectual property rights strongly associated with the historical R&D projects for multiple historical R&D projects;

[0042] Sample data is constructed based on the investment amount, number of R&D personnel, time span, and IPC classification number of historical R&D projects, and training data is constructed by using the number of intellectual property rights that are strongly correlated with historical R&D projects as labels.

[0043] The gradient boosting decision tree model is trained based on the training data and a pre-built loss function to obtain the output prediction model, wherein the mathematical expression of the loss function is:

[0044]

[0045]

[0046] In the formula, For the overall optimization objective, The total number of training samples, For training sample index, For the first Labels of each training sample For the first The model prediction value for each training sample. To predict losses, For the index of the tree, For the total number of trees in the regression tree, The complexity penalty is applied to a single tree. is the exponential parameter of the Tweedie distribution.

[0047] In one embodiment of this application, predicting the intellectual property output of each R&D project based on a pre-built output prediction model includes:

[0048] The input vector is obtained by extracting the investment amount, number of R&D personnel, time span, and IPC classification number from the R&D project data.

[0049] The input vector is fed into a pre-built output prediction model to predict the quantity of intellectual property output from the R&D project.

[0050] In one embodiment of this application, output disconnect detection is performed based on the predicted quantity of intellectual property output and the associated intellectual property data, including:

[0051] Extract the number of related intellectual property rights from the related intellectual property data;

[0052] The predicted quantity of intellectual property output is compared with the quantity of associated intellectual property. If the predicted quantity of intellectual property output exceeds the quantity of associated intellectual property, and the excess is greater than a set threshold, it is determined that there is a disconnect in output.

[0053] This application also provides an automatic detection system for the disconnect between R&D and intellectual property based on the spatiotemporal correlation of multi-source heterogeneous data, including:

[0054] The acquisition module is used to acquire R&D project data for multiple R&D projects and intellectual property data for multiple intellectual property rights of the target enterprise. The R&D project data includes project initiation time, completion time, project description text, list of R&D personnel, and investment amount; the intellectual property data includes application date, list of inventors, abstract text, and IPC classification.

[0055] The matching module is used to perform semantic similarity and relationship graph matching on the R&D project data and the intellectual property data to obtain multiple project data groups, wherein each project data group contains the R&D project data of one R&D project and the associated intellectual property data.

[0056] The correlation decoupling detection module is used to construct a bipartite time series graph of multiple data groups, calculate a three-dimensional correlation comprehensive score based on the bipartite time series graph, and perform correlation decoupling detection based on the correlation comprehensive score. The correlation comprehensive score includes the generation of time series consistency score, personnel consistency score, and technology relevance score.

[0057] The output disconnect detection module is used to predict the number of intellectual property outputs for each R&D project based on a pre-built output prediction model, and to perform output disconnect detection based on the predicted number of intellectual property outputs and the associated intellectual property data.

[0058] The beneficial effects of this application are as follows: The automatic detection method and system for R&D-intellectual property disconnect based on the spatiotemporal correlation of multi-source heterogeneous data effectively solves the cross-system barriers of multi-source heterogeneous data in terms of time anchors and technical semantics by constructing a unified data collection and entity alignment mechanism, realizing the automatic correlation and standardized grouping of R&D projects and intellectual property data. Based on this, a three-dimensional comprehensive scoring model is innovatively introduced to quantitatively evaluate the project-patent correlation strength from three dimensions: temporal consistency, personnel consistency, and technical relevance, accurately identifying hidden disconnect risks caused by misalignment between project cycles and patent examination cycles, drift in technical routes, or decoupling between inventors and R&D personnel. Simultaneously, by combining an output prediction model to regress and estimate the expected number of patents for the project and comparing it with the actual associated output, anomalies in the efficiency of R&D investment conversion are detected. This application upgrades the traditional post-event sampling inspection that relies on human experience to a systematic, multi-dimensional proactive early warning and quantitative assessment. It can effectively help companies comprehensively investigate logical flaws between R&D and intellectual property before submitting their IPO applications, significantly reduce the inquiry risks and on-the-spot remedial costs caused by insufficient relevance of core assets, and greatly improve the efficiency and compliance quality of listing review. Attached Figure Description

[0059] The present application will be further described below with reference to the accompanying drawings and embodiments:

[0060] Figure 1 This is a flowchart illustrating the overall concept of an automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data in one embodiment of this application.

[0061] Figure 2This is a flowchart illustrating an embodiment of the automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data in this application;

[0062] Figure 3 This is a schematic diagram of the tone alignment matching process in one embodiment of this application;

[0063] Figure 4 This is an example of the disconnection detection result of the automatic detection method for R&D-intellectual property disconnection based on spatiotemporal correlation of multi-source heterogeneous data in this application;

[0064] Figure 5 This is a logical structure diagram of an automatic detection system for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data, as shown in one embodiment of this application. Detailed Implementation

[0065] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0066] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the layers related to this application and are not drawn according to the actual number, shape and size of the layers in the actual implementation. In the actual implementation, the form, number and proportion of each layer can be arbitrarily changed, and the layer layout may also be more complex.

[0067] Numerous details are explored in the following description to provide a more thorough explanation of embodiments of this application; however, it will be apparent to those skilled in the art that embodiments of this application may be practiced without these specific details.

[0068] Figure 1 This is a flowchart illustrating the overall concept of an automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data in one embodiment of this application. Figure 1 As shown, the overall idea of ​​this application is as follows:

[0069] Data Acquisition and Standardization: Collect R&D project data (project initiation / completion time, description text, personnel list, investment amount) and intellectual property data (application date, inventor list, abstract, IPC classification) from multiple systems such as PLM, ERP and intellectual property management platforms, and perform standardization operations such as date alignment, personnel disambiguation, text cleaning and missing data handling to eliminate data barriers between heterogeneous systems.

[0070] Project-Patent Dual-Path Association Matching: A dual-path matching strategy combining semantic similarity and relationship graph model is adopted. First, the cosine similarity between the project description and the patent abstract is calculated using a fine-tuned BERT model to make explicit associations. Then, the R-GCN relationship graph model is introduced to mine implicit associations by using relationship edges such as personnel overlap, time proximity, and IPC classification, so as to complete isolated patents to the corresponding projects and build structured project data groups.

[0071] 3D correlation disconnect detection: For each project data group, calculate the temporal consistency, personnel consistency and technical relevance scores between the R&D project and related patents, weight and merge them into a comprehensive correlation score, and then aggregate it into a project-level score by the patent importance weight specified by the management system. Compare it with the preset threshold to determine whether there is a correlation disconnect.

[0072] Output Disconnection Warning: Based on XGBoost, an output prediction model is built. Using historical project investment amount, number of personnel, cycle and IPC as input, the expected number of patent outputs of the current project is predicted and compared with the actual number of related patents. When the actual output is significantly lower than expected, an output disconnection warning is triggered.

[0073] Figure 2 This is a flowchart of an automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data in one embodiment of this application, as shown below. Figure 2 As shown, the automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data in this application mainly includes the following steps:

[0074] S110, Obtain R&D project data for multiple R&D projects of the target company and intellectual property data for multiple intellectual property rights. The R&D project data includes project initiation time, completion time, project description text, list of R&D personnel, and investment amount; the intellectual property data includes application date, list of inventors, abstract text, and IPC classification.

[0075] The R&D project data comes from the company's internal PLM (Product Lifecycle Management), ERP (Enterprise Resource Planning), and financial / human resource management systems; the intellectual property data comes from the company's intellectual property management platform and is connected to the State Intellectual Property Office (CNIPA) or commercial patent databases for cross-verification and supplementation of legal status and information.

[0076] The data encompasses both structured fields and unstructured text. R&D project data includes numeric / date-based structured fields such as project initiation / completion dates and investment amounts, as well as project description text; intellectual property data includes structured tags such as application dates and IPC classification numbers, along with a list of inventors and abstract text.

[0077] After collecting the data, the system performs the following standardized operations:

[0078] (1) Date alignment: Convert the project initiation date, project completion date, and application date into Unix timestamps;

[0079] (2) Personnel disambiguation: Merge names with the same author and standardize aliases for the list of R&D personnel and inventors;

[0080] (3) Text cleaning: The project description and summary text are segmented, stop words are removed, and a custom domain dictionary is added;

[0081] (4) Missing data handling: For samples with missing key fields, rule filling or label removal are used to ensure the integrity and consistency of the input data, laying the foundation for subsequent entity alignment and model calculation.

[0082] S120, perform semantic similarity and relationship graph matching on the R&D project data and the intellectual property data to obtain multiple project data groups, wherein each project data group contains the R&D project data of one R&D project and the associated intellectual property data.

[0083] R&D project data is stored in PLM / ERP systems, while intellectual property data is distributed in IP management platforms and external patent databases. The two are not only physically isolated, but their textual descriptions also have a significant “jargon-term” gap (such as the difference between project codes and standard technical terms).

[0084] This step employs a dual-path matching strategy: first, it utilizes the BERT semantic similarity model to measure the textual relevance between the project description and the patent abstract, addressing the issue of heterogeneous technical expressions; second, it introduces the R-GCN relationship graph model, leveraging multi-dimensional relationship edges such as personnel overlap, temporal proximity, and IPC classification to uncover implicit connections that are difficult to express in text. By combining the explicit results of semantic matching with the completion results of relationship graph reasoning, the system automatically associates each R&D project with its corresponding set of intellectual property data, constructing a structured "project-patent" data set. This provides a precise data foundation for subsequent three-dimensional comprehensive scoring and disjoint risk detection for each data set, effectively avoiding omissions and biases caused by manual association.

[0085] Figure 3 This is a schematic diagram of the tone alignment matching process in one embodiment of this application, as shown below. Figure 3As shown, the specific matching process includes:

[0086] S121, Based on the fine-tuned language model, the R&D project data and the intellectual property data are vectorized to obtain the first semantic feature vector and the second semantic feature vector.

[0087] Select a pre-trained semantic model (such as BERT, RoBERTa, or Patent-BERT) and fine-tune it using historically labeled "project-patent" association data to adapt it to the technical text expression habits of the company's specific domain. After fine-tuning, input the project description text of the R&D project into the model, and it will output a fixed-dimensional first semantic feature vector. Input the "abstract text" of intellectual property into the model, and output the second semantic feature vector. .

[0088] S122, Calculate the cosine similarity between the first semantic feature vector and the second semantic feature vector;

[0089] For each pair of project-patent vectors Calculate the cosine similarity:

[0090]

[0091] Cosine similarity measures the directional consistency between two vectors. It is insensitive to the magnitude of the vectors and therefore unaffected by differences in text length. Its value range is [-1, 1], and after normalization of actual text vectors, it usually falls within the [0, 1] interval. The closer the value is to 1, the more semantically relevant the project description is to the patent abstract.

[0092] S123, for each first semantic feature vector, a second semantic feature vector with a cosine similarity greater than or equal to a preset threshold is used as the associated second semantic feature vector; and the associated intellectual property data of each R&D project is determined based on the association relationship of the semantic feature vectors;

[0093] Set preset threshold (Default 0.7, can be dynamically adjusted according to enterprise data distribution). For each project Filter out all patents This serves as its associated intellectual property data. A patent can be matched with multiple projects simultaneously (many-to-many relationship).

[0094] S124, Based on a pre-built relational graph model, match intellectual property data that is not associated with R&D projects with R&D projects, and use the intellectual property data that matches the R&D projects as associated intellectual property data;

[0095] The matching process based on the relational graph model includes:

[0096] S124-1, Construct a graph input vector based on the second semantic feature vector of intellectual property data not associated with R&D projects and the first semantic feature vectors of multiple R&D project data respectively;

[0097] For isolated patents in S123 that are not associated with any project, their second semantic feature vectors are respectively... With the first semantic feature vector of each project The vectors are concatenated to form the graph input vector. , as input samples for the relational graph model.

[0098] S124-2, Input the graph input vector into the pre-constructed relational graph model to obtain the correlation score;

[0099] Input the graph as a vector The input is fed into a pre-trained R-GCN model, which outputs a correlation score between 0 and 1, representing the probability that the patent belongs to that project under that pair of combinations. R-GCN uses multi-layer graph convolution operations to message-pass and aggregate the semantic vector of each node with its neighbor information in the heterogeneous graph (other nodes connected by relational edges such as people, time, and IPC), thereby generating node embeddings that integrate multi-dimensional heterogeneous relationships. The scoring layer calculates the correlation probability based on the node embeddings, enabling the discovery of implicit correlations that pure text matching cannot find.

[0100] S124-3, For the same intellectual property data, when there is target R&D project data with the highest correlation score that is greater than or equal to the preset correlation threshold, it is determined that the intellectual property data matches the target R&D project data.

[0101] This application employs a dual judgment logic of highest score + threshold to ensure that the matching results have both relative optimality (taking the strongest association) and absolute confidence (only those exceeding the threshold are adopted), avoiding misjudging weak signals as valid associations and ensuring the accuracy of data grouping.

[0102] S125, construct project data groups based on the R&D project data and related intellectual property data of each R&D project.

[0103] Finally, each research and development project The original data (project initiation / completion time, description text, personnel list, investment amount) is merged and packaged with all related patent data (application date, inventor list, abstract, IPC classification) matched in steps S123 and S124 to form a structured project data group. :

[0104]

[0105] The project is stored in the project-patent association database using the project's primary key as an index, for subsequent use by the three-dimensional comprehensive scoring and discontinuity detection modules.

[0106] The above methods rely on pre-built relationship graph models, which are constructed using the following methods:

[0107] S1, define node features and relation edges, wherein the node features include project text semantic vectors and intellectual property text semantic vectors, and the relation edges include personnel intersection, time window, IPC classification and text similarity;

[0108] The project description text of each R&D project and the summary text of each intellectual property are respectively input into the fine-tuned BERT / Doc2Vec model, and the output is a dense vector of fixed dimensions (such as 128-dimensional or 768-dimensional), which serves as the initial semantic feature of the node.

[0109] In addition, four types of edges are established between project nodes and intellectual property nodes:

[0110] (1) Personnel overlap: If at least one member of the project team overlaps with the inventor of the patent, an edge is established.

[0111] (2) Time window: If the patent application date falls within 6 months before or after the project cycle (from the project initiation date to the project completion date), then an edge is established.

[0112] (3) IPC classification: If the IPC main classification number of the patent matches the label of the technical department to which the project belongs, then an edge is established.

[0113] (4) Text similarity: If the BERT cosine similarity between the project description and the patent abstract is >0.7, then an edge is established (as an explicit strong signal).

[0114] S2, obtain a training sample set, wherein the training sample set includes multiple training samples, and each training sample includes node feature samples and correlation labels;

[0115] Multiple training samples include positive and negative samples. Positive samples are extracted from historical enterprise ownership data, focusing on project-patent pairs with known real connections (e.g., those confirmed by manual audit or with high confidence (BERT similarity > 0.9)). These pairs are labeled with a correlation tag y = 1. Furthermore, a random negative sampling strategy is employed, matching each positive sample with k negative samples (typically k = 3~5). Patents unrelated to this project are randomly selected and labeled y = 0.

[0116] Each training sample is a triplet. ,in, Indexing project nodes, For patent node indexing, This is a relevance label.

[0117] S3, Substitute the training samples into the relational graph convolutional network, and propagate them in the relational graph convolutional network to obtain the project node embedding vector and the intellectual property node embedding vector;

[0118] The initial feature vector of all nodes (all projects and all patents) A single input to R-GCN is performed, followed by multi-layer graph convolution. The propagation formula for each layer is as follows:

[0119]

[0120] In the formula, Represents a node After the first The new embedding vector output after the convolutional layer For nodes In the The current embedding vector of the layer, For nodes neighboring nodes In the The current embedding vector of the layer, For a set of relation types, For nodes In relation The set of neighboring nodes below, The normalization constant is For the relation-specific linear transformation weight matrix, This is a weight matrix specific to the self-loop. This represents the activation function.

[0121] After propagation is complete, the corresponding final embedding vector is extracted based on the node indices recorded in the training samples. and , as input to the scoring layer.

[0122] S4, based on the predicted correlation between the project node embedding vector and the intellectual property node embedding vector mapped by the scoring layer, wherein the scoring layer is a multilayer perceptron or a bilinear transform function;

[0123] Specifically, taking a multilayer perceptron (MLP) as an example: the final embedded vector and spliced ​​as The input consists of a 2-3 layer fully connected network, followed by a Sigmoid activation function, and the output is the predicted correlation between 0 and 1.

[0124] S5, calculate the loss between the predicted correlation degree and the correlation degree label based on the pre-constructed loss function, and adjust the internal parameters of the relationship graph convolutional network and the scoring layer based on the backpropagation of the loss;

[0125] This application uses binary cross-entropy loss, expressed as:

[0126]

[0127] S6. Repeat steps S3-S5 until training is complete, resulting in a relation graph model that includes a relation graph convolutional network and a scoring layer.

[0128] Repeat steps S3 through S5. After each complete data traversal (one epoch), calculate evaluation metrics (e.g., AUC, Hit@10) on the reserved validation set. When the AUC on the validation set no longer improves for n consecutive epochs (e.g., 20 epochs), terminate training and save the model parameters that achieved the best performance on the validation set as the final model. This yields a graph model containing the relationship between the weights of the fully trained R-GCN convolutional layers and the weights of the scoring layers, which can be used in steps S1, 2, and 4 in the production environment.

[0129] S130, calculate the comprehensive correlation score of multiple project data groups, and perform correlation disconnect detection based on the comprehensive correlation score, wherein the comprehensive correlation score includes the generation of time sequence consistency score, personnel consistency score and technology relevance score;

[0130] The calculation of the overall correlation score and the process of detecting correlation disconnection include:

[0131] S131, Based on the R&D project data and intellectual property data in the project data group, calculate the temporal consistency score between the R&D project and each intellectual property. Personnel consistency score Technology relevance score The time-series consistency score Personnel consistency score Technology relevance score The mathematical expressions are as follows:

[0132]

[0133]

[0134]

[0135] In the formula, The application date of the intellectual property right. The project's completion date, This is a gathering of project R&D personnel. It is a collection of inventors of intellectual property rights. For project text semantic vectors, For the semantic vector of the intellectual property abstract text;

[0136] Among them, time consistency score The decay function with a one-year grace period is adopted. The principle is that from the completion of a project to the completion of a patent application, the company objectively needs time for internal technical disclosure and for the agency to draft the application. If the patent is filed within one year after the project completion, it is considered a reasonable delay, and the score decreases linearly; if it exceeds one year, it is directly reduced to zero.

[0137] Personnel Consistency Score The Jaccard coefficient is used to calculate the overlap rate between the project team and the inventor. The principle is that the core technology must be generated by the company's R&D personnel. If there is no overlap between the two personnel lists, it strongly suggests that the patent is purchased or registered under someone else's name, which does not meet the review requirements for independent R&D.

[0138] Technology relevance score The semantic consistency between the project description and the patent abstract is calculated using cosine similarity. The principle is that the smaller the angle between the two in the vector space, the more similar their technical themes are, which can effectively identify technology drift.

[0139] S132, based on the aforementioned timing consistency score Personnel consistency score Technology relevance score Calculate the overall correlation degree The comprehensive correlation degree The mathematical expression is:

[0140]

[0141] In the formula, As the first weight, As the second weight, It is the third weight;

[0142] The sum of the three weights is 1 to ensure the overall correlation. The values ​​always fall within the [0,1] range, forming a unified evaluation benchmark. Weight values ​​can be adjusted according to industry characteristics (e.g., the biopharmaceutical industry can adjust weights accordingly). The system is adjusted to be more industry-adaptable because the R&D cycle is long and patent layout is usually completed before clinical trials.

[0143] S133, Comprehensive correlation of multiple intellectual property rights A weighted summation is performed to obtain the overall relevance score of the R&D project, which includes the overall relevance of multiple intellectual property rights. The weights are based on those specified in the management system;

[0144] In S133, multiple patents associated with the project are... Values ​​are aggregated into a single score for the project dimension, specifically using weights specified by the management system. (For example, the system assigns a weight of 1 to high-value invention patents, 0.5 to ordinary invention patents, and 0.2 to utility model patents.) A weighted average is applied based on the principle that different patents have significantly different strategic values ​​for the project. Core foundational patents have a higher weight than peripheral defensive patents, preventing two biases: high-value patents being diluted by low-value patents or filler patents lowering the score of core patents. The aggregated result is a comprehensive score at the project level, used for cross-project comparison and risk ranking.

[0145] S134, compare the correlation comprehensive score with a preset comprehensive score threshold. If the correlation comprehensive score is greater than or equal to the comprehensive score threshold, determine that there is no correlation or disconnection of intellectual property rights in the R&D project; otherwise, determine that there is correlation or disconnection of intellectual property rights in the R&D project.

[0146] threshold Essentially, this is a mathematical expression of the tolerance level for review. Projects falling below this value indicate a significant disconnect between project R&D investment and intellectual property output in at least one dimension—either in terms of timing, personnel, or technology—failing to meet the self-proving requirements for hard technology attributes. Based on this, the system automatically triggers subsequent remedial tasks (such as mandatory patent application plans and personnel attribution verification), shifting from passive post-event sampling to proactive quantitative early warning, providing repeatable and traceable judgment criteria for risk screening before IPO filing.

[0147] S140, predicting the intellectual property output of each R&D project based on a pre-built output prediction model, and performing output disconnect detection based on the predicted intellectual property output and the associated intellectual property data, specifically including:

[0148] S141, Extract the investment amount, number of R&D personnel, time span, and IPC classification number from the R&D project data to obtain the input vector;

[0149] First, four core input characteristics are extracted from R&D project data. The principle behind this is that the intellectual property output of R&D activities is essentially driven by four factors: resource input intensity, manpower scale, R&D cycle, and technological complexity. The amount of investment reflects the level of financial support, and the difference in dimensions can be eliminated after logarithmic transformation; the number of R&D personnel represents the upper limit of manpower available for participating in technological breakthroughs; the time span (completion date - project initiation date) determines the length of the window period for technological iteration; and the IPC classification number (converted into the number of categories or coding vectors) characterizes the breadth of the project's technical field and the difficulty of breakthroughs.

[0150] By integrating four types of heterogeneous features into a unified numerical input vector This provides a calculable quantitative description for subsequent regression models, enabling the models to establish a statistical mapping relationship between "input and output".

[0151] S142, the input vector is input into the pre-built output prediction model to predict the number of intellectual property outputs of the R&D project.

[0152] This step will input vector Substitute the values ​​into the pre-trained XGBoost regression model to output the predicted number of intellectual property outputs for the project. .

[0153] In one embodiment of this application, the method for constructing the output prediction model includes:

[0154] (1) Obtain the investment amount, number of R&D personnel, time span, IPC classification number, and number of intellectual property rights that are strongly related to the historical R&D projects for multiple historical R&D projects;

[0155] (2) Sample data was constructed based on the investment amount, number of R&D personnel, time span, and IPC classification number of historical R&D projects, and the number of intellectual property rights that are strongly correlated with historical R&D projects was used as labels to construct training data;

[0156] (3) The gradient boosting decision tree model is trained based on the training data and the pre-constructed loss function to obtain the output prediction model, wherein the mathematical expression of the loss function is:

[0157]

[0158]

[0159] In the formula, For the overall optimization objective, The total number of training samples, For training sample index, For the first Labels of each training sample For the first The model prediction value for each training sample. To predict losses, For the index of the tree, For the total number of trees in the regression tree, The complexity penalty is applied to a single tree. is the exponential parameter of the Tweedie distribution.

[0160] The algorithm uses XGBoost (Extreme Gradient Boosting) as its base algorithm, and its core principle is additive ensemble: by adding new regression trees in each round... Each tree specifically fits the residual predicted by the previous tree, and the predictions of multiple weak learners are accumulated to form a strong predictive ability. Its advantages are: it naturally supports mixed numerical and categorical features (IPC classification numbers are input after One-Hot encoding); it automatically captures nonlinear interactions between features through the tree splitting mechanism (such as high input + short cycle and low input + long cycle may produce similar outputs); and it has good robustness to missing values ​​and outliers.

[0161] Through an iterative gradient boosting process, a new tree is added in each round. The first-order gradient of the loss with respect to the predicted value and the second-order Hessian matrix are calculated to guide the growth direction of the tree and the calculation of the weights of the leaf nodes, thus maximizing the total loss. The model decreases in successive rounds. Combined with an early stopping mechanism (training is terminated when the validation set error does not decrease for several consecutive rounds), a converged model is finally produced, which is used in step S140-2 to perform real-time inference and prediction of the expected number of patents for the target project.

[0162] S143, Extract the number of associated intellectual property rights from the associated intellectual property rights data;

[0163] Specifically, from the project data group constructed in step S125 In the process, the actual number of intellectual property rights associated with the project is extracted (i.e., the total number of patents successfully associated with the project after semantic matching and relationship graph completion).

[0164] S144, compare the predicted quantity of intellectual property output with the quantity of related intellectual property. If the predicted quantity of intellectual property output exceeds the quantity of related intellectual property, and the excess is greater than a set quantity threshold, it is determined that there is a disconnect in output.

[0165] This step will predict the quantity of intellectual property output. Number of intellectual property rights actually related The comparison is based on the following rule: when When the preset quantity threshold is 2 (default value), it is determined that there is a disconnect between output and production in the R&D project.

[0166] If expected output Significantly higher than actual related output This indicates that although the project invested a significant amount of resources, it failed to translate into corresponding intellectual property results as scheduled. This abnormal state of high input and low output constitutes a serious disconnect in the output dimension, and the reasons may include: incorrect research and development direction, failure to overcome technical challenges, weak awareness of intellectual property layout, or lagging management processes.

[0167] Figure 4This is an example of the disconnection detection result of the automatic detection method for R&D-intellectual property disconnection based on spatiotemporal correlation of multi-source heterogeneous data in this application. The final output disconnection detection result is as follows: Figure 4 As shown.

[0168] This application presents an automatic detection method for R&D-intellectual property disconnect based on the spatiotemporal correlation of multi-source heterogeneous data. By constructing a unified data collection and entity alignment mechanism, it effectively solves the cross-system barriers of multi-source heterogeneous data in terms of time anchors and technical semantics, realizing the automatic correlation and standardized grouping of R&D projects and intellectual property data. Building upon this, it innovatively introduces a three-dimensional comprehensive scoring model to quantitatively evaluate the project-patent correlation strength from three dimensions: temporal consistency, personnel consistency, and technical relevance. This accurately identifies hidden disconnect risks caused by misalignment between project cycles and patent examination cycles, drifting technology routes, or disconnection between inventors and R&D personnel. Simultaneously, it combines an output prediction model to regress and estimate the expected number of patents for a project and compares this with actual correlated output, achieving anomaly detection in R&D investment conversion efficiency. This application upgrades the traditional ex-post sampling inspection relying on manual experience to a systematic, multi-dimensional proactive early warning and quantitative assessment. It can effectively help companies comprehensively investigate logical flaws between R&D and intellectual property before IPO filing, significantly reducing the inquiry risks and on-the-spot remedial costs caused by insufficient correlation of core assets, and greatly improving the efficiency and compliance quality of listing review.

[0169] like Figure 5 As shown, this application also provides an automatic detection system for the disconnect between R&D and intellectual property based on the spatiotemporal correlation of multi-source heterogeneous data, including:

[0170] The acquisition module is used to acquire R&D project data for multiple R&D projects and intellectual property data for multiple intellectual property rights of the target enterprise. The R&D project data includes project initiation time, completion time, project description text, list of R&D personnel, and investment amount; the intellectual property data includes application date, list of inventors, abstract text, and IPC classification.

[0171] The matching module is used to perform semantic similarity and relationship graph matching on the R&D project data and the intellectual property data to obtain multiple project data groups, wherein each project data group contains the R&D project data of one R&D project and the associated intellectual property data.

[0172] The correlation decoupling detection module is used to construct a bipartite time series graph of multiple data groups, calculate a three-dimensional correlation comprehensive score based on the bipartite time series graph, and perform correlation decoupling detection based on the correlation comprehensive score. The correlation comprehensive score includes the generation of time series consistency score, personnel consistency score, and technology relevance score.

[0173] The output disconnect detection module is used to predict the number of intellectual property outputs for each R&D project based on a pre-built output prediction model, and to perform output disconnect detection based on the predicted number of intellectual property outputs and the associated intellectual property data.

[0174] This application's automatic detection system for R&D-intellectual property disconnect based on the spatiotemporal correlation of multi-source heterogeneous data effectively solves the cross-system barriers of multi-source heterogeneous data in terms of time anchors and technical semantics by constructing a unified data collection and entity alignment mechanism, realizing the automatic correlation and standardized grouping of R&D projects and intellectual property data. Based on this, it innovatively introduces a three-dimensional comprehensive scoring model to quantitatively evaluate the project-patent correlation strength from three dimensions: temporal consistency, personnel consistency, and technical relevance. This accurately identifies hidden disconnect risks caused by misalignment between project cycles and patent examination cycles, drifting technology routes, or decoupling between inventors and R&D personnel. Simultaneously, it combines an output prediction model to regress and estimate the expected number of patents for a project and compares this with actual associated output, achieving anomaly detection in R&D investment conversion efficiency. This application upgrades the traditional ex-post sampling inspection relying on manual experience to a systematic, multi-dimensional proactive early warning and quantitative assessment, effectively helping companies comprehensively investigate logical flaws between R&D and intellectual property before IPO filing, significantly reducing the inquiry risks and on-the-spot remedial costs caused by insufficient correlation of core assets, and greatly improving the efficiency and compliance quality of listing review.

[0175] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0176] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory so that the terminal performs any of the methods in this embodiment.

[0177] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0178] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0179] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0180] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0181] In the above embodiments, although the present application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. The embodiments of the present application are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0182] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. An automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data, characterized in that, include: The system acquires R&D project data for multiple R&D projects and intellectual property data for multiple intellectual property rights from the target company. The R&D project data includes project initiation time, completion time, project description text, list of R&D personnel, and investment amount. The intellectual property data includes application date, list of inventors, abstract text, and IPC classification. The R&D project data and the intellectual property data are matched based on semantic similarity and relationship graph to obtain multiple project data groups, wherein each project data group contains the R&D project data of one R&D project and the associated intellectual property data; Calculate the overall correlation score of multiple project data groups, and perform correlation decoupling detection based on the overall correlation score, wherein the overall correlation score includes time sequence consistency score, personnel consistency score, and technology relevance score; The system predicts the intellectual property output of each R&D project based on a pre-built output prediction model, and detects output disconnect based on the predicted intellectual property output and the associated intellectual property data.

2. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 1, characterized in that, The R&D project data and the intellectual property data are matched based on semantic similarity and relationship graphs to obtain multiple project data groups, including: Based on the fine-tuned language model, the R&D project data and the intellectual property data are vectorized to obtain the first semantic feature vector and the second semantic feature vector. Calculate the cosine similarity between the first semantic feature vector and the second semantic feature vector; For each first semantic feature vector, a second semantic feature vector with a cosine similarity greater than or equal to a preset threshold is used as the associated second semantic feature vector; and the associated intellectual property data of each R&D project is determined based on the association relationship of the semantic feature vectors. Based on a pre-built relational graph model, intellectual property data that is not associated with R&D projects is matched with R&D projects, and the intellectual property data that matches the R&D projects is used as associated intellectual property data. Project data groups are constructed based on the R&D project data and related intellectual property data of each R&D project.

3. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 2, characterized in that, The method for constructing the relational graph model includes: S1, define node features and relation edges, wherein the node features include project text semantic vectors and intellectual property text semantic vectors, and the relation edges include personnel intersection, time window, IPC classification and text similarity; S2, obtain a training sample set, wherein the training sample set includes multiple training samples, and each training sample includes node feature samples and correlation labels; S3, Substitute the training samples into the relational graph convolutional network, and propagate them in the relational graph convolutional network to obtain the project node embedding vector and the intellectual property node embedding vector; S4, based on the predicted correlation between the project node embedding vector and the intellectual property node embedding vector mapped by the scoring layer, wherein the scoring layer is a multilayer perceptron or a bilinear transform function; S5, calculate the loss between the predicted correlation degree and the correlation degree label based on the pre-constructed loss function, and adjust the internal parameters of the relationship graph convolutional network and the scoring layer based on the backpropagation of the loss; S6. Repeat steps S3-S5 until training is complete, resulting in a relation graph model that includes a relation graph convolutional network and a scoring layer.

4. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 2, characterized in that, Based on a pre-built relationship graph model, intellectual property data not associated with R&D projects is matched with R&D projects, including: A graph input vector is constructed based on the second semantic feature vector of intellectual property data that is not associated with R&D projects and the first semantic feature vector of multiple R&D project data, respectively. The graph input vector is input into a pre-constructed relational graph model to obtain a correlation score. For the same intellectual property data, if there is a target R&D project data with the highest relevance score that is greater than or equal to a preset relevance threshold, the intellectual property data is determined to match the target R&D project data.

5. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 1, characterized in that, Calculate the composite score of multiple project data groups, including: Based on the R&D project data and intellectual property data in the aforementioned project data group, a time-series consistency score is calculated between the R&D project and each intellectual property. Personnel consistency score Technology relevance score The time-series consistency score Personnel consistency score Technology relevance score The mathematical expressions are as follows: In the formula, The application date of the intellectual property right. The project's completion date, This is a gathering of project R&D personnel. It is a collection of inventors of intellectual property rights. For project text semantic vectors, For the semantic vector of the intellectual property abstract text; Based on the time consistency score Personnel consistency score Technology relevance score Calculate the overall correlation degree The comprehensive correlation degree The mathematical expression is: In the formula, As the first weight, As the second weight, It is the third weight; Comprehensive correlation of multiple intellectual property rights A weighted summation is performed to obtain the overall relevance score of the R&D project, which includes the overall relevance of multiple intellectual property rights. The weights are based on the specifications in the management system.

6. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 1, characterized in that, Based on the aforementioned comprehensive correlation score, correlation disconnection detection is performed, including: The correlation comprehensive score is compared with a preset comprehensive score threshold. If the correlation comprehensive score is greater than or equal to the comprehensive score threshold, it is determined that there is no disconnection of intellectual property rights in the R&D project; otherwise, it is determined that there is a disconnection of intellectual property rights in the R&D project.

7. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 1, characterized in that, The method for constructing the output prediction model includes: Obtain the investment amount, number of R&D personnel, time span, IPC classification number, and number of intellectual property rights strongly associated with the historical R&D projects for multiple historical R&D projects; Sample data is constructed based on the investment amount, number of R&D personnel, time span, and IPC classification number of historical R&D projects, and training data is constructed by using the number of intellectual property rights that are strongly correlated with historical R&D projects as labels. The gradient boosting decision tree model is trained based on the training data and a pre-built loss function to obtain the output prediction model, wherein the mathematical expression of the loss function is: In the formula, For the overall optimization objective, The total number of training samples, For training sample index, For the first Labels of each training sample For the first The model prediction value for each training sample. To predict losses, For the index of the tree, For the total number of trees in the regression tree, The complexity penalty is applied to a single tree. is the exponential parameter of the Tweedie distribution.

8. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 1, characterized in that, Based on a pre-built output prediction model, the predicted number of intellectual property outputs for each R&D project is calculated, including: The input vector is obtained by extracting the investment amount, number of R&D personnel, time span, and IPC classification number from the R&D project data. The input vector is fed into a pre-built output prediction model to predict the quantity of intellectual property output from the R&D project.

9. The automatic detection method for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data according to claim 1, characterized in that, Output disconnect detection is performed based on the predicted quantity of intellectual property output and the associated intellectual property data, including: Extract the number of related intellectual property rights from the related intellectual property data; The predicted quantity of intellectual property output is compared with the quantity of associated intellectual property. If the predicted quantity of intellectual property output exceeds the quantity of associated intellectual property, and the excess is greater than a set threshold, it is determined that there is a disconnect in output.

10. An automatic detection system for R&D-intellectual property disconnect based on spatiotemporal correlation of multi-source heterogeneous data, characterized in that, include: The acquisition module is used to acquire R&D project data for multiple R&D projects and intellectual property data for multiple intellectual property rights of the target enterprise. The R&D project data includes project initiation time, completion time, project description text, list of R&D personnel, and investment amount; the intellectual property data includes application date, list of inventors, abstract text, and IPC classification. The matching module is used to perform semantic similarity and relationship graph matching on the R&D project data and the intellectual property data to obtain multiple project data groups, wherein each project data group contains the R&D project data of one R&D project and the associated intellectual property data. The correlation decoupling detection module is used to construct a bipartite time series graph of multiple data groups, calculate a three-dimensional correlation comprehensive score based on the bipartite time series graph, and perform correlation decoupling detection based on the correlation comprehensive score. The correlation comprehensive score includes the generation of time series consistency score, personnel consistency score, and technology relevance score. The output disconnect detection module is used to predict the number of intellectual property outputs for each R&D project based on a pre-built output prediction model, and to perform output disconnect detection based on the predicted number of intellectual property outputs and the associated intellectual property data.

Images