Method and system for predicting transformation potential of scientific and technological achievements based on multi-source data fusion

Abstract

The application relates to the technical field of artificial intelligence and machine learning, and discloses a scientific and technological achievement transformation potential prediction method and system based on multi-source data fusion, which comprises the following steps: collecting non-public early scientific research activity multi-source heterogeneous data; performing multi-modal feature extraction and semantic alignment; constructing a dynamic causal graph combining Granger causality test and counterfactual reasoning; generating a scientific and technological achievement state embedding through a time sequence perception graph attention network; decoding output of three-dimensional prediction values of technical maturity, market adaptability and industrialization obstacles, and triggering graded early warning. The system comprises corresponding function modules, supports real-time and interpretable early potential identification, and significantly improves the timeliness, causal reliability and decision support capability of prediction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning technology, specifically relating to a method and system for predicting the potential for the transformation of scientific and technological achievements based on multi-source data fusion. Background Technology

[0002] With the accelerated evolution of technological innovation, the transformation of scientific and technological achievements has become a core engine driving industrial upgrading and economic growth. Currently, the assessment of the potential value of scientific and technological achievements generally relies on publicly available structured data, such as the number of patents granted, the frequency of citations of academic papers, and corporate investment and financing records. While such methods have a certain statistical basis, they are inherently limited by the lag in information release, making it difficult to capture cutting-edge technological trends that are still in the laboratory stage or have not yet been formally disclosed. More importantly, existing predictive models are mostly based on linear regression or simple correlation analysis, implicitly assuming that "co-occurrence of indicators equates to causal relationship," which cannot effectively distinguish between real driving factors and accidental synchronous phenomena, leading to misjudgments or omissions of the potential of disruptive technologies.

[0003] The method for predicting the commercialization potential of scientific and technological achievements based on multi-source heterogeneous data fusion aims to construct a cross-level early warning mechanism by integrating dynamic signals from three dimensions: scientific discovery, technological research and development, and industrial application. This direction emphasizes extracting forward-looking features from non-traditional channels (such as preprint platforms, recruitment data, and patent preliminary examination texts) and exploring their leading relationships and structural dependencies in time series, in order to break through the limitations of the traditional static evaluation paradigm.

[0004] Achieving the above goals still faces the following challenges: There is a lack of effective causal modeling methods between scientific signals (such as the level of attention to emerging theories) and technological behaviors (such as the breadth of patent claim coverage). Most systems only perform weighted summation or shallow fusion, neglecting the complex temporal dependencies and feedback mechanisms between cross-domain indicators. Industry dynamics (such as the investment pace of sub-sectors and the talent layout of startups) are often simplified into lagging verification indicators, failing to participate in early judgments as proactive predictive factors. When faced with "emergent" innovation—that is, when a fundamental discovery is suddenly cross-referenced or technologically adopted by multiple unrelated fields—existing models lack the ability to identify mutation patterns and cross-domain diffusion paths, thus failing to transform it into a key signal for potential multiplication. Summary of the Invention

[0005] This invention provides a method and system for predicting the commercialization potential of scientific and technological achievements based on multi-source data fusion. It aims to solve the signal lag problem caused by reliance on publicly available patents, papers, and investment data in existing technologies, as well as the misjudgment problem caused by the inability to effectively distinguish between causal relationships and accidental correlations due to the use of linear correlation assumptions. This invention constructs a data collection channel for non-public early-stage scientific research activities, integrates structured and unstructured multi-source heterogeneous data, and establishes a dynamic graph neural network model with causal inference as its core. This enables real-time, accurate, and interpretable quantitative prediction of the commercialization potential of scientific and technological achievements in their nascent stages.

[0006] This invention provides a method for predicting the commercialization potential of scientific and technological achievements based on multi-source data fusion, which includes:

[0007] Collect raw data streams of non-public early scientific research activities. The raw data streams include research project application texts, laboratory equipment usage logs, internal collaborative communication records of researchers, academic conference preprint submission records, and metadata of technology prototype demonstration videos.

[0008] The original data stream is subjected to multimodal feature extraction and semantic alignment processing to generate a standardized multi-source feature vector sequence;

[0009] A dynamic causal graph structure is constructed, wherein the dynamic causal graph structure uses scientific research entities as nodes and cross-modal event triggering relationships as directed edges, and the edge weights are jointly determined by Granger causality test and counterfactual reasoning.

[0010] The multi-source feature vector sequence is input into a time-aware graph attention network. At each time step, the network aggregates neighborhood node information according to the dynamic causal graph structure and outputs a hidden layer embedding vector representing the current state of the scientific and technological achievement.

[0011] Based on the hidden layer embedding vector, a predicted value of the potential for the transformation of scientific and technological achievements is generated through the decoder module. The predicted value includes three quantitative indicators: technology maturity index, market adaptability score, and industrialization barrier coefficient.

[0012] Based on the predicted value, a graded early warning mechanism is triggered. When the industrialization barrier coefficient is lower than a preset threshold and the technology maturity index is higher than a preset threshold, a high-potential scientific and technological achievement identification signal is generated.

[0013] As one embodiment of the present invention, the collection of raw data streams from non-public early scientific research activities specifically includes:

[0014] The research project application text is obtained through the interface of the scientific research management information system. The text includes a description of the technical route, the expected form of the results, and information on the team composition.

[0015] The laboratory IoT gateway collects device usage logs, which record device type, usage duration, operation frequency, and parameter configuration change sequence.

[0016] The internal collaborative communication records of researchers are intercepted by an encrypted communication proxy module. After the records are de-identified, the co-occurrence frequency of keywords and the evolution path of topics are retained.

[0017] Access to preprint submission records via academic conference digital infrastructure, including submission timestamps, subject classification tags, and reference citation maps;

[0018] Metadata from the technology prototype demonstration video is extracted using a video analytics server. This metadata includes action recognition tag sequences, interface interaction heatmaps, and sentiment polarity scores of the speech-to-text transcription.

[0019] As one embodiment of the present invention, the multimodal feature extraction and semantic alignment processing of the original data stream specifically includes:

[0020] Adaptive word embedding is performed on the text of scientific research project application, and sentence vectors are generated by a Chinese character-level convolutional neural network pre-trained on a corpus of science and technology policies.

[0021] The laboratory equipment usage logs are sliced ​​into time windows, and the equipment activity entropy value and operation mode transition matrix are calculated within each window.

[0022] A dynamic topic model is constructed from internal collaborative communication records to extract the topic distribution vector and the transition probability between topics within each 7-day cycle.

[0023] Analyze the reference citation map of preprint submission records, calculate the node centrality index and generate citation influence decay curves;

[0024] A multi-channel temporal feature tensor is constructed by fusing action recognition tag sequences and speech emotion polarity scores with metadata from the technical prototype demonstration video.

[0025] The feature vectors of each modality are mapped to a unified dimension through a shared latent space projection matrix and aligned by timestamps to form a multi-source feature vector sequence.

[0026] As one embodiment of the present invention, the construction of the dynamic causal graph structure specifically includes:

[0027] Define a set of scientific research entities, which includes five types of nodes: researchers, research institutions, technical keywords, equipment types, and project numbers.

[0028] Perform Granger causality test on the event sequence between any two types of nodes. If the null hypothesis of no causality is rejected at a significance level of 0.05, then a directed edge is established.

[0029] Further counterfactual reasoning verification is performed on the established directed edges. The change in the probability of the target node event is observed after the source node event is removed by intervention simulation. If the absolute value of the change is greater than the preset sensitivity threshold, the directed edge is retained.

[0030] Each retained directed edge is assigned a weight equal to the weighted sum of the Granger causality F-statistic and the counterfactual effect, and the weights are updated every 24 hours.

[0031] As one embodiment of the present invention, the time-aware graph attention network includes an input layer, a multi-head graph attention layer, a time-gated recurrent unit layer, and an output layer;

[0032] In the input layer, the multi-source feature vector sequence is input together with the adjacency matrix of the dynamic causal graph structure;

[0033] In the multi-head graph attention layer, each attention head independently calculates the importance coefficient between nodes. The importance coefficient is determined by the node feature similarity and edge weights, as shown in the formula:

[0034]

[0035] in , For nodes , eigenvectors, For learnable weight matrix, For attention vectors, For edge weights, This is the adjustment coefficient;

[0036] In the temporal gated recurrent unit layer, the aggregated feature sequence output by the multi-head graph attention layer is input into the gated recurrent unit. This unit contains update gates and reset gates to control the retention and forgetting of historical state information.

[0037] In the output layer, a fully connected transformation is applied to the hidden state at the final moment of the gated recurrent unit to generate the hidden layer embedding vector.

[0038] As one embodiment of the present invention, the decoder module includes three parallel sub-networks, which respectively correspond to the prediction of the technology maturity index, market adaptability score, and industrialization barrier coefficient.

[0039] The technology maturity index subnetwork receives hidden layer embedding vectors and outputs continuous values ​​ranging from 0 to 10. Its loss function is the sum of mean squared error and grade constraint regularization term.

[0040] The market fit scoring subnetwork introduces external macroeconomic indicators as auxiliary inputs, and integrates the hidden layer embedding vectors and macroeconomic indicators through a cross-attention mechanism to output a 100-point score ranging from 0 to 100.

[0041] The industrialization barrier coefficient subnetwork adopts a quantile regression structure to output the barrier coefficient estimates at the 15th, 50th, and 85th quantiles, which are used to characterize the prediction uncertainty.

[0042] As one embodiment of the present invention, the graded early warning mechanism sets three response thresholds:

[0043] When the technology maturity index is greater than 6.5 and the industrialization barrier coefficient is less than 0.3, a level one warning is triggered, a list of high-potential scientific and technological achievements is generated and pushed to the technology transfer office;

[0044] When the technology maturity index is between 5.0 and 6.5 and the industrialization barrier coefficient is between 0.3 and 0.5, a level-two warning is triggered, and the expert review process is initiated.

[0045] When the technology maturity index is less than 5.0 or the industrialization barrier coefficient is greater than 0.5, it is classified into the regular monitoring queue and no active intervention is triggered.

[0046] This invention provides a system for predicting the potential for the transformation of scientific and technological achievements based on multi-source data fusion, which includes:

[0047] The non-public early research activity data collection module is used to collect research project application texts, laboratory equipment usage logs, internal collaborative communication records of researchers, academic conference preprint submission records, and metadata of technology prototype demonstration videos.

[0048] The multimodal feature extraction and semantic alignment module is used to perform multimodal feature extraction and semantic alignment processing on the original data stream to generate a standardized multi-source feature vector sequence;

[0049] The dynamic causal graph construction module is used to construct a dynamic causal graph structure with scientific research entities as nodes and cross-modal event triggering relationships as directed edges. The edge weights are jointly determined by Granger causality testing and counterfactual reasoning.

[0050] The time-aware graph neural network processing module is used to input the multi-source feature vector sequence into a time-aware graph attention network and output a hidden layer embedding vector representing the current state of the scientific and technological achievement.

[0051] The potential prediction decoding module is used to generate a predicted value of the potential for the transformation of scientific and technological achievements, which includes a technology maturity index, a market adaptability score, and an industrialization barrier coefficient, based on the hidden layer embedding vector through the decoder module.

[0052] The graded early warning triggering module is used to trigger the graded early warning mechanism based on the predicted value and generate a high-potential scientific and technological achievement identification signal.

[0053] As one embodiment of the present invention, the non-public early scientific research activity data collection module is implemented through a data proxy node deployed on the intranet of the research institution. The data proxy node adopts a zero-trust security architecture, collects only authorized data fields, and completes preliminary de-identification processing locally.

[0054] As one embodiment of the present invention, the dynamic causal graph construction module has a built-in causal discovery engine, which periodically performs a full graph causal relationship scan. Each scan covers newly added scientific research entities and events in the past 30 days and updates the graph structure topology and edge weights.

[0055] As one embodiment of the present invention, the time-aware graph neural network processing module is deployed on a distributed computing cluster, supporting real-time inference of dynamic causal graphs with a scale of more than 100,000 nodes, with a single inference latency of no more than 500 milliseconds.

[0056] As one embodiment of the present invention, the output of the potential prediction decoding module is accompanied by an interpretability report. The report identifies the original data segment that contributes the most to the prediction result through the gradient backpropagation method and marks its data source and time window.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] 1. This invention breaks through the limitations of traditional scientific and technological achievement transformation potential assessment relying on lagging public data. By directly accessing non-public early scientific research activity data streams, the prediction window is moved forward to the technology bud stage, realizing the early capture of disruptive potential.

[0059] 2. This invention abandons the linear correlation assumption and constructs a dynamic causal graph structure that integrates Granger causality test and counterfactual reasoning, which effectively distinguishes between real driving factors and accidental correlations and significantly improves the causal reliability of prediction results.

[0060] 3. The temporal-aware graph attention network proposed in this invention can adaptively aggregate cross-modal correlation information from multi-source heterogeneous data, and output multi-dimensional quantitative indicators through the decoder module, providing a refined and operable basis for technology transfer decisions.

[0061] 4. The hierarchical early warning mechanism and interpretability reporting function of the present invention further enhance the system's practicality and decision support capabilities, and solve the core defects of the prior art in terms of early warning accuracy, causal inference rigor, and result interpretability. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method and system for predicting the transformation potential of scientific and technological achievements based on multi-source data fusion proposed in this invention.

[0063] Figure 2 This is a schematic diagram of the core principle framework of the construction of dynamic causal graph structure and the joint verification of Granger causality test and counterfactual reasoning in this invention;

[0064] Figure 3 This is a flowchart illustrating the logical process of multimodal feature extraction and semantic alignment processing of the original data stream of non-public early scientific research activities in this invention.

[0065] Figure 4 This is a flowchart illustrating the collaborative processing flow of the multi-head attention mechanism and the temporal gating loop unit in the temporal-aware graph attention network of this invention.

[0066] Figure 5 This is a schematic diagram of the logical framework for the decoder module to generate a three-dimensional transformation potential index for the hidden layer embedding vector in this invention.

[0067] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the graded early warning mechanism and the high-potential scientific and technological achievement identification signal in this invention. Detailed Implementation

[0068] Please refer to the attached document. Figures 1 to 6 This invention provides a method and system for predicting the commercialization potential of scientific and technological achievements based on multi-source data fusion. Its core lies in collecting raw data streams from non-public early-stage scientific research activities and constructing a dynamic graph neural network model based on causal inference to achieve real-time, accurate, and interpretable quantitative prediction of the commercialization potential of scientific and technological achievements in their nascent stages. The specific implementation methods will be described in detail below, combining the various technical steps and system modules of this invention.

[0069] The method includes the following steps:

[0070] S1, collecting raw data streams from non-public early scientific research activities;

[0071] S2, perform multimodal feature extraction and semantic alignment processing on the original data stream to generate a standardized multi-source feature vector sequence;

[0072] S3, construct a dynamic cause-effect graph structure;

[0073] S4, input the multi-source feature vector sequence into a time-aware graph attention network, and output a hidden layer embedding vector representing the current state of the scientific and technological achievement;

[0074] S5, Based on the hidden layer embedding vector, a predicted value of the potential for transformation of scientific and technological achievements is generated through the decoder module;

[0075] S6, trigger the graded early warning mechanism based on the predicted value.

[0076] In step S1, raw data streams of non-public early-stage research activities are collected. These raw data streams encompass five types of heterogeneous data sources: research project application documents, laboratory equipment usage logs, internal collaborative communication records of researchers, preprint submission records of academic conferences, and metadata of technology prototype demonstration videos. All of these data sources originate from the research institution's intranet or controlled digital infrastructure, and are highly sensitive and non-public, making them impossible to obtain through traditional evaluation systems.

[0077] Specifically, the research project application text is obtained through the research management information system interface. This text contains structured and semi-structured fields such as technical route description, expected outcome, and team composition information. The system is deployed on the data proxy node at the backend of the research management platform, reading only authorized field content and performing preliminary anonymization processing locally, such as removing privacy information like applicant's name and contact information, while retaining core elements that can be used for subsequent modeling, such as technical keywords, research objectives, and collaborating institutions.

[0078] Laboratory equipment usage logs are collected in real time by the laboratory's IoT gateway. Each piece of experimental equipment is equipped with embedded sensors and communication modules, continuously uploading log entries such as equipment type identification, usage start and end timestamps, operation frequency, and parameter configuration change sequences. The system streams the log streams, slicing them into fixed time windows. Within each window, the system calculates the equipment activity entropy value to measure the regularity or suddenness of equipment usage behavior. Simultaneously, it calculates an operation mode transition matrix to characterize the transition probability between different operation commands, reflecting the complexity and exploratory nature of the experimental process.

[0079] Internal collaborative communication records among researchers are intercepted via an encrypted communication proxy module. This module, deployed at the front end of the research institution's internal instant messaging platform or email server, employs a zero-trust security architecture. Without storing the original message content, it extracts the co-occurrence frequency of de-identified keywords and the evolution path of topics. Keyword extraction is based on joint filtering using a domain dictionary and a named entity recognition model, ensuring the preservation of key semantic units such as technical terms, project codes, and equipment names. The topic evolution path is obtained through topic clustering and topic shift modeling within a sliding time window, used to track the migration trajectory of research focus.

[0080] Preprint submission records for academic conferences are accessed through the conference's digital infrastructure. The system interfaces with the APIs of mainstream preprint platforms, simultaneously acquiring metadata such as submission timestamps, subject classification tags, and citation graphs when users submit their initial drafts. The citation graph is stored as a directed graph, with nodes representing cited documents and edges representing citation relationships. Based on this, the system calculates centrality metrics for each node, such as PageRank and betweenness centrality, and fits a citation impact decay curve to assess the breadth and sustained impact of technical ideas.

[0081] Metadata from the technology prototype demonstration videos is extracted by a video analytics server. This server receives technology demonstration video streams uploaded by research teams and uses a multimodal perception model to perform action recognition, interface interaction heatmap generation, speech-to-text transcription, and sentiment polarity analysis. Action recognition tag sequences record the time distribution of key operation steps; the interface interaction heatmap reflects the user's focus on system functions; and the speech-to-text, after being processed by a sentiment analysis model, outputs a sentiment polarity score, used to determine the level of confidence or technological maturity cues during the demonstration.

[0082] In step S2, multimodal feature extraction and semantic alignment are performed on the original data stream to generate a standardized multi-source feature vector sequence. This process designs dedicated feature extractors for different modalities and achieves cross-modal alignment by sharing latent space projection.

[0083] For research project proposal texts, the system employs a Chinese character-level convolutional neural network pre-trained on a science and technology policy corpus to perform domain-adaptive word embedding. This network uses characters as the basic input unit and captures local semantic combinations through multi-layer convolution and pooling operations, ultimately outputting fixed-dimensional sentence vectors. This effectively avoids the semantic drift problem of traditional word vectors in scenarios with sparse technical terms.

[0084] For laboratory equipment usage logs, the system calculates the equipment activity entropy value within each time window, defined as follows: ,in For the first The frequency of occurrence of each operation mode within the window; and the construction of the operation mode transition matrix. ,in Indicates from pattern Transfer to mode The probability of the transition mode is calculated. After singular value decomposition, the first k principal components of the matrix are taken as eigenvectors, preserving the main transition mode information.

[0085] For internal collaborative communication records, the system constructs a dynamic topic model, which is updated on a sliding basis every 7 days. Within each period, a variational inference algorithm is used to estimate the topic distribution vector. and the transition probability matrix between topics . It reflects the activity level of each research topic within the current cycle. Depicts the directionality and stability of the theme's evolution.

[0086] For preprint submission records, the system parses their reference citation map and calculates the PageRank value of each cited node. Defined as ,in The damping coefficient is... This represents the total number of nodes. Pointing to a node The set of nodes, For nodes The system further fits the exponential function of citation influence decaying over time. Extract the attenuation rate As a characteristic, For initial citation influence, for The impact of citations at any given moment.

[0087] For metadata from technology prototype demonstration videos, the system integrates action recognition tag sequences. Interface interaction heatmap , The dimensions of the heatmap and the emotional polarity score of the voice. Constructing a multi-channel temporal feature tensor ,That The dimensions of the multi-channel temporal feature tensor are shown in the figure. The number of channels includes action category encoding, heatmap statistical features (such as mean, variance, and peak position), sentiment score, and its first-order difference.

[0088] The feature vectors of all modalities are projected through a learnable shared latent space matrix. Mapping to a unified dimension ,in For the first The original feature dimensions of the modality. After mapping, the features are aligned by timestamps, and missing time points are filled using forward padding or linear interpolation, ultimately forming a multi-source feature vector sequence. , =1,2,..., .

[0089] In step S3, a dynamic causal graph structure is constructed. This graph uses research entities as nodes and cross-modal event triggering relationships as directed edges, with edge weights determined jointly by Granger causality testing and counterfactual reasoning.

[0090] The research entity set contains five types of nodes: researchers, research institutions, technical keywords, equipment types, and project numbers. The system assigns a unique identifier to each new entity and maintains an entity attribute database, recording metadata such as its type, creation time, and associated events.

[0091] For any two types of nodes u and v, the system collects their event time series over the past 30 days. and An event is defined as an observable action that occurs at a specific point in time, such as "submitting a preprint," "using a device," or "initiating a collaborative session." The system converts the event sequence into a binary time series. and ,like The value is 1 if the event exists at any given time, and 0 otherwise.

[0092] Subsequently, the system and Perform Granger causality testing. Null hypothesis. for" Granger caused ",Right now The future value cannot be The past values ​​are significant predictors. The test uses a vector autoregression model. Calculate the F-statistic .like If the value is less than 0.05, then reject. A preliminary directed edge system has been established. .

[0093] To further verify the causal validity, the system performs counterfactual reasoning on the established directed edges. Through intervention simulation, exist Force the value to 0 within the interval, and rerun the prediction model to obtain the result. counterfactual predictions Calculate the actual observed values. Difference from counterfactual predictions .like If the sensitivity threshold is greater than a preset threshold (e.g., 0.1), the directed edge is retained; otherwise, it is deleted to eliminate spurious correlations.

[0094] Each retained directed edge is assigned a weight. ,in This is an adjustment coefficient, with a value of 0.6. The weights are updated daily to reflect the dynamic evolution of causal relationships.

[0095] In step S4, the multi-source feature vector sequence The input is fed into a time-aware graph attention network. This network consists of an input layer, a multi-head graph attention layer, a time-gated recurrent unit layer, and an output layer.

[0096] At the input layer, the system will With dynamic cause-effect graphs in time adjacency matrix Common input. For a weighted directed graph, elements If an edge exists → Otherwise, it is 0.

[0097] In the multi-head graph attention layer, each attention head h independently computes nodes. To the neighbors Importance coefficient The formula is:

[0098]

[0099] in For the first The learnable weight matrix of the head, Let || be the attention vector, and || denotes vector concatenation. This is the edge weight adjustment coefficient, with a value of 0.3. (The final node value is missing from the original text.) The representation of is , for The neighbor set is obtained by concatenating the outputs of multiple sources and then performing a linear transformation. .

[0100] In the timing-gated loop unit layer Sequence input gated recurrent unit (GRU). Update gate. Reset the door Candidate state Ultimately hidden state ,in, for The input sequence at time , for The final hidden state vector at time step 1. for The candidate hidden state vector at time step 1. for Reset gate vector at time step / For input / History hidden state To the update gate The weight matrix, / For input / History hidden state To the update gate The weight matrix, / For input / History hidden state To the update gate The weight matrix, For the sigmoid function, This is an element-wise multiplication.

[0101] In the output layer, take the final time step. Hidden state Hidden layer embedding vectors are generated through fully connected layer transformation. As a comprehensive representation of the current state of scientific and technological achievements.

[0102] In step S5, based on the hidden layer embedding vector The system generates predicted values ​​for the commercialization potential of scientific and technological achievements through a decoder module. This module contains three parallel sub-networks.

[0103] Technology Maturity Index Sub-network Reception The output range is a continuous value from 0 to 10. Its loss function is... ,in For the true value vector, For the predicted value vector, For the rank constraint regularization term, the penalty is the inconsistent ranking of the predicted values ​​with the true rank labels. The regularization coefficient... Take 0.5.

[0104] Market fit scoring subnetwork introduces external macroeconomic indicators As an auxiliary input, such as GDP growth rate Industry Investment Popularity Index Score for policy support This subnetwork is fused through a cross-attention mechanism. and :

[0105]

[0106]

[0107]

[0108]

[0109] Output a percentage score ranging from 0 to 100.

[0110] The industrialization barrier coefficient subnetwork uses a quantile regression structure to output barrier coefficient estimates at the 15th, 50th, and 85th quantiles. The loss function is ,in The quantile loss function, This represents the estimated barrier coefficient output by the quantile regression model. This represents the actual observed value of the industrialization barrier coefficient. This design is used to characterize forecasting uncertainty and support risk-based decision-making.

[0111] In step S6, a tiered early warning mechanism is triggered based on the predicted value. The system sets three response thresholds:

[0112] When the technology maturity index is greater than 6.5 and the industrialization barrier coefficient is less than 0.3, a level one warning is triggered, a list of high-potential scientific and technological achievements is generated and sent to the technology transfer office, along with an interpretable report.

[0113] When the technology maturity index is between 5.0 and 6.5 and the industrialization barrier coefficient is between 0.3 and 0.5, a level-two warning is triggered, the expert review process is initiated, and field experts are invited to conduct manual analysis of the original data fragments.

[0114] When the technology maturity index is less than 5.0 or the industrialization barrier coefficient is greater than 0.5, it is included in the regular monitoring queue, and the system continues to track its subsequent data flow without triggering active intervention.

[0115] The interpretability report is generated using the gradient backpropagation method. The system calculates the predicted output against the original multi-source feature vector sequence. gradient Take the original data segments corresponding to the five time windows with the largest absolute values, and label their data sources, timestamps, entity nodes, and causal edge paths to form a visual explanation chain.

[0116] The system includes a non-public early scientific research activity data acquisition module, a multimodal feature extraction and semantic alignment module, a dynamic causal graph construction module, a time-series-aware graph neural network processing module, a potential prediction decoding module, and a graded early warning triggering module.

[0117] The non-public early research activity data collection module is deployed on the research institution's intranet and includes multiple data proxy nodes. Each node corresponds to a type of data source. It adopts a zero-trust security architecture, collects only authorized fields, and performs anonymization locally.

[0118] The multimodal feature extraction and semantic alignment module has a built-in dedicated feature extractor and a shared projection matrix, supporting online learning and model updates to ensure the timeliness and consistency of feature representation.

[0119] The dynamic cause-effect graph construction module has a built-in cause-effect discovery engine that performs a full graph scan every 24 hours, covering new entities and events added in the past 30 days, and updating the graph topology and edge weights.

[0120] The time-aware graph neural network processing module is deployed on a distributed computing cluster. It adopts a graph partitioning and pipeline parallelism strategy, supports real-time inference on dynamic causal graphs with a scale of more than 100,000 nodes, and the latency of a single inference is no more than 500 milliseconds.

[0121] The potential prediction decoding module integrates three sub-networks and an interpretability generator to output structured prediction results and natural language interpretation summaries.

[0122] The tiered early warning trigger module connects to the organization's internal workflow system, automatically distributes early warning signals to the relevant responsible persons, and records the processing feedback loop.

[0123] This embodiment, through the above-described method and system, achieves early, accurate, and causally explainable prediction of the potential for the transformation of scientific and technological achievements, solving two major technical challenges: signal lag and misjudgment of linear correlation.

Claims

1. A method for predicting the potential for technology transfer based on multi-source data fusion, characterized in that, include: Collect raw data streams of non-public early scientific research activities. The raw data streams include research project application texts, laboratory equipment usage logs, internal collaborative communication records of researchers, academic conference preprint submission records, and metadata of technology prototype demonstration videos. The original data stream is subjected to multimodal feature extraction and semantic alignment processing to generate a standardized multi-source feature vector sequence; A dynamic causal graph structure is constructed, wherein the dynamic causal graph structure uses scientific research entities as nodes and cross-modal event triggering relationships as directed edges, and the edge weights are jointly determined by Granger causality test and counterfactual reasoning. The multi-source feature vector sequence is input into a time-aware graph attention network. At each time step, the network aggregates neighborhood node information according to the dynamic causal graph structure and outputs a hidden layer embedding vector representing the current state of the scientific and technological achievement. Based on the hidden layer embedding vector, a predicted value of the potential for the transformation of scientific and technological achievements is generated through the decoder module. The predicted value includes three quantitative indicators: technology maturity index, market adaptability score, and industrialization barrier coefficient. Based on the predicted value, a graded early warning mechanism is triggered. When the industrialization barrier coefficient is lower than a preset threshold and the technology maturity index is higher than a preset threshold, a high-potential scientific and technological achievement identification signal is generated.

2. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 1, characterized in that, Collect raw data streams from non-public early research activities, including: The research project application text is obtained through the interface of the scientific research management information system. The text includes a description of the technical route, the expected form of the results, and information on the team composition. The laboratory IoT gateway collects device usage logs, which record device type, usage duration, operation frequency, and parameter configuration change sequence. The internal collaborative communication records of researchers are intercepted by an encrypted communication proxy module. After the records are de-identified, the co-occurrence frequency of keywords and the evolution path of topics are retained. Access to preprint submission records via academic conference digital infrastructure, including submission timestamps, subject classification tags, and reference citation maps; Metadata from the technology prototype demonstration video is extracted using a video analytics server. This metadata includes action recognition tag sequences, interface interaction heatmaps, and sentiment polarity scores of the speech-to-text transcription.

3. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 2, characterized in that, Multimodal feature extraction and semantic alignment processing are performed on the original data stream, including: Adaptive word embedding is performed on the text of scientific research project application, and sentence vectors are generated by a Chinese character-level convolutional neural network pre-trained on a corpus of science and technology policies. The laboratory equipment usage logs are sliced ​​into time windows, and the equipment activity entropy value and operation mode transition matrix are calculated within each window. A dynamic topic model is constructed from internal collaborative communication records to extract the topic distribution vector and the transition probability between topics within each 7-day cycle. Analyze the reference citation map of preprint submission records, calculate the node centrality index and generate citation influence decay curves; A multi-channel temporal feature tensor is constructed by fusing action recognition tag sequences and speech emotion polarity scores with metadata from the technical prototype demonstration video. The feature vectors of each modality are mapped to a unified dimension through a shared latent space projection matrix and aligned by timestamps to form a multi-source feature vector sequence.

4. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 3, characterized in that, Constructing a dynamic cause-effect graph structure includes: Define a set of scientific research entities, which includes five types of nodes: researchers, research institutions, technical keywords, equipment types, and project numbers. Perform Granger causality test on the event sequence between any two types of nodes. If the null hypothesis of no causality is rejected at a significance level of 0.05, then a directed edge is established. Further counterfactual reasoning verification is performed on the established directed edges. The change in the probability of the target node event is observed after the source node event is removed by intervention simulation. If the absolute value of the change is greater than the preset sensitivity threshold, the directed edge is retained. Each retained directed edge is assigned a weight equal to the weighted sum of the Granger causality F-statistic and the counterfactual effect, and the weights are updated every 24 hours.

5. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 4, characterized in that, The time-aware graph attention network includes an input layer, a multi-head graph attention layer, a time-gated recurrent unit layer, and an output layer. In the input layer, the multi-source feature vector sequence is input together with the adjacency matrix of the dynamic causal graph structure; In the multi-head graph attention layer, each attention head independently calculates the importance coefficient between nodes. The importance coefficient is determined by the node feature similarity and edge weight, and the formula is: in , For nodes , eigenvectors, For learnable weight matrix, For attention vectors, For edge weights, This is the adjustment coefficient; In the temporal gated recurrent unit layer, the aggregated feature sequence output by the multi-head graph attention layer is input into the gated recurrent unit. This unit contains update gates and reset gates to control the retention and forgetting of historical state information. In the output layer, a fully connected transformation is applied to the hidden state at the final moment of the gated recurrent unit to generate the hidden layer embedding vector.

6. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 5, characterized in that, Based on the hidden layer embedding vector, a predicted value for the potential for technology transfer is generated through the decoder module, including: The hidden layer embedding vector is received through the technology maturity index subnetwork, and the output is a continuous value ranging from 0 to 10. Its loss function is the sum of mean squared error and level constraint regularization term. External macroeconomic indicators are introduced as auxiliary inputs through a market fit scoring subnetwork. The hidden layer embedding vector and macroeconomic indicators are fused through a cross-attention mechanism to output a 100-point score ranging from 0 to 100. By employing a quantile regression structure through the industrialization barrier coefficient subnetwork, the estimated barrier coefficients at the 15th, 50th, and 85th quantiles are output to characterize the prediction uncertainty.

7. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 6, characterized in that, The tiered early warning mechanism is triggered based on the predicted value, including: When the technology maturity index is greater than 6.5 and the industrialization barrier coefficient is less than 0.3, a level one warning is triggered, a list of high-potential scientific and technological achievements is generated and pushed to the technology transfer office; When the technology maturity index is between 5.0 and 6.5 and the industrialization barrier coefficient is between 0.3 and 0.5, a level-two warning is triggered, and the expert review process is initiated. When the technology maturity index is less than 5.0 or the industrialization barrier coefficient is greater than 0.5, it is classified into the regular monitoring queue and no active intervention is triggered.

8. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 7, characterized in that, Domain-adaptive word embedding is applied to research project application texts, including: Using characters as the basic input unit, it captures local semantic combinations through multi-layer convolution and pooling operations, and outputs sentence vectors of fixed dimensions, thus avoiding semantic drift in sparse professional terminology scenarios.

9. The method for predicting the potential for technology transfer based on multi-source data fusion according to claim 8, characterized in that, Time window slicing of laboratory equipment usage logs, including: Calculate device activity entropy value within each time window ,in For the first The frequency of occurrence of each operation mode within the window; Construct the operation mode transition matrix ,in Indicates from pattern Transfer to mode The probability, and for After performing singular value decomposition, the first k principal components are taken as feature vectors.

10. A technology transfer potential prediction system based on multi-source data fusion, characterized in that, include: The non-public early research activity data collection module is used to collect research project application texts, laboratory equipment usage logs, internal collaborative communication records of researchers, academic conference preprint submission records, and metadata of technology prototype demonstration videos. The multimodal feature extraction and semantic alignment module is used to perform multimodal feature extraction and semantic alignment processing on the original data stream to generate a standardized multi-source feature vector sequence; The dynamic causal graph construction module is used to construct a dynamic causal graph structure with scientific research entities as nodes and cross-modal event triggering relationships as directed edges. The edge weights are jointly determined by Granger causality testing and counterfactual reasoning. The time-aware graph neural network processing module is used to input the multi-source feature vector sequence into a time-aware graph attention network and output a hidden layer embedding vector representing the current state of the scientific and technological achievement. The potential prediction decoding module is used to generate a predicted value of the potential for the transformation of scientific and technological achievements, which includes a technology maturity index, a market adaptability score, and an industrialization barrier coefficient, based on the hidden layer embedding vector through the decoder module. The graded early warning triggering module is used to trigger the graded early warning mechanism based on the predicted value and generate a high-potential scientific and technological achievement identification signal.

Images