A technical forecasting method, system, and computing device

Abstract

The application discloses a kind of technical prediction method, system and computing device, it is related to computer technology and technical prediction field, including: constructing target technology semantic space;Reconstruction technology enrichment and blank area;Design technology evolution network, construct the technology evolution prediction model of coupling evolution trend state analysis;Realize trend prediction and evaluation.The application realizes all-around, full automation, strong explainable technology prediction by target technology deep semantic representation, reconstruction, evolution trend evolution analysis and dynamics modeling with technology evolution logic, solves the problems that existing method is difficult in field technology semantic understanding, prediction data is scarce and not easy to obtain, technical trend and breakthrough prediction accuracy is low, can comprehensively support technology roadmap planning, technology layout analysis and research and development strategic decision-making etc. Scene.

Description

Technical Field

[0001] This invention relates to the field of computer technology and technology prediction, specifically to the field of technology prediction based on technical text semantic analysis and technology evolution network analysis. Background Technology

[0002] Technology forecasting analyzes technical information such as patents and scientific papers to identify technological development trends and potential opportunities, and can be used for technology roadmap planning, technology layout analysis, and R&D strategic decision-making. As technical textual data continues to grow, technology forecasting is gradually evolving from qualitative analysis methods relying on expert experience to data-driven approaches based on text mining, machine learning, and network analysis.

[0003] Existing technology prediction methods often employ bibliometrics, keyword statistics, or topic models to identify target technologies. However, methods based on word frequency statistics or bag-of-words models struggle to fully represent the contextual semantics of technical texts. Their ability to handle synonyms, near-synonyms, variations of technical terms, and the sparsity of short texts is limited, easily affecting the accuracy of target technology term identification and technical semantic space construction. In identifying technological hotspots and opportunities, existing methods typically use dimensionality reduction, clustering, and density analysis to display technology distribution and identify potential technological gaps based on low-density regions. Some dimensionality reduction methods lack an effective mechanism for restoring the low-dimensional space to a high-dimensional technology feature space, making it difficult to reconstruct technology-rich or technology-gap regions in the low-dimensional space into corresponding high-dimensional technical features and semantics. Consequently, the identification results are difficult to directly use for the analysis and interpretation of potential technological directions.

[0004] In predicting the state of technological evolution, existing methods typically construct time series based on the number of technical documents, topic intensity, keyword weight, or growth rate, and employ life cycle models, Markov models, or Hidden Markov models for trend extrapolation. Some methods require pre-dividing continuous technological features into discrete states, which can easily lead to the loss of some semantic information and trend features. When predicting solely based on historical state transition probabilities, it is also difficult to simultaneously utilize technological semantic features, technological state features, and technological evolution rules, potentially resulting in unstable state changes, premature convergence of prediction paths, or a lack of technological logical explanation in the prediction results.

[0005] Furthermore, existing technology trend prediction methods typically analyze single technical terms or themes, neglecting the relationships, dependencies, and state propagation effects between different technology nodes. Existing technology evolution network analysis primarily focuses on node relationship identification, network structure statistics, or static visualization, often failing to dynamically determine directed edge weights based on node states and relationships across different time windows. It also fails to ensure these directed edge weights actually participate in the training, parameter adjustment, and node state updates of the dynamic prediction model. Therefore, existing methods struggle to characterize the directional impact of a technology node's state change on other related technology nodes, and are also inadequate for propagation analysis of technology trends within evolutionary networks and for predicting future trends at the node level. In the absence of quantitative data such as candidate technical terms, state probabilities, evolutionary paths, and network analysis results, directly generating technology prediction conclusions may suffer from insufficient stability, unclear interpretive basis, and difficulty in verifying the results.

[0006] Therefore, existing technologies still need to further improve the synergy of target technology semantic representation, semantic interpretation of technology-rich regions and technology-blank regions, prediction of technology evolution status, and analysis of technology node correlation and propagation, in order to achieve interpretable prediction of technology status, technology evolution path, and technology node trend changes. Summary of the Invention

[0007] In view of the shortcomings of existing technology prediction methods in terms of technology semantic representation, technology gap semantic interpretation, technology evolution state prediction, and technology node association and propagation analysis, this invention provides a technology prediction method, system, and computing device. Through the construction of target technology semantic space, identification of technology-rich areas and technology gap areas, analysis of technology evolution state, and prediction of technology evolution network dynamics, it can predict existing technology hotspots, potential technology opportunities, technology evolution paths, and technology node trends.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution.

[0009] In a first aspect, the present invention provides a technology prediction method, comprising: constructing a target technology semantic space; defining technology-rich areas and technology-blank areas; designing a technology evolution network and constructing a technology evolution prediction model coupled with evolutionary trend state analysis; performing evolutionary dynamics analysis based on the technology evolution network; and realizing technology trend prediction and evaluation based on the target technology characteristics, technology evolution state, technology evolution path, and technology evolution network analysis results.

[0010] As a preferred embodiment of the technology prediction method of the present invention, the construction of the target technology semantic space includes parsing, cleaning and characterizing patents, scientific papers, technical reports or other technical texts in the target technology field, extracting technical semantic features and forming a high-dimensional feature representation space; and filtering a set of target technical terms based on the technical semantic features.

[0011] Furthermore, the target technical terms are comprehensively screened based on at least one of the following: thesaurus matching results, terminology degree of technical terms, domain relevance, document coverage, time change characteristics, and semantic representativeness. Preferably, the above features of different technical terms are normalized and then a comprehensive evaluation value is calculated, and the target technical term set is determined according to a preset screening threshold or ranking result.

[0012] Furthermore, for synonyms, near-synonyms, abbreviations, variations of technical terms, and terms with the same or similar technical meanings in different languages, a correspondence is established through domain thesaurus matching and semantic vector similarity; when the semantic similarity between two technical terms reaches a preset threshold, they are merged or a term association is established to reduce the impact of different expressions on the construction results of the technical semantic space.

[0013] As a preferred embodiment of the technology prediction method of the present invention, the high-dimensional technology features are mapped to a low-dimensional technology distribution space, a technology enrichment area is determined based on the local density of technology points in the low-dimensional technology distribution space, and a technology blank area is determined based on the low-density area and the spatial relationship between the low-density area and the adjacent technology enrichment area.

[0014] Furthermore, a generative topology mapping model is preferably used to perform low-dimensional mapping on high-dimensional technical features. The high-dimensional technical features of the target technical terms or technical text are used as input to the generative topology mapping model, and the distribution position of technical points in the two-dimensional topology space is obtained through the probabilistic mapping relationship between potential nodes and high-dimensional technical features.

[0015] Furthermore, a kernel density estimation method is preferably used to calculate the local density of technology points in the two-dimensional technology distribution space. For any target location in the two-dimensional space, the kernel density value is calculated based on the distance between that location and its neighboring technology points. Regions with a density higher than a first density threshold and a continuous distribution of technology points are identified as technology-rich areas, while regions with a density lower than a second density threshold, located within the existing technology distribution range, and adjacent to one or more technology-rich areas are identified as technology-blank candidate areas. The first and second density thresholds can be determined based on the quantiles of the overall density distribution or the density distribution of historical technology data.

[0016] As a preferred embodiment of the technology prediction method of the present invention, the method includes: obtaining the low-dimensional topological spatial coordinates corresponding to the technology-rich area and the technology-blank area, and establishing an inverse mapping relationship between the low-dimensional technology distribution coordinates and the high-dimensional technology semantic features, converting the target low-dimensional coordinates into the corresponding high-dimensional reconstructed technology features; and obtaining representative technology words, candidate technology word combinations, or potential technology directions based on the similarity between the high-dimensional reconstructed technology features and the candidate technology word features.

[0017] Furthermore, a regression model with regularization constraints is preferably used to establish the inverse mapping relationship. The low-dimensional topological coordinates of known technical terms are used as model input, and the corresponding high-dimensional technical features are used as model output for training. The regularization coefficient of the regression model is determined based on the cross-validation error, enabling the model to maintain the correspondence between low-dimensional coordinates and high-dimensional technical features while reducing overfitting.

[0018] Furthermore, the cosine similarity between the high-dimensional reconstructed technical features corresponding to the technical gap area and the technical term features in the target technical term library is calculated, and a preset number of candidate technical terms are selected according to the similarity from high to low; general terms, cross-domain terms and ambiguous terms are filtered through domain term library, technical classification information or contextual semantics to generate potential technical semantic information corresponding to the technical gap area.

[0019] As the design technology evolution network described in this invention, a technology evolution network node, node characteristics, and relationship weight strategy based on trend labels are designed;

[0020] The technology evolution network nodes, node features, and relationship weight strategies include, but are not limited to, target technology node screening, node hidden state set, observation feature set, emission probability model, and inter-node relationship weight design method.

[0021] As a preferred embodiment of the technology prediction method of the present invention, a continuous technology observation sequence is constructed based on at least one of the feature word weights, theme strengths, technology activity, document coverage or growth rates of target technology terms, technology themes or technology branches in different time windows, and a set of latent states of technology evolution is established based on the preset technology system evolution trend.

[0022] Furthermore, the preset technological system evolution trend preferably adopts the TRIZ technology system evolution trend. The technological evolutionary hidden states include at least one of the following: increased system completeness, increased system coordination, enhanced dynamism, increased controllability, flow optimization, transition to a supersystem, reduced human intervention, increased value, and increased tailoring. The TRIZ technology system evolution trend is used to determine the technological meaning corresponding to different hidden states and to explain the evolutionary transition relationships between hidden states.

[0023] As a preferred embodiment of the technology prediction method described in this invention, the continuous technology observation sequence is input into a continuous hidden Markov model, the hidden state probability distribution in each time window is determined based on the initial state probability, state transition probability, and the observation feature distribution under different hidden states, and the technology evolution state and technology evolution path in future time windows are deduced based on the state transition relationship.

[0024] Furthermore, it is preferable to set time weights based on the time distance between historical time windows and the target prediction time window. For the t-th historical time window, its time weight can be expressed as:

[0025]

[0026] Where T represents the current prediction baseline time window, T−t represents the time distance between the historical time window and the current prediction time window, and λ is the time decay coefficient. Historical data closer to the prediction time window has a greater time weight, while historical data further away has a smaller time weight. The time decay coefficient λ is preferably determined through rolling verification of historical time series, with the parameter corresponding to a smaller prediction error or state identification error being the preferred parameter.

[0027] Furthermore, during the parameter training process of the continuous hidden Markov model, the time weights are applied to the parameter estimation process of observation data in different time windows, so that more recent technological data has a greater impact on the estimation of observation distribution parameters and state transition parameters, thereby improving the model's ability to respond to recent technological evolution characteristics.

[0028] Furthermore, it is preferable to adjust the hidden state transition probabilities by enhancing the diagonal elements in the state transition matrix. For the trained state transition matrix... The state transition probability is adjusted according to the following formula:

[0029]

[0030] in, Let represent the original probability of transitioning from the i-th hidden state to the j-th hidden state. Represents the self-state enhancement coefficient. For indicator functions; when The value is 1 if the condition is met, and 0 otherwise. By adding diagonal elements to the state transition matrix and re-normalizing the rows, the probability of the current hidden state continuing to maintain that state in the next time window is increased.

[0031] Furthermore, the self-state enhancement coefficient Determined through rolling verification based on historical state sequences. When the data is too small, the model is more sensitive to short-term data fluctuations and is prone to frequent switching of hidden states; when... When the value is too large, the hidden states may excessively maintain their original state, thus reducing their responsiveness to real technological changes. Therefore, it is preferable to select the self-state enhancement coefficient based on a joint evaluation criterion of state prediction error and state switching frequency, so as to preserve the responsiveness to significant technological changes while maintaining the continuity of technological state evolution.

[0032] In a preferred embodiment of the technology prediction method described in this invention, the target technology, target technical term, technical theme, technical branch, or technical theme cluster formed by multiple semantically similar technical terms are used as nodes in the technology evolution network. Node relationships are established based on at least one of the following: semantic association relationship, technical document co-occurrence relationship, time-leading relationship, and state transmission relationship between nodes.

[0033] Furthermore, the node status corresponding to each technical node is determined based on the technical text and technical characteristics in different time windows. The node status is represented by at least one of the following: technical node proportion, node activity, technical theme intensity, probability of technical evolution state, technical growth rate, or comprehensive node status value.

[0034] Furthermore, it is preferable to perform non-negative normalization on the status of each technical node within the same time window, so that the status of each node is on a uniform numerical scale; when the status of a technical node is represented by a node percentage, the sum of the status values ​​of all technical nodes within the same time window is made to be 1, so as to compare the relative changes of different technical nodes in different time windows.

[0035] As a preferred embodiment of the technology prediction method described in this invention, the directed edge weights are determined based on at least one of the state transmission relationship, time lead relationship, technology co-occurrence relationship and semantic association relationship between the source technology node and the target technology node, and a directed edge weight matrix is ​​formed by the directed edge weights between each technology node.

[0036] Furthermore, it is preferable to calculate the state propagation weight, time lead weight, and technology co-occurrence weight separately. The state propagation weight is determined based on the probability distribution of the hidden states corresponding to the two technology nodes, the co-occurrence relationship of the hidden states, and the state transition relationship; the time lead weight is determined based on the directional correlation between the state change of the source technology node in the previous time window and the state change of the target technology node in the subsequent time window; the technology co-occurrence weight is determined based on the intensity of co-occurrence of the corresponding technical terms of the two technology nodes in the same technical document, the same time window, or the same technical topic.

[0037] Furthermore, the weight of the directed edge from the i-th technical node to the j-th technical node can be expressed as:

[0038]

[0039] in, For state propagation weights, Prioritize time. For the co-occurrence weight of technologies, , and The corresponding fusion parameters are preferably satisfied as follows:

[0040]

[0041] The fusion parameters are determined based on the node trend prediction errors within the historical time window. Preferably, multiple candidate fusion parameter combinations are set, and corresponding directed edge weight matrices are established for each combination. Historical rolling predictions are then performed. The root mean square error, mean absolute error, or a combination of both of these evaluation indicators are used as the parameter selection criteria, and the parameter combination with the smaller prediction error is selected as the directed edge weight fusion parameter.

[0042] As a preferred embodiment of the technology prediction method described in this invention, the historical node state sequence, node relationships, and directed edge weight matrix are input into the technology evolution dynamics prediction model. Based on the error between the actual node state and the predicted node state, at least one of the node self-change parameters, network transmission parameters, state update parameters, and directed edge weight fusion parameters is trained or optimized.

[0043] Furthermore, the node state change is preferably determined by combining the node's recent changes, historical average state, and the propagation effect of the technology evolution network. The technology evolution network propagation effect is calculated based on the directed edge weight matrix and the node state within the current time window, and is used to represent the net state transmission impact received by the target node from other associated nodes.

[0044] A preferred form of node state update is as follows:

[0045]

[0046] in, This represents the node state vector within the current time window. Indicates the recent trend of node changes. This represents the average state of historical nodes. This indicates the net propagation effect generated by the technology evolution network. , and These represent the model parameters corresponding to recent changes, historical state adjustments, and network propagation effects, respectively.

[0047] The predicted node state in the next time window is updated based on the change in node state, and the prediction result is subject to non-negative constraints. When the node state is represented by the node percentage, the prediction result is further normalized so that the sum of the predicted states of all technical nodes is 1.

[0048] Furthermore, it is preferable to use a rolling time validation method to determine the model parameters, that is, to train the model using continuous historical data before the target prediction time window, predict the node state in the next time window, and compare the prediction results with the corresponding historical actual node states; to move the training window in chronological order and repeat the above process, and to select the model parameters based on the average prediction error in multiple time windows.

[0049] As a preferred embodiment of the technical prediction method described in this invention, the node trend is identified based on the node state difference between the prediction time window and the baseline time window.

[0050] Furthermore, a trend determination threshold is preferably set based on the degree of fluctuation in the historical changes of nodes. For the i-th technical node, the standard deviation is calculated based on the changes in the node state in adjacent historical time windows. and with As a threshold for trend determination, A preset proportional coefficient is used. When the predicted change in node state is greater than the trend determination threshold, the corresponding technical node is identified as having an upward trend; when the predicted change in node state is less than the negative trend determination threshold, the corresponding technical node is identified as having a downward trend; when the change in node state is between the two thresholds, the corresponding technical node is identified as having a relatively stable trend. This allows technical nodes with different fluctuation levels to have trend determination criteria adapted to their historical change characteristics.

[0051] Furthermore, node state prediction can be re-executed by deleting, reducing, or increasing the weight of specified directed edges, and the target node state prediction values ​​before and after adjustment can be compared to evaluate the degree of influence of different source technology nodes and directed edges on the trend changes of the target node and determine the evolution relationship of the target technology.

[0052] As a preferred embodiment of the technology prediction method of the present invention, structured technology evolution information is formed based on at least one of candidate technology feature words, technology semantic information, technology evolution state probability, predicted evolution path, technology node state, node trend and technology evolution network relationship.

[0053] Furthermore, the structured technology evolution information preferably includes at least one of the following: target technology name, candidate technology terms, current technology evolution state, future technology evolution state, state transition probability, current state of technology nodes, predicted state of technology nodes, direction of node change, and key related technology nodes. A technology evolution reasoning chain is formed based on the structured technology evolution information to generate technology prediction results or potential technology solutions.

[0054] Furthermore, at least one of the following methods—large model, intelligent agent, or manual evaluation—is used to evaluate the technology prediction results. Preferably, a comprehensive evaluation is conducted based on at least one of the following: consistency of technology semantics, consistency of evolutionary state, rationality of technology evolution path, and consistency of node trends, in order to reduce the possibility of inconsistencies between a single generated result and the aforementioned quantitative analysis results.

[0055] Secondly, the present invention provides a technology prediction system, including a semantic space construction module, a region identification and decoding module, a technology evolution state analysis module, a technology evolution network construction module, a dynamic prediction module, and a prediction generation and evaluation module.

[0056] The semantic space construction module is used to acquire technical text data of the target technology field, extract technical semantic features, and construct the target technology semantic space; the region identification and decoding module is used to identify technology-rich areas and technology-blank areas, and perform semantic inverse reconstruction of the target region; the technology evolution state analysis module is used to construct a continuous technology observation sequence and predict the technology evolution state and evolution path; the technology evolution network construction module is used to determine technology nodes, node relationships, and directed edge weights; the dynamics prediction module is used to train or optimize the technology evolution dynamics prediction model based on historical node states and directed edge weight matrices, and predict node trend changes; the prediction generation and evaluation module is used to generate and evaluate technology prediction results based on the above results.

[0057] Thirdly, the present invention provides a computing device, including a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the above-mentioned technical prediction method.

[0058] Compared with existing technologies, this invention integrates the construction of the target technology semantic space, the identification of technology-rich and technology-blank areas, the inverse reconstruction of technology region semantics, the deduction of technology evolution states, and the analysis of technology evolution network dynamics into a unified technology prediction process. On the one hand, through the inverse mapping between low-dimensional technology distribution space and high-dimensional technology semantic features, technology-blank locations are transformed into candidate technology directions with clear technological semantics. On the other hand, through time-weighted continuous hidden Markov models and self-state transition probability adjustment, the continuity of technology evolution states in the time series and the responsiveness to recent technological changes are improved. Simultaneously, through the directed edge weights between technology nodes, the directional transmission of technology states participates in the training of the dynamic model and the updating of node states, thereby realizing the prediction of technology evolution paths and node-level future trends, and improving the interpretability of technology prediction results. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a flowchart of the comprehensive technology prediction framework of the present invention;

[0061] Figure 2 A schematic diagram illustrating the mathematical principles and process of the blank decoding mechanism based on manifold-inverse mapping technology;

[0062] Figure 3 This is a technology enrichment distribution diagram in the field of power batteries in the embodiments of the present invention;

[0063] Figure 4 This is a detection diagram of the technological gap in the field of power batteries in this embodiment of the invention;

[0064] Figure 5 The topological structure diagram of the continuous hidden Markov model (TRIZ-CHMM) coupled with the TRIZ rule;

[0065] Figure 6 A time-series evolution prediction curve for multidimensional technological evolution trends;

[0066] Figure 7 Sankey diagram for the intertemporal transfer of the hidden state of core technologies.

[0067] Figure 8 This is a comparison chart of prediction errors of different technology evolution prediction models in the embodiments of the present invention;

[0068] Figure 9This is a comparison diagram of the status of various technical nodes in the field of power batteries during the baseline period and the predicted period in an embodiment of the present invention. Detailed Implementation

[0069] To make the technical solution of this invention clearer, the specific implementation process is described below using the field of power batteries as an example. The following embodiments are used to illustrate the technical solution of this invention and are not intended to limit the scope of protection of this invention.

[0070] Example 1: Construction of target technology semantic space and identification of technology-rich and blank areas

[0071] Historical patent texts in the field of power batteries were collected, using textual information from patent titles, abstracts, claims, and specifications as the basic data. The original patent data underwent deduplication, character cleaning, format standardization, publication year extraction, and patent identifier matching. Technical terms were also segmented, standardized, and semantically encoded.

[0072] like Figure 1 As shown, semantic encoding and keyword extraction are performed on the processed technical text. Specifically, SBERT can be used to obtain high-dimensional semantic vectors of the technical text, and target technical terms can be selected based on TF-IDF, domain relevance, document coverage, and time change features, thereby establishing a target technical terminology library and a high-dimensional technical semantic space.

[0073] The technical semantic vectors are aggregated according to the patent publication year to obtain technical feature representations for different time windows, and time series data are formed for subsequent evolutionary state analysis.

[0074] like Figure 2 As shown, after preprocessing the high-dimensional technical features, a generative topological mapping model is used to map them to a two-dimensional latent space. The generative topological mapping model forms a regular latent grid in the low-dimensional space, and the location of the technical object in the two-dimensional space is determined by calculating the probability distribution of the technical samples near each grid.

[0075] The kernel density estimation method is used to calculate the local technology density in two-dimensional space. For example... Figure 3 As shown, areas with high local density and concentrated distribution of technology samples are identified as technology-rich areas, used to represent technology themes that have achieved a certain level of technological accumulation; for example... Figure 4 As shown, regions with low local density within the existing technology distribution range are identified as candidate technology gaps.

[0076] For candidate regions lacking technical expertise, Ridge regression is used to establish an inverse mapping relationship between low-dimensional grid coordinates and high-dimensional technical features. The low-dimensional coordinates of the candidate regions are input into the inverse mapping model to obtain the corresponding high-dimensional reconstructed features. Then, the semantic similarity between the high-dimensional reconstructed features and the target technical word vectors is calculated to obtain candidate technical words.

[0077] After filtering general mathematical terms, cross-domain terms, and ambiguous terms in the output technical terms, candidate technical terms related to power battery structure, cell simulation, energy efficiency optimization, group control, and intelligent diagnosis are retained, thereby transforming the blank areas in the two-dimensional technical space into potential technical directions that can be directly explained.

[0078] Example 2: Technological Evolution Analysis Based on the TRIZ-CHMM Model

[0079] Based on the target technical terms and their time attributes in the field of power batteries obtained in Example 1, the weight, theme intensity, technical document coverage, technical activity and growth rate of each technical term are statistically analyzed according to the year, and the features of different dimensions are standardized to form the continuous technical observation vector and continuous technical observation sequence corresponding to each year.

[0080] According to the technology evolution state model described in claim 5, a set of hidden states is established using the TRIZ technology system evolution trend as the trend label. This embodiment sets nine types of technology evolution states: increased system completeness, increased system coordination, enhanced dynamism, increased controllability, flow optimization, transition to a supersystem, reduced human intervention, increased value, and increased tailoring. Continuous technological characteristics from different years in the power battery field are used as the set of observed characteristics. Figure 5 As shown, the annual observation characteristics are established with the nine types of latent states of technological evolution through the emission probability model, thus forming a technological evolution state network based on time series.

[0081] The aforementioned continuous observation sequence is input into TRIZ-CHMM for training. In this embodiment, a continuous probability distribution is used to describe the observation feature distribution corresponding to each hidden state, and the initial state probability, emission probability parameters, and state transition matrix are estimated using the Baum-Welch algorithm. Based on the trained model, the hidden state probability distribution corresponding to each year is calculated, and the Viterbi algorithm is used to obtain the technology evolution state sequence with higher probability.

[0082] During model training, observation data from different years are weighted according to the time-weighting method described in the invention, giving higher weights to technical data closer to the prediction time. The time decay parameter is determined through historical time window rolling verification, i.e., candidate parameters are used for model training and state prediction respectively, and the corresponding parameters are selected based on the prediction error. For the state transition matrix obtained from training, the diagonal state probability is increased according to the aforementioned self-state transition probability adjustment method and then re-normalized; the self-state enhancement parameter is also selected based on historical prediction error and state switching frequency to reduce frequent state jumps caused by short-term data fluctuations.

[0083] After completing model training, the probability and technological strength of the evolution states of various TRIZ technologies in different years in the field of power batteries are obtained. For example... Figure 6 As shown, different technological evolution states exhibit different enhancements, weakenings, and phased changes in the time series, thereby identifying the main evolutionary trends of the target technology at different development stages.

[0084] Further state deduction is performed for subsequent time windows based on the state transition matrix and the current hidden state probability. For example... Figure 7 As shown, a Sankey diagram of intertemporal transitions is formed by identifying the main hidden states and their state transition relationships within adjacent time windows. The connection direction represents the transition direction of the technological evolution state, and the connection width represents the transition intensity of the corresponding state. This yields the main evolutionary path of the target technology for power batteries from its current technological evolution state to subsequent technological evolution states, and this result is used as the state characteristic input for subsequent technological evolution networks and dynamic predictions.

[0085] Example 3: Construction of Technology Evolution Network and Directed Edge Weights

[0086] Based on the results of Examples 1 and 2, target technical terms with similar semantics, similar technical functions, and similar temporal change characteristics are aggregated. This example aggregates 38 modeling technical terms into six technical nodes, denoted as follows:

[0087] T01, Electrode Structure and Phase Transition;

[0088] T02, Mechanism Analysis and Performance Modeling;

[0089] T03, Cell Simulation and Energy Efficiency Optimization;

[0090] T04, Group Integration and Network Control;

[0091] T05, intelligent diagnostics and low power consumption sensing;

[0092] T06, Mathematical Modeling and Optimization Analysis.

[0093] We statistically analyzed the technical term weights, patent counts, and technical theme intensity of each technology node in different time windows, and then normalized them to obtain the annual status value of each technology node.

[0094] Further calculations are performed on the technology co-occurrence relationship, time-leading relationship, and CHMM state transmission relationship between nodes. Specifically, the technology co-occurrence relationship is calculated based on the intensity of co-occurrence of corresponding technical terms of two nodes within the same patent or the same time window; the time-leading relationship is calculated based on the changes of the source node in the previous time window and the changes of the target node in the subsequent time window; and the state transmission relationship is determined based on the TRIZ technology evolution state probabilities and state transition relationships of the two nodes.

[0095] According to the directed edge weight calculation method described in the invention, the state propagation weight, time lead weight, and technology co-occurrence weight are normalized and fused to obtain the directed edge weights from the source node to the target node, and the directed edge weights between the six technology nodes constitute the directed edge weight matrix of the technology evolution network.

[0096] The changes in node state within adjacent time windows are calculated based on the node's historical state sequence, and the network propagation effect is calculated in conjunction with the directed edge weight matrix. Recent changes in nodes, historical average states, and network propagation effects are used together as the basis for node state updates to construct a predictive model for technological evolution dynamics.

[0097] Example 4: Training and Trend Prediction of Technology Evolution Dynamics Model

[0098] Based on the technology evolution network constructed in Example 3, the technology node state sequence, node relationship and directed edge weight matrix in different time windows are used as inputs to the dynamic prediction model to train and predict the future state of technology nodes.

[0099] To determine the node self-change parameters, network propagation parameters, and state correction parameters in the dynamic model, rolling time verification is performed using historical node state sequences. This embodiment divides continuous historical windows according to time sequence. Each time, the model is trained using the node states prior to the prediction time point and the corresponding technical evolution network, and the node states in the next time window are predicted. The prediction results are then compared with the actual node states in the corresponding time window, and the prediction window is moved forward according to time sequence, repeating the model training and prediction process.

[0100] Model prediction errors are expressed as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Evaluation is conducted. Different node state update methods are set for the node's own historical changes, time changes, network propagation effects, and residual correction methods, and the corresponding model parameters are determined by rolling prediction results.

[0101] like Figure 8 As shown, the overall prediction errors differ among different state update methods. Specifically, the RMSE of the safe residual update model is 0.023691, the MAE is 0.018147, and the coefficient of determination R0 is [missing data]. 2 The RMSE of the persistence prediction model is 0.924329; the RMSE of the persistence prediction model is 0.025654. This indicates that, within the historical time window, correcting the node state based on recent node changes and historical prediction residuals can reduce the overall prediction error. Therefore, the corresponding parameters can be used for parameter adjustment in subsequent technology evolution dynamics prediction models.

[0102] After completing historical rolling verification, the state of the technical nodes in the baseline time window is used as the initial state, and the directed edge weight matrix of the complete technology evolution network is introduced into the node state update process. For each technical node, the node state in the next time window is calculated based on its recent state changes, historical state, and network propagation effects received from other related technical nodes. After prediction, the node states are subjected to non-negativity constraints and normalization to obtain the predicted state of each technical node in subsequent time windows.

[0103] Furthermore, the node trend is determined based on the difference in node states between the predicted time window and the baseline time window. When the predicted state significantly increases relative to the baseline state, the corresponding node is identified as having an upward trend; when the predicted state significantly decreases, the corresponding node is identified as having a downward trend; when the change is within a preset trend threshold range, the corresponding node is identified as having a relatively stable trend. The trend threshold can be determined based on the historical state change range of each node or the statistical distribution of state changes in historical adjacent time windows.

[0104] like Figure 9 As shown, using the node status in 2025 as the baseline, the six technology nodes in 2028 are projected. Among them, the T03 cell simulation and energy efficiency optimization node increases from 0.300290 to 0.329548, and the T04 group integration and network control node increases from 0.095232 to 0.110229, showing an upward trend; the T01 electrode structure and phase change node decreases from 0.236134 to 0.208027; the T05 intelligent diagnosis and low power consumption sensing node and the T06 mathematical modeling and optimization analysis node also decrease; the T02 mechanism analysis and performance modeling node changes from 0.232652 to 0.226730, with a relatively small overall change.

[0105] The results show that, within the prediction time window, the relative states of the two technology nodes—cell simulation and energy efficiency optimization, and group integration and network control—have increased, indicating that the activity and network role of these related technologies in the technology evolution network have improved. Conversely, the relative states of some infrastructure, intelligent sensing, and mathematical modeling nodes have decreased. By comparing the predicted states and magnitudes of change of different nodes, a node-level technology trend ranking can be established, providing a quantitative basis for subsequent technology direction predictions.

[0106] To further analyze the sources of trend changes in technology nodes, perturbation analysis can be performed on the directed edges in the technology evolution network. For the target technology node, a specified directed edge is deleted, the corresponding edge weight is reduced, or the corresponding edge weight is increased, while keeping other model conditions unchanged, and the state deduction is repeated; the changes in the predicted state of the target node before and after the perturbation are compared. When the prediction result of the target node changes significantly after deleting or reducing the weight of a certain directed edge, the corresponding source node and directed edge are identified as important technical relationships affecting the trend of the target node.

[0107] For example, for the T03 cell simulation and energy efficiency optimization node, which is predicted to show an upward trend, the weights of the associated edges pointing to T03 can be adjusted, and its predicted state can be recalculated to identify the propagation effect of mechanism analysis, group control, or other associated nodes on the changes in the cell simulation and energy efficiency optimization trend. This not only yields the future trend of the technology node but also identifies the source technology nodes and key evolutionary relationships influencing this trend.

[0108] Therefore, this embodiment achieves continuous deduction from the technology evolution network to the node-level future trend through historical rolling verification, dynamic parameter adjustment, directed network propagation, node state prediction, and key directed edge analysis.

[0109] Example 5: Generation of Potential Technology Directions Based on Technology Evolution Reasoning Chain

[0110] Based on the candidate technical terms, technical evolution status, predicted evolution path, technical node status, directed edge weights and node trend results obtained in Examples 1 to 4, structured technical evolution information is constructed, and potential technical directions are generated accordingly.

[0111] In this embodiment, the T03 cell simulation and energy efficiency optimization node is selected as the target node. The state value of this node increases from 0.300290 in the baseline time window to 0.329548 in 2028, showing an upward trend. The related technical nodes include T02 mechanism analysis and performance modeling and T04 group integration and network control. The state value of the T04 node increases from 0.095232 to 0.110229, also showing an upward trend.

[0112] Based on the technological evolution status and path obtained in Example 2, the following structured technological evolution information is formed: the target technology field is power batteries; the target technology node is cell simulation and energy efficiency optimization; the node trend is upward; the main technical contents include cell simulation, operation status modeling and energy efficiency optimization; the related technical directions include mechanism analysis and performance modeling, group integration and network control; the evolution basis includes technological state transition, node historical changes and technological evolution network propagation results.

[0113] A technology evolution inference chain is formed in the order of "current technical characteristics - technology evolution status - network correlation - node prediction trend - candidate technology direction", and the following technology prediction instructions are generated:

[0114] Based on the above predictions of power battery technology, a potential technology direction with cell simulation and energy efficiency optimization as its core is generated. At the same time, combined with related directions such as mechanism analysis and performance modeling, group integration and network control, the name of the technology direction, its main technical components and implementation process are given. Content unrelated to the field of power batteries or unverified performance indicators must not be generated.

[0115] After inputting the technology prediction instruction into the technology prediction generation model, the following potential technology direction is obtained: a digital twin technology for electro-thermal-aging multiphysics fields for cross-level collaborative optimization of cell-battery pack.

[0116] This technical approach establishes a coupled model of the cell's electrochemical, thermal characteristics, and aging state, and combines it with operational data such as voltage, current, temperature, and state of charge to dynamically correct the model parameters, forming a digital twin model that can reflect the cell's operating state.

[0117] Based on the differences in capacity, internal resistance, temperature, and aging degree of different cells, the cell-level simulation results are transmitted to the battery pack control module to coordinate and adjust the parameters of equalization control, charge and discharge power distribution, thermal management, and energy management, thereby forming a cross-level optimization process that combines cell state prediction, energy loss analysis, and battery pack control.

[0118] Specifically, electrochemical mechanisms and aging modeling correspond to the T02 mechanism analysis and performance modeling node; multiphysics simulation and energy efficiency analysis of the battery cell correspond to the T03 battery cell simulation and energy efficiency optimization node; and equalization control, thermal management, and power coordination correspond to the T04 group integration and network control node. The results demonstrate the correlation between the T02, T03, and T04 technology nodes, and correspond to the upward trend of the T03 and T04 nodes.

[0119] After verifying the consistency of technical semantics, evolution path, and node relationships, the aforementioned technical directions were selected as candidate technical directions generated in this embodiment. The results are jointly constrained by candidate technical terms, technical evolution status, node prediction trends, and technical evolution network relationships; their actual technical effects need further verification through simulation or engineering experiments. This embodiment achieves a seamless connection between technical semantic analysis, technical evolution status prediction, and node trend analysis, leading to the generation of specific potential technical directions.

Claims

1. A technology forecasting method, characterized in that, Includes the following steps: Construct a target technology semantic space, define technology enrichment and gap areas, design a technology evolution network, and build a technology evolution prediction model coupled with evolutionary trend state analysis to achieve trend prediction and evaluation.

2. The technology forecasting method according to claim 1, characterized in that, The construction of the target technology semantic space specifically includes: parsing, cleaning and feature processing a set of technical texts, extracting technical semantic features, screening the target technical word set and forming a high-dimensional feature representation space; The selection of target technical terms is based on the semantic features of the technology, and a comprehensive judgment is made using lexicon matching, terminology degree calculation, and domain features.

3. The technology forecasting method according to claim 1, characterized in that, The definition of technology-rich and technology-blank areas specifically includes: reducing the high-dimensional feature representation space to a low-dimensional distribution to achieve semantic reconstruction of technology-rich and technology-blank areas; The technology enrichment region is determined using the local density enrichment criteria of the technology point distribution area. The aforementioned technology gap area is determined by analyzing the distribution area of ​​technology points that are not covered, thereby further identifying potential target technologies.

4. The technology forecasting method according to claim 3, characterized in that, The semantic reconstruction of the technology-rich region and the technology-blank region includes: designing a low-dimensional topological space coordinate system corresponding to the technology-rich region and the technology-blank region, determining candidate technology feature words, completing the reverse reconstruction and decoding of the semantic connotation of the technology space, and converting it into the corresponding technology word vector. The process involves determining candidate technical feature words and generating corresponding technical semantic information. The semantic connotation is used to characterize existing technological hotspots or potential technological opportunities.

5. The technology forecasting method according to claim 1, characterized in that, The design of the technology evolution network includes: designing trend-label-based technology evolution network nodes, node characteristics, and relationship weighting strategies; The technology evolution network nodes, node features, and relationship weight strategies include, but are not limited to, target technology node screening, node hidden state set, observation feature set, emission probability model, and inter-node relationship weight design method. The trend-label-based technology evolution network constructs a network set consisting of target technologies, a set of target technologies, and trend relationships between target technologies under a time series. The set of hidden states consists of multiple technological evolution states corresponding to the preset technological system evolution trend; The set of observed features consists of at least one of the feature weights, topic strengths, semantic vectors, or growth rates of the target technology in different time windows. The emission probability model is used to characterize the correspondence between the set of observed features and the set of hidden states.

6. The technology prediction method according to claim 1, wherein the construction of the technology evolution prediction model coupled with evolutionary trend state analysis is characterized in that, Based on the observed feature set and the emission probability model, determine the hidden state probability distribution of the target technology in the current time window; Design a state transition matrix to characterize the evolutionary transition relationship between different hidden states, and use it to simulate and predict the temporal state transition process of the target technology; The technology evolution prediction model predicts the technology evolution state and evolution path based on the state transition matrix of the target technology.

7. The technology prediction method according to claim 1 or 6, wherein the technology evolution prediction model coupled with evolutionary trend state analysis is characterized in that, It also includes prediction of technology evolution networks based on evolutionary dynamics analysis; The evolutionary dynamics analysis, based on the nodes, relationships, and directed edge weights in the technological evolution network, trains or optimizes a complex network dynamics prediction model to identify network node trends and predict changes in those trends.

8. The technology forecasting method according to claim 1, characterized in that, The aforementioned trend prediction and evaluation includes: forming a technology evolution reasoning chain based on the target technology characteristics, state probability distribution, technology evolution path, or evolution network analysis results, and generating technology prediction results or potential technology solutions; The assessment can be conducted by comprehensively evaluating existing large models, intelligent agents, and human prediction results.

9. A technology prediction system, characterized in that, include: The semantic space construction module is used to acquire technical text data in the target technology field and construct the semantic space of the target technology. The region identification and decoding module is used to identify technology-rich areas and technology-blank areas, and to generate a set of target technology terms; The model inference module integrates a technology evolution prediction model coupled with evolutionary trend state analysis, and performs technology evolution reasoning; The prediction generation module generates a comprehensive evaluation of the technology prediction results based on the target technology set and the technology deduction results.

10. A computing device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the technical prediction method according to any one of claims 1 to 8.

Images