A Method for Predicting and Early Warning of Technological Trends Based on Big Data Analysis
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-19
- Publication Date
- 2026-08-11
AI Technical Summary
[0002]现有科技趋势预测方法主要存在以下缺陷：一是数据源单一，多仅依赖专利数据，忽略了论文、基金、政策、舆情等蕴含的多维度科技信号，导致预测信息不全面；二是预测与预警割裂，两者独立运行，缺乏信息反馈与联动机制，预测结果无法引导预警，预警信号也无法校正预测模型；三是预测结果可信度不足，缺乏系统性的多层级置信度评估与可解释性分析，决策者难以判断预测的可靠性；四是模型缺乏动态自适应机制，部署后无法根据新数据和预警反馈在线更新，性能随时间衰退；五是预警维度单一，尚未建立面向科技发展特有风险（如技术过度集中、创新泡沫、关键路径断裂等）的多维预警指标体系
[0031]本发明通过融合多模态异构科技大数据构建跨模态知识图谱，克服单一数据源的信息不完备问题，提升预测的全面性与准确性；首创双向协同预测预警神经网络，通过正向先验知识注入将预测结果主动引导预警、反向预警反馈驱动预测模型在线增量更新，打破预测与预警长期割裂的局限，实现从被动异常检测到主动风险预判的升级，并赋予模型持续自适应能力；结合蒙特卡洛方法与特征归因技术构建多层级置信度与可解释性评估体系，使决策者能依据不同置信等级分级采纳预测结果，增强决策透明度；同时针对科技发展特有规律设计多维专属预警指标，填补了科技领域专用预警能力的空白，使预测性能较现有方法获得进一步提升，且系统具备良好的可扩展性与工业实用性。
Smart Images
Figure CN122547950A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data analytics, and in particular to a method for predicting and warning of technological trends based on big data analytics. Background Technology
[0002] Existing methods for predicting science and technology trends suffer from the following main shortcomings: First, they rely on a single data source, often solely on patent data, neglecting multi-dimensional science and technology signals from papers, grants, policies, and public opinion, resulting in incomplete prediction information. Second, prediction and early warning are disconnected, operating independently without information feedback or linkage mechanisms; prediction results cannot guide early warnings, and early warning signals cannot correct prediction models. Third, the reliability of prediction results is insufficient, lacking systematic multi-level confidence assessments and interpretability analysis, making it difficult for decision-makers to judge the reliability of predictions. Fourth, the models lack dynamic adaptive mechanisms, failing to update online based on new data and early warning feedback after deployment, leading to performance degradation over time. Fifth, early warning dimensions are limited, lacking a multi-dimensional early warning indicator system addressing risks specific to science and technology development (such as over-concentration of technology, innovation bubbles, and critical path disruption). Therefore, there is an urgent need for a highly reliable and adaptive method for predicting and warning science and technology trends that integrates multi-source data, facilitates two-way collaboration between prediction and early warning, and is tailored to specific risks. Summary of the Invention
[0003] The main objective of this invention is to provide a method for predicting and warning of technological trends based on big data analysis, which can effectively solve the problems mentioned in the background art.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0005] A method for predicting and warning of technological trends based on big data analysis includes the following steps:
[0006] S1: Simultaneously collect multimodal heterogeneous big data on science and technology from multiple data sources, and perform data cleaning, standardization alignment, and feature extraction;
[0007] S2: Map the multimodal data processed by S1 into a unified knowledge representation framework, construct a cross-modal science and technology knowledge graph, and incrementally update the newly incoming data during the operation;
[0008] S3: Construct a two-way collaborative prediction and early warning network, which consists of a trend prediction sub-network and a risk early warning sub-network, and realizes information interaction and joint optimization between the two sub-networks through a two-way feedback mechanism; wherein, the trend prediction sub-network outputs multi-time-scale technology trend prediction results based on the knowledge graph of S2, and the risk early warning sub-network monitors multi-dimensional early warning indicators in real time and generates multi-level early warning signals.
[0009] S4: Based on time series interpretability and confidence assessment, generate multi-level confidence scores and interpretability reports for each prediction result output by the trend prediction sub-network in S3;
[0010] S5: The risk warning sub-network outputs the warning level and provides corresponding handling suggestions based on real-time monitoring and comprehensive analysis of multi-dimensional warning indicators;
[0011] S6: Based on the bidirectional feedback mechanism in S3 and the early warning signal output by S5, a model continuous update strategy combining incremental learning and periodic retraining is adopted to enable the system to adapt to the dynamic changes in technological development.
[0012] Preferably, the multimodal heterogeneous science and technology big data in S1 includes: patent data, academic paper data, science and technology funding data, industrial policy data, and science and technology media sentiment data. These five types of data respectively reflect different dimensions of science and technology development signals such as technology research and development, theoretical exploration, strategic funding, policy guidance, and public attention, and together constitute a more comprehensive information foundation than single patent data.
[0013] Preferably, the bidirectional feedback mechanism in S3 includes a forward channel and a reverse channel:
[0014] The positive channel injects the technical field popularity prediction and / or technical fusion probability distribution output by the trend prediction subnetwork into the risk warning subnetwork as prior knowledge, which is used to dynamically adjust the threshold of the warning indicators and realize the upgrade from passive anomaly detection to proactive risk prediction.
[0015] When the risk warning subnetwork issues a Level II or higher warning for T consecutive time steps, the reverse channel triggers a local incremental update of the trend prediction subnetwork. The parameters of at least the last layer of the prediction subnetwork are fine-tuned through the knowledge distillation method, so that the prediction model incorporates the warning signal into the consideration of future predictions.
[0016] The forward channel uses the prediction results as prior knowledge to dynamically adjust the warning threshold, achieving an upgrade from passive detection to proactive prediction; the reverse channel triggers local incremental updates to the prediction model when warning signals continue to appear. This two-way information interaction mechanism breaks the traditional model of separating prediction and warning, enabling the two sub-networks to promote each other and jointly optimize.
[0017] Preferably, the multi-dimensional early warning indicators in S5 include at least one of the following: an abnormal speed indicator for technological integration, a risk indicator for the concentration of technological themes, a disruptive impact index, a deviation indicator for the technological life cycle, and a risk indicator for international cooperation; wherein:
[0018] The anomaly index of technology convergence speed is calculated by comparing the co-occurrence growth rate among different technology fields with the degree of deviation of similar technologies from the average growth rate.
[0019] The technology theme concentration risk index is based on the Herfindahl index to calculate the distribution concentration of patents and papers within a research theme area;
[0020] The Disruptive Impact Index quantifies the impact of new technologies on existing knowledge systems by analyzing the rate of change in the reference network topology.
[0021] The technology lifecycle deviation index is calculated by comparing the deviation between actual observed values and the fitted curve of the Logistic growth model;
[0022] International cooperation risk indicators monitor changes in the number of cross-border patent collaborations and the connectivity of cooperation networks in key areas.
[0023] These indicators differ from general warning indicators; they are specifically designed to capture asymmetric risks in the technology sector, such as technological homogenization, innovation bubbles, and path lock-in.
[0024] Preferably, the confidence assessment in S4 employs the Monte Carlo Dropout method: during the prediction phase, the model undergoes K forward propagations, with some neurons randomly masked at a predetermined Dropout rate during each propagation. The prediction mean and variance are calculated based on the statistical distribution of the K prediction results, with the confidence score being 1 − (variance / maximum possible variance). Furthermore, the SHAP method is used to calculate the marginal contribution of each input feature to the prediction result, providing interpretability information in an importance-ranked manner. The multi-level confidence score is divided into three levels: high confidence (≥0.85), medium confidence (0.60~0.85), and low confidence (<0.60). By determining the confidence score and classifying it into high, medium, and low levels through the statistical distribution of multiple forward propagations, decision-makers can adopt prediction results at different levels based on confidence levels, significantly improving the credibility and transparency of the prediction results.
[0025] Preferably, the trend prediction sub-network in S3 specifically includes: a modality alignment encoder, used to map feature vectors extracted from different modalities to a unified semantic space; a temporal graph Transformer encoder, used to capture long-distance temporal dependencies between technology nodes in the knowledge graph using time steps as sequence units; a technology fusion prediction head, used to output the probability of candidate technology pairs forming technology fusion in the next time step; and a topic evolution prediction head, used to predict the trajectory of attention changes for each technology topic within a future time window. These modules work together to map multimodal data to a unified semantic space, capture long-distance temporal dependencies, and simultaneously output technology fusion probabilities and topic evolution trajectories, achieving multi-granularity trend prediction.
[0026] Preferably, the risk warning sub-network in S3 specifically includes: a warning indicator sensor layer, used to extract real-time values of multi-dimensional warning indicators from the knowledge graph in a time window sliding manner; a heterogeneous graph attention warning encoder, used to learn the mutual influence weights and dependencies between different warning indicators; a time-series anomaly detection module, used to model the normal pattern of the indicator sequence based on a long short-term memory autoencoder and determine anomalies through reconstruction errors, while combining control chart technology in statistical process control methods to capture trend anomalies early; and a warning decision fusion module, used to weightedly fuse the output of the graph attention encoder with the anomaly detection results to comprehensively output the warning level. This architecture can dynamically extract real-time indicators, learn the mutual influence between indicators, combine reconstruction errors and statistical process control to determine anomalies, and fuse multi-source information to output a comprehensive warning level.
[0027] Preferably, the cross-modal science and technology knowledge graph in S2 adopts an attribute graph model G=(V, E, A), where V is a set of nodes, including four types: technology nodes, institutional nodes, personnel nodes, and national nodes; E is a set of edges, including reference relationships, collaborative research and development relationships, co-occurrence relationships, and technology inheritance relationships; and A is a set of attributes for nodes and edges, including timestamps, type labels, and confidence metadata. This model uniformly represents entities and their relationships in different modalities, providing a structured knowledge foundation for subsequent graph neural network prediction and early warning.
[0028] Preferably, the warning levels in S5 include: Level I warning / attention warning, Level II warning / alert warning, and Level III warning / emergency warning, and corresponding handling suggestions are provided according to the warning level; wherein, when the Herfindahl-Hirschman Index of the technology theme concentration risk indicator is greater than 0.3 and shows an upward trend for three consecutive months, at least a Level II warning is triggered. Through tiered warnings and corresponding handling suggestions, the system can provide differentiated response strategies according to the severity of the risk, improving the practicality and operability of the warning system.
[0029] Preferably, the model continuous update strategy combining incremental learning and periodic retraining in S6 specifically involves: performing a lightweight incremental update every 7 days, adjusting the model parameters using newly collected data since the most recent update and accumulated warning feedback signals from the positive channel; and performing a full retraining every 90 days, completely retraining the model using all historical data. This defines the specific rhythm of the model update strategy combining incremental learning and periodic retraining. Lightweight incremental updates are used to quickly respond to recent data and warning feedback, while full retraining is used to maintain the model's overall grasp of long-term historical patterns. The combination of the two ensures both timeliness and stability, achieving dynamic adaptive capability.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] This invention constructs a cross-modal knowledge graph by integrating multimodal heterogeneous scientific and technological big data, overcoming the problem of incomplete information from a single data source and improving the comprehensiveness and accuracy of predictions. It pioneers a bidirectional collaborative prediction and early warning neural network, which proactively guides early warnings through positive prior knowledge injection and drives online incremental updates of the prediction model through reverse early warning feedback. This breaks the long-standing separation between prediction and early warning, achieving an upgrade from passive anomaly detection to proactive risk prediction, and endowing the model with continuous adaptive capabilities. Combining Monte Carlo methods and feature attribution techniques, it constructs a multi-level confidence and interpretability evaluation system, enabling decision-makers to adopt prediction results at different confidence levels, enhancing decision-making transparency. Simultaneously, it designs multi-dimensional, dedicated early warning indicators specific to the unique laws of scientific and technological development, filling the gap in dedicated early warning capabilities in the scientific and technological field. This further improves prediction performance compared to existing methods, and the system possesses good scalability and industrial applicability. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0033] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0034] like Figure 1 As shown, the following is a detailed embodiment.
[0035] This invention first performs the collection and preprocessing of multimodal heterogeneous scientific and technological big data. The system deploys multiple collection sub-modules in parallel, connecting to patent databases, academic paper databases, science and technology funding information platforms, government policy release channels, and science and technology news and social media, simultaneously acquiring multi-source heterogeneous data. The collected raw data undergoes preprocessing operations such as deduplication, normalization of institution and personnel names, extraction of textual semantic features, and unified timestamp alignment to form a standardized multimodal dataset.
[0036] Building upon this foundation, a cross-modal science and technology knowledge graph is constructed. An attribute graph model is used to define technology nodes, institutional nodes, personnel nodes, and national nodes, as well as edge types such as citation relationships, collaborative R&D relationships, co-occurrence relationships, and technology inheritance relationships. Nodes and edges are extracted from multimodal data using entity recognition and disambiguation algorithms and stored in a native graph database. The knowledge graph supports incremental updates: new data is appended, existing nodes have their statistical attributes updated, and new nodes or edges are created and their occurrence time recorded. All elements are timestamped to support time-series backtracking.
[0037] Subsequently, a bidirectional collaborative prediction and early warning network is constructed. This network consists of a trend prediction subnetwork and a risk early warning subnetwork, and achieves information interaction and joint optimization through a bidirectional feedback mechanism. In traditional methods, prediction and early warning are independent of each other. This invention uses bidirectional feedback to enable prior knowledge of prediction to drive the adaptive adjustment of early warning parameters, while the early warning signal triggers online incremental learning of the prediction model.
[0038] The trend prediction subnetwork comprises a modality alignment encoder, a temporal graph Transformer encoder, a technology fusion prediction head, and a topic evolution prediction head. The modality alignment encoder maps features from various modalities to a unified semantic space using a shared multilayer perceptron. The temporal graph Transformer encoder discretizes the time axis into time steps, aggregates node neighborhood information using a graph attention network, and then captures the temporal evolution and cross-temporal dependencies of node states using a Transformer. The technology fusion prediction head embeds nodes of candidate technology pairs into the multilayer perceptron and outputs the fusion probability for the next time step. The topic evolution prediction head clusters technology nodes into topics and then extrapolates the future popularity of each topic using a sequence prediction model. The output includes a ranking of technology fusion probabilities, topic popularity evolution curves, and dependency information.
[0039] The risk warning subnetwork comprises a warning indicator sensor layer, a heterogeneous graph attention warning encoder, a time-series anomaly detection module, and a warning decision fusion module. The sensor layer calculates multi-dimensional warning indicators using a sliding window approach, including anomalies in technology integration speed, technology theme concentration risk, disruptive impact index, technology lifecycle deviation, and international cooperation risk. Each indicator is output after normalization. The heterogeneous graph attention warning encoder treats each indicator as a graph node, learns the mutual influence weights between indicators through a heterogeneous graph attention network, and outputs a comprehensive risk feature vector. The time-series anomaly detection module uses a long short-term memory autoencoder to calculate reconstruction errors for anomaly determination and combines cumulative sum control charts to capture small, persistent drifts. The warning decision fusion module fuses the risk feature vector with the anomaly score and outputs a Level I (attention), Level II (warning), or Level III (emergency) warning via a classification network.
[0040] The specific implementation of the two-way feedback mechanism is as follows: In the forward channel, the prediction results output by the trend prediction subnetwork (such as the probability of technology integration and topic popularity) are pushed to the early warning subnetwork in real time. The latter dynamically adjusts the sensitivity threshold of the early warning indicators accordingly—relaxing the threshold for predicted active directions to reduce false alarms, and tightening the threshold for predicted declining directions to increase sensitivity. In the reverse channel, when the early warning subnetwork outputs Level II or higher warnings for multiple consecutive time steps, a local incremental update of the trend prediction subnetwork is triggered: the underlying parameters are frozen, only the last few layers are fine-tuned, recent new data and early warning signals are used as supervision signals, and a knowledge distillation method is used to prevent catastrophic amnesia.
[0041] For each prediction result output by the trend prediction subnetwork, this invention performs temporal interpretability and confidence assessment. During the prediction phase, the same input is forward-propagated multiple times while maintaining Dropout activation to obtain a set of prediction results. The mean and variance are calculated, and the inverse mapping of the normalized variance is used as the confidence score, thereby classifying the predictions into high, medium, and low levels. Simultaneously, the SHAP framework is invoked to calculate the Shapley values of each input feature. After sorting by absolute value, several features with the largest positive and negative contributions are selected to generate an interpretability report, presented as a feature importance list or evolution path diagram.
[0042] For warning signals output by the risk warning sub-network, the system executes multi-level warnings and graded responses. The warning decision fusion module outputs the probability of each level, and the level corresponding to the maximum value is taken as the final warning level. Level I warnings indicate mild anomalies, and are pushed out through regular reports with a suggestion to increase the frequency of tracking. Level II warnings indicate medium risk, and are pushed out through email or system notifications, along with suggestions for diversifying investments and assessing the diversity of technical routes. Level III warnings indicate high risk, and are immediately delivered to the decision-making level through multiple channels such as instant messages and SMS, along with emergency suggestions for initiating alternative solutions and adjusting R&D investment. The warning system supports tracing and viewing the original indicator sequence and the influence weights between indicators.
[0043] Finally, the system employs a continuous model update strategy combining incremental learning and periodic retraining. Incremental updates are performed at shorter intervals (e.g., every 7 days), using newly acquired data and early warning feedback signals to adjust the last few layers of the trend prediction subnetwork, the baseline of the early warning indicator sensor, and the normal mode of the anomaly detection module, while using elastic weight consolidation to prevent forgetting. Full retraining is performed at longer intervals (e.g., every 90 days), using all historical data to re-optimize all network parameters, and replacing the online model after offline validation. Version management is implemented for each update, supporting retrospective comparison, thus enabling the system to continuously adapt.
[0044] In summary, this invention, through the organic combination of the above steps, forms a complete technical closed loop from data collection, knowledge graph construction, bidirectional collaborative prediction and early warning, confidence assessment, hierarchical response to adaptive learning, achieving integrated and continuous optimization of trend prediction and risk early warning in the field of science and technology. The basic principles, main features, and advantages of this invention have been shown and described above.
[0045] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting and warning technological trends based on big data analysis, characterized in that: Includes the following steps: S1: Simultaneously collect multimodal heterogeneous scientific and technological big data from multiple data sources, and perform data cleaning, standardization alignment, and feature extraction; S2: Map the multimodal data processed by S1 into a unified knowledge representation framework, construct a cross-modal science and technology knowledge graph, and incrementally update the newly incoming data during the operation; S3: Construct a two-way collaborative prediction and early warning network, which consists of a trend prediction sub-network and a risk early warning sub-network, and realizes information interaction and joint optimization between the two sub-networks through a two-way feedback mechanism; wherein, the trend prediction sub-network outputs multi-time-scale technology trend prediction results based on the knowledge graph of S2, and the risk early warning sub-network monitors multi-dimensional early warning indicators in real time and generates multi-level early warning signals. S4: Based on time series interpretability and confidence assessment, generate multi-level confidence scores and interpretability reports for each prediction result output by the trend prediction sub-network in S3; S5: The risk warning sub-network outputs the warning level and provides corresponding handling suggestions based on real-time monitoring and comprehensive analysis of multi-dimensional warning indicators; S6: Based on the bidirectional feedback mechanism in S3 and the early warning signal output by S5, a model continuous update strategy combining incremental learning and periodic retraining is adopted to enable the system to adapt to the dynamic changes in technological development. 2.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The multimodal heterogeneous science and technology big data mentioned in S1 includes: patent data, academic paper data, science and technology fund support data, industrial policy data, and science and technology media public opinion data. 3.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The bidirectional feedback mechanism in S3 includes a positive channel and a negative channel: The positive channel injects the technical field popularity prediction and / or technical fusion probability distribution output by the trend prediction subnetwork into the risk warning subnetwork as prior knowledge, which is used to dynamically adjust the threshold of the warning indicators and realize the upgrade from passive anomaly detection to proactive risk prediction. When the risk warning subnetwork issues a Level II or higher warning for T consecutive time steps, the reverse channel triggers a local incremental update of the trend prediction subnetwork. The parameters of at least the last layer of the prediction subnetwork are fine-tuned through the knowledge distillation method, so that the prediction model incorporates the warning signal into the consideration of future predictions. 4.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The multidimensional early warning indicators in S5 include at least one of the following: an abnormal speed indicator for technological integration, a risk indicator for the concentration of technological themes, a disruptive impact index, a deviation index from the technological life cycle, and a risk indicator for international cooperation; wherein: The anomaly index of technology convergence speed is calculated by comparing the co-occurrence growth rate among different technology fields with the degree of deviation of similar technologies from the average growth rate. The technology theme concentration risk index is based on the Herfindahl index to calculate the distribution concentration of patents and papers within a research theme area; The Disruptive Impact Index quantifies the impact of new technologies on existing knowledge systems by analyzing the rate of change in the reference network topology. The technology lifecycle deviation index is calculated by comparing the deviation between actual observed values and the fitted curve of the Logistic growth model; International cooperation risk indicators monitor changes in the number of cross-border patent collaborations and the connectivity of cooperation networks in key areas. 5.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The confidence assessment in S4 uses the Monte Carlo Dropout method: during the prediction phase, the model is propagated K times forward, and some neurons are randomly masked at a predetermined Dropout rate during each propagation. The prediction mean and prediction variance are calculated based on the statistical distribution of the K prediction results, and the confidence score is calculated as 1 − (variance / maximum possible variance). Furthermore, the SHAP method is used to calculate the marginal contribution of each input feature to the prediction result, providing interpretability information in the order of importance. The multi-level confidence score is divided into three levels: high confidence (≥0.85), medium confidence (0.60~0.85), and low confidence (<0.60).
6. The method for predicting and early warning of technological trends based on big data analysis according to claim 1, characterized in that: The trend prediction sub-network in S3 specifically includes: a modality alignment encoder, used to map feature vectors extracted from different modal data to a unified semantic space; a temporal graph Transformer encoder, used to capture long-distance temporal dependencies between technology nodes in the knowledge graph with time steps as sequence units; a technology fusion prediction head, used to output the probability of candidate technology pairs forming technology fusion in the next time step; and a topic evolution prediction head, used to predict the trajectory of attention changes for each technology topic within a future time window. 7.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The risk warning sub-network in S3 specifically includes: a warning indicator sensor layer, used to extract real-time values of multi-dimensional warning indicators from the knowledge graph in a time window sliding manner; a heterogeneous graph attention warning encoder, used to learn the mutual influence weights and dependencies between different warning indicators; a time-series anomaly detection module, used to model the normal pattern of the indicator sequence based on the long short-term memory autoencoder and determine anomalies through reconstruction errors, while combining control chart technology in statistical process control methods to capture trend anomalies early; and a warning decision fusion module, used to weight and fuse the output of the graph attention encoder with the anomaly detection results to comprehensively output the warning level. 8.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The cross-modal science and technology knowledge graph in S2 adopts an attribute graph model G=(V, E, A), where V is a set of nodes, including four types: technology nodes, institutional nodes, personnel nodes, and national nodes; E is a set of edges, including reference relationships, collaborative research and development relationships, co-occurrence relationships, and technology inheritance relationships; and A is a set of attributes for nodes and edges, including timestamps, type labels, and confidence metadata. 9.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The warning levels mentioned in S5 include: Level I warning / attention warning, Level II warning / alert warning and Level III warning / emergency warning, and corresponding handling suggestions are provided according to the warning level; among them, when the Herfindahl index of the technology theme concentration risk indicator is greater than 0.3 and shows an upward trend for three consecutive months, at least a Level II warning is triggered. 10.The big data analysis based technology field trend prediction and early warning method according to claim 1, characterized in that: The model continuous update strategy combining incremental learning and periodic retraining described in S6 is as follows: a lightweight incremental update is performed every 7 days, using newly collected data since the last update and the early warning feedback signals accumulated in the positive channel to adjust the model parameters; a full retraining is performed every 90 days, using all historical data to completely retrain the model.