A scientific and technological event intelligence analysis method based on multi-source data fusion

Abstract

The application provides a scientific and technological event information analysis method based on multi-source data fusion, comprising: acquiring multi-source heterogeneous data, analyzing time difference characteristics and corresponding trust level characteristics of the multi-source heterogeneous data in multiple publishing channels to obtain a first feature set; if the signal comprehensive superposition strength value is greater than a preset fluctuation determination threshold, it is determined that there is an abnormal fluctuation phenomenon in the target field, and a first fluctuation event is recorded; the initial weight coefficient is dynamically corrected and adjusted according to the information arrival time difference characteristics to obtain a target weight coefficient; if the key node activity index in the deep analysis result meets a pre-established response trigger condition, a dynamic trigger instruction for the target field is generated; the dynamic trigger instruction is acquired, the response mechanism for the target field is started according to the dynamic trigger instruction, and final decision layout information is output.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for analyzing intelligence on scientific and technological events based on multi-source data fusion. Background Technology

[0002] Technology intelligence analysis plays a crucial role in national security, industrial decision-making, and strategic planning. Accurately identifying and promptly responding to major breakthroughs, technological shifts, and industrial transformations in the technology field directly impacts the strategic planning and response efficiency of decision-making departments. However, existing technology intelligence monitoring methods generally suffer from delayed responses and inaccurate judgments. Most systems rely on manually set fixed rules or single thresholds for event identification, a practice lacking flexibility in the face of complex and ever-changing technology events. More significantly, current methods often operate independently when processing information from different channels such as social media, industry forums, and official announcements, failing to establish unified judgment criteria. This results in important signals being drowned out by noise or key events being discovered late.

[0003] The deeper technical challenge lies in the differences in timeliness and credibility among different data sources. Social media information spreads rapidly but its authenticity varies greatly; official announcements are highly authoritative but relatively time-consuming; industry forums fall somewhere in between. This mismatch between timeliness and credibility makes it difficult for the system to determine when to trigger an alert. Over-reliance on rapidly spreading social media information can lead to numerous false alarms and wasted resources; while waiting for official confirmation may result in missing the optimal response opportunity. Further complicating matters, different dimensions of indicators are inherently interrelated. A surge in patent applications for a particular technology may be accompanied by fluctuations in the capital market and increased academic discussion; the cumulative effect of these signals is difficult to measure through simple numerical summation.

[0004] Specifically, when an emerging technology field suddenly experiences abnormal fluctuations across multiple dimensions—for example, when a quantum computing technology sees top-tier journal publications, multiple companies announcing related investments, and a large amount of social media discussion within the same week—the system needs to comprehensively assess the correlation strength and overall importance of these signals, rather than viewing each data point in isolation. Therefore, establishing an intelligent monitoring mechanism capable of dynamically integrating multi-source heterogeneous data, accurately quantifying the weights of different signals, and automatically triggering in-depth analysis at appropriate times becomes a key issue in improving the timeliness and accuracy of intelligence analysis on science and technology events. Summary of the Invention

[0005] This invention provides a method for analyzing science and technology event intelligence based on multi-source data fusion, mainly including: Acquire multi-source heterogeneous data, analyze the timeliness differences and corresponding trust level characteristics of the multi-source heterogeneous data under multiple publishing channels, and obtain a first feature set; The first feature set is processed using a machine learning model to identify the information association relationships of multiple information dimensions in the multi-source heterogeneous data and to calculate the initial weight coefficients corresponding to each information dimension. Calculate the comprehensive correlation strength value of the multiple information dimensions within the same monitoring time window based on the initial weighting coefficients, and obtain the comprehensive superposition strength value of the signal for the target domain by accumulating the comprehensive correlation strength value. If the combined intensity value of the signals is greater than the preset fluctuation judgment threshold, then it is determined that there is an abnormal fluctuation phenomenon in the target area, and the abnormal fluctuation phenomenon is recorded to obtain the first fluctuation event. A time-series prediction model is used to process the historical evolution time-series data of the first fluctuation event to determine the information arrival time difference characteristics of the first fluctuation event under the multiple release channels. The initial weight coefficients are dynamically corrected and adjusted based on the information arrival time difference characteristics to obtain the target weight coefficients; By processing the correlation data between the target weight coefficient and the first fluctuation event through a relational analysis model, the relational network structure between the core entity nodes corresponding to the first fluctuation event is extracted to obtain in-depth analysis results. If the activity index of key nodes in the deep analysis results meets the pre-established response triggering conditions, then a dynamic triggering command for the target domain is generated. Obtain the dynamic triggering instruction, initiate a response mechanism for the target domain based on the dynamic triggering instruction, and output the final decision layout information.

[0006] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a method for analyzing science and technology events based on multi-source data fusion, primarily addressing the issues of delayed intelligence analysis and misjudgments caused by the timeliness differences of multi-source heterogeneous data across different release channels. This invention analyzes the timeliness differences and trust level characteristics of multi-source heterogeneous data, utilizes a machine learning model to identify the correlations across multiple information dimensions and calculates initial weight coefficients. Within the same monitoring time window, it accumulates comprehensive correlation strength values ​​to form a signal superposition strength value. When this value exceeds a fluctuation judgment threshold, it promptly captures abnormal fluctuations in the target domain. For captured fluctuation events, this invention employs a time-series prediction model to analyze the information arrival time differences across multiple release channels, dynamically adjusting the weight coefficients accordingly. A relationship analysis model extracts the relationship network structure between core entity nodes, automatically generating dynamic trigger commands and initiating a response mechanism when the activity of key nodes meets the response trigger conditions. This invention achieves intelligent fusion and real-time monitoring of multi-source heterogeneous data, improving the timeliness and accuracy of science and technology event intelligence analysis and providing timely and reliable intelligence support for decision-making in the target domain. Attached Figure Description

[0007] Figure 1 This is a flowchart of a technology event intelligence analysis method based on multi-source data fusion according to the present invention.

[0008] Figure 2 This is a schematic diagram of a technology event intelligence analysis method based on multi-source data fusion according to the present invention.

[0009] Figure 3 This is another schematic diagram of a technology event intelligence analysis method based on multi-source data fusion according to the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0011] like Figures 1-3 This embodiment of a technology event intelligence analysis method based on multi-source data fusion may specifically include: Step S101: Obtain multi-source heterogeneous data, analyze the timeliness differences and corresponding trust level characteristics of the multi-source heterogeneous data under multiple publishing channels, and obtain the first feature set.

[0012] Multi-source heterogeneous data is acquired, and the publication timestamps of the multi-source heterogeneous data on multiple publishing ends are extracted to obtain a first timeliness sequence. Based on the first timeliness sequence, the absolute value of the time difference between different publishing ends is calculated to obtain the timeliness difference. For the timeliness difference, if the timeliness difference is greater than a preset threshold, it is determined that the corresponding publishing end has a delay, and a set of delayed publishing ends is determined. The set of delayed publishing ends is classified using a random forest algorithm to obtain the trust level of each publishing end. The trust level is obtained, and if the trust level is lower than a preset trust threshold, the trust level is downgraded to obtain a level score. The level score and the timeliness difference are concatenated to obtain a first feature set.

[0013] For example, in a financial market information monitoring scenario, acquiring multi-source heterogeneous data includes obtaining text news from official media and market data snapshots from data terminals. The timestamps of these data across multiple publishing platforms are extracted to form a first timeliness sequence. For instance, the timestamp of an official media article about a merger and acquisition might be extracted as 10:05:00, while the timestamp of a self-media article might be 10:05:30. This unifies the time dimension of different data sources, providing foundational data support for subsequent timeliness analysis.

[0014] In one possible implementation, the time difference is obtained by calculating the absolute value of the time difference between different publishing platforms based on the first timeliness sequence. A preset threshold of 15 seconds is set. If the time difference between self-media and official media is 30 seconds, which is greater than the 15-second threshold, then the self-media platform is determined to have a delay and is included in the set of delayed publishing platforms. By setting a reasonable threshold, it is possible to accurately filter out publishing channels that truly have information lag, thereby improving the accuracy of monitoring.

[0015] It should be noted that for the set of delayed publishers, a random forest algorithm is used for classification to obtain trust scores. Input features include the number of historical delays, average delay duration, and data source type. The random forest algorithm performs ensemble learning by constructing multiple decision trees, outputting a trust score for each publisher, quantified from 0 to 100. For example, a certain self-media outlet received a score of 45 due to frequent delays. This process effectively utilizes multi-dimensional historical features, avoiding the bias of a single indicator evaluation, and significantly improving the objectivity and robustness of the trust assessment.

[0016] For example, after obtaining the aforementioned trust level, a preset trust threshold of 60 points is set. If the trust level of a publishing end is 45 points, which is lower than this threshold, a downgrade process is triggered. The downgrade rule can be based on mapping trust level ranges to different level scores, such as mapping 40 to 50 points to a level score of 2 points, and 60 points and above to a level score of 5 points. This tiered downgrade process can intuitively reflect the reliability of the publishing end, facilitating the rapid invocation of subsequent strategies.

[0017] In one possible implementation, a first feature set is obtained by concatenating the credit rating score and the timeliness difference. The credit rating score of 2 points is combined with the timeliness difference of 30 seconds to form the first feature set. This set integrates the static credit rating and dynamic timeliness performance at the publishing end, providing comprehensive and high-quality input features for downstream information recommendation ranking or risk warning models, effectively enhancing the data processing efficiency of the overall business chain.

[0018] Step S102: The first feature set is processed using a machine learning model to identify the information correlation relationships of multiple information dimensions in the multi-source heterogeneous data, and the initial weight coefficients corresponding to each information dimension are calculated.

[0019] Obtain heterogeneous source data, extract the dimensional values ​​of the heterogeneous source data, and obtain a first feature set based on the dimensional values. Process the first feature set using the cosine distance algorithm, calculate the distance value of each feature within the first feature set, and obtain an initial similarity. Determine the association state between features based on the initial similarity; if the initial similarity is greater than a preset similarity threshold, a first association degree is determined between the features. Receive the first association degree through a decision tree model, analyze the number of nodes corresponding to the first association degree, and determine the initial weight value corresponding to each node number. Obtain the initial weight values, extract the deviation contained in the initial weight values, and obtain a first compensation value based on the deviation. Process the initial weight values ​​using the first compensation value to obtain the target weight value.

[0020] Specifically, in financial credit risk control scenarios, acquiring heterogeneous source data refers to collecting background information on borrowers in different formats. This heterogeneous source data includes structured bank transaction records and unstructured transcripts of face-to-face interviews. The process involves extracting dimensional values ​​from this heterogeneous data: extracting the average monthly income from transaction records and the repayment intention / sentiment dimension from the transcripts. These dimensional values ​​provide the first feature set. The advantage of this approach is that it breaks down data silos, unifying multi-source, differentiated data into a computable feature vector, thus comprehensively depicting the borrower's credit profile.

[0021] For example, a cosine distance algorithm is used to process the first feature set. Its inputs are the extracted monthly average income feature vector and repayment willingness feature vector, and its output is the cosine of the angle between them in multidimensional space. Initial similarity is obtained by calculating the distance between each feature in the first feature set. Assume the initial similarity between two features is 85, with a maximum score of 100. The association between features is determined based on the initial similarity. If the initial similarity is greater than the preset similarity threshold of 80, a first degree of association is determined between the features. This process effectively quantifies the collinearity between different risk control features, eliminates redundant features, and improves the computational efficiency of subsequent models.

[0022] It should be noted that the decision tree model receives the first degree of association, analyzes the number of nodes corresponding to the first degree of association, and determines the initial weight value corresponding to each node. The input is the first degree of association, the algorithm is a classification and regression tree algorithm, and the output is the initial weight value. When a feature with a high degree of association is assigned to a node at the 3rd level of the decision tree, due to the deeper level, it is assigned an initial weight value of 10. After obtaining the initial weight value, the bias contained in the initial weight value is extracted. Since uneven sample distribution may cause the model to over-rely on certain features and produce bias, assuming that the extracted bias is 2, a first compensation value of -2 is obtained based on this bias. The initial weight value is processed using the first compensation value, that is, the initial weight value and the compensation value are added together to obtain the target weight value of 8. This mechanism corrects the weight allocation of decision tree nodes through dynamic compensation, avoids overfitting of the risk control model, and significantly improves the accuracy of credit default prediction.

[0023] Step S103: Calculate the comprehensive correlation strength value of the multiple information dimensions within the same monitoring time window based on the initial weight coefficients, and obtain the comprehensive superposition strength value of the signal for the target domain by accumulating the comprehensive correlation strength values.

[0024] First signal source data within the target domain is acquired. Second signal source data is obtained by truncating the first signal source data according to a preset time window. Multiple information dimensions are obtained by parsing the second signal source data. Principal component analysis (PCA) is used to reduce the dimensionality of these first information dimensions to obtain second information dimensions. A first product term is calculated by multiplying the second information dimensions with pre-established initial weights. A first correlation degree is obtained by summing the first product term. If the first correlation degree is greater than a preset correlation degree threshold, it is determined to be the target correlation degree, and a first intensity value corresponding to the target correlation degree is acquired. A first superposition value is obtained by accumulating multiple first intensity values ​​within the same monitoring period. A comprehensive signal superposition intensity value for the target domain is determined based on the first superposition value. If the comprehensive signal superposition intensity value is greater than a preset intensity threshold, an abnormal state of the target domain corresponding to the comprehensive signal superposition intensity value is determined, and an alarm command corresponding to the abnormal state is output.

[0025] For example, regarding the acquisition and interception process of the first signal source data within the target domain, taking industrial motor operation status monitoring as an example, continuous vibration signals collected by motor sensors are acquired as the first signal source data. A time window length of 5 seconds is set, and the continuous vibration signal is intercepted by sliding within this time window to obtain multiple independent 5-second intervals of second signal source data. This interception method transforms an infinitely long continuous signal into a finite-length analysis sample, facilitating subsequent feature extraction and dimensionality reduction processing.

[0026] Specifically, after acquiring the data from the second signal source, it is analyzed to extract the first values ​​of multiple information dimensions. The extracted information dimensions include vibration frequency, amplitude, phase, etc., totaling 20 feature values. Due to the high dimensionality and information redundancy, principal component analysis (PCA) is used to reduce the dimensionality of these 20 first values. The extracted feature value matrix is ​​input into the PCA algorithm model, and orthogonal transformations are used to convert highly correlated variables into linearly independent principal components, outputting the second values ​​of the three core information dimensions after dimensionality reduction. Initial weight values ​​are pre-set for these three core information dimensions, for example, 8, 5, and 2. The first product term is calculated by multiplying each of these three second values ​​with its corresponding initial weight value. Then, all first product terms are summed, and the output result is the first correlation degree.

[0027] In one possible implementation, a correlation threshold of 50 is set. When the calculated first correlation is greater than 50, it is determined to be the target correlation, and the corresponding first intensity value is extracted, for example, a first intensity value of 15. Within a set 10-minute monitoring period, there may be multiple target correlations that meet the conditions. All the first intensity values ​​extracted within the monitoring period are summed to obtain a first superposition value. Based on this first superposition value, the comprehensive superposition strength value of the signal for the motor target domain can be determined. With an intensity threshold of 100, if the final calculated comprehensive superposition strength value reaches 120, which is greater than the preset intensity threshold, an abnormal state corresponding to this value in the motor target domain is determined, such as bearing wear or rotor imbalance. An alarm command corresponding to this abnormal state is then output, driving the on-site alarm equipment to activate.

[0028] Step S104: If the combined intensity value of the signals is greater than the preset fluctuation judgment threshold, it is determined that there is an abnormal fluctuation phenomenon in the target area, and the abnormal fluctuation phenomenon is recorded to obtain the first fluctuation event.

[0029] Initial signal values ​​are acquired from the acquisition end of the target domain, and a first sequence group is obtained by truncating the initial signal values ​​through a sliding window. Based on the first sequence group, the outlier score of the first sequence group is calculated using the isolated forest algorithm to obtain the signal strength value. A preset benchmark strength value is obtained, and the absolute difference between the signal strength value and the benchmark strength value is calculated to obtain a first difference term. Based on the first difference term, a smoothing filter is applied to the first difference term to obtain a smoothed fluctuation amount. If the smoothed fluctuation amount is greater than a preset judgment threshold, the target domain is determined to be in a first anomalous state. For the first anomalous state, the duration of the first anomalous state is extracted to obtain an anomalous state quantity. Based on the anomalous state quantity, the anomalous state quantity is classified according to a preset classification rule to obtain a first event set.

[0030] For example, the continuous initial signal values ​​transmitted in real time by the target domain acquisition terminal are obtained, and a sliding window mechanism is used to segment the continuous initial signal values. The length of the sliding window is set to 60 sampling points, and the sliding step size is 10 sampling points. By sliding this window on the time axis, the continuous signal stream is divided into multiple first sequence groups containing 60 discrete values.

[0031] In one possible implementation, an isolation forest algorithm is introduced to mine outlier features for the first sequence group obtained from the extraction. Multiple isolation trees are constructed using the first sequence group as input data, and the data is recursively partitioned by randomly selecting features and split points. The average path length required for the first sequence group to be isolated in the isolation trees is calculated and converted into an outlier score. This outlier score is directly mapped to the signal strength value within the current time window; a higher score indicates a greater deviation of the signal from the normal distribution.

[0032] It should be noted that a pre-calibrated reference strength value is obtained, set to 20. The absolute difference between the current signal strength value and this reference strength value is calculated to obtain the first difference term. Subsequently, a moving average filtering algorithm is used to smooth the first difference term for multiple consecutive time windows, with the filter window size set to 5, and the smoothed fluctuation is output to reflect the true trend of signal strength change.

[0033] For example, the smoothed fluctuation amount is compared with a preset threshold. If the threshold is set to 15, when the smoothed fluctuation amount reaches 18, the target domain is confirmed to have entered a first abnormal state. The duration of this first abnormal state from triggering to returning to normal is further extracted, for example, lasting 45 seconds, and this duration is used as the abnormal state quantity. Finally, according to preset classification rules, if the abnormal state quantity is between 30 and 60 seconds, it is classified as a moderately persistent abnormal event, thereby generating a first event set containing specific abnormality levels and durations.

[0034] Step S105: Use a time series prediction model to process the historical evolution time series data of the first fluctuation event to determine the information arrival time difference characteristics of the first fluctuation event under the multiple release channels.

[0035] A historical set of the first fluctuation event is acquired and processed using data cleaning tools to obtain standard timeline data. Based on the standard timeline data, the information flow of the first fluctuation event at multiple channels is extracted to obtain an initial sequence set. A time-series prediction model is used to process the initial sequence set to obtain the evolution points of the first fluctuation event at the multiple channels. The evolution points at the multiple channels are analyzed using comparison tools to determine the information arrival points at each channel. Based on the information arrival points, the delay between different channels is calculated to obtain an initial difference value. If the initial difference value is greater than a preset threshold, it is smoothed to obtain a target difference value. Based on the target difference value, the information arrival time difference characteristics of the first fluctuation event at the multiple channels are determined.

[0036] In one possible implementation, data cleaning is performed on the historical set of the first fluctuation event, with the input being the original fluctuation records containing noise and missing values. The K-nearest neighbor interpolation algorithm is used to fill in the missing signal strength values, and the non-standard timestamps from different acquisition nodes are uniformly converted to Coordinated Universal Time (UTC) format, thus outputting standard timeline data. By constructing a unified time base, fundamental data support can be provided for subsequent multi-dimensional signal alignment.

[0037] For example, based on the above standard timeline data, the information flow of the first fluctuation event is extracted across multiple channels.

[0038] Specifically, signal intensity change sequences within corresponding time periods are extracted from sensor nodes A and B in the target domain. These multi-source heterogeneous sequence data are then spliced ​​and recombined to obtain an initial sequence set. This initial sequence set completely preserves the propagation pattern of the same wave event at different spatial locations.

[0039] It should be noted that after obtaining the initial sequence set, a Long Short-Term Memory (LSTM) network model is used to process it. The initial sequence set is used as the input to this time series prediction model, and the long-term dependencies in the signal sequence are captured through the model's internal gating mechanism to output the evolution points of the first fluctuation event at each channel.

[0040] For example, the model identifies that the signal abrupt change at node A occurs at a certain moment, while the change at node B occurs at another subsequent moment.

[0041] In one possible implementation, a dynamic time warping algorithm is used as a comparison tool to analyze the evolution points of node A and node B, determining the information arrival points of the multiple channels. Based on the extracted information arrival point timestamps, the time difference between node A and node B is calculated to obtain an initial difference value. Assume the calculated initial difference value is 45 milliseconds.

[0042] For example, a preset threshold of 30 milliseconds is set. Since 45 milliseconds is greater than this threshold, an exponential smoothing algorithm is used to filter the initial difference value to eliminate random errors caused by network transmission jitter, resulting in a target difference value of 42.5 milliseconds. Based on this target difference value, the information arrival time difference characteristics of the first fluctuation event under the multi-channel network are finally determined. This characteristic can directly reflect the asymmetric law of signal propagation within the target domain.

[0043] Step S106: Dynamically adjust the initial weight coefficients according to the information arrival time difference characteristics to obtain the target weight coefficients.

[0044] Upon arrival of the first information stream, the time difference between adjacent information streams is calculated to obtain an initial time difference. An anomaly detection is performed on the initial time difference using the Isolation Forest algorithm. If the initial time difference is determined to be abnormal, it is discarded to obtain a standard time difference. The time dimension difference is extracted based on the standard time difference to obtain a first difference value. A preset initial weight is obtained, and a correction amount is calculated for the initial weight based on the first difference value to obtain a first correction amount. The initial weight is dynamically compensated using the first correction amount to obtain a compensation value. A weight adjustment rate is calculated based on the compensation value to obtain a first adjustment rate. The initial weight is updated using the first adjustment rate to obtain the target weight.

[0045] Specifically, for a specific fluctuation event in the information flow of a certain information channel, the absolute timestamp of each piece of information arriving at the receiving end is recorded as the arrival time. The initial time difference is calculated by taking the arrival time of two adjacent pieces of information; for example, if the first piece arrives at 10:05:10 and the second at 10:05:15, the initial time difference is 5 seconds. Since network jitter can cause extreme data delays, an isolated forest algorithm is used to construct multiple isolated trees to spatially partition the initial time difference set. If a time difference, such as 120 seconds, is isolated after only a few path partitions in the isolated trees, and its anomaly score is greater than the 60% threshold, it is judged as an anomaly and removed. Normal fluctuation data within the range of 2 to 10 seconds are retained as the standard time difference.

[0046] For example, based on the aforementioned standard time difference, the time dimension difference value between different channels within the same time window is extracted. Assuming the average standard time difference for the first channel is 3 seconds and for the second channel it is 8 seconds, the extracted first difference value is 5 seconds. In multi-channel information fusion services, each channel is pre-assigned an initial weight; for example, the initial weight for the first channel is 80. Based on the first difference value of 5 seconds, a correction step of 2 is set for each 1-second difference, calculating a correction amount of 10 for the initial weight of the first channel. This correction amount is then used to dynamically compensate the initial weight, i.e., the initial weight and the correction amount are linearly superimposed to obtain a compensation value of 90.

[0047] In one possible implementation, to prevent sudden weight changes from causing instability in subsequent information processing, a weight adjustment rate needs to be calculated based on the compensation value. A smoothing coefficient is set to half, and the difference between the compensation value of 90 and the original initial weight of 80 is multiplied by this smoothing coefficient to obtain a first adjustment rate of 5. Finally, this first adjustment rate is used to update the initial weight of 80, i.e., by adding the first adjustment rate of 5, to obtain a target weight of 85. Through the above quantification process, dynamic weight adaptation of the channel information flow arrival characteristics is completed.

[0048] Step S107: Process the correlation data between the target weight coefficient and the first fluctuation event through a relational analysis model, extract the relational network structure between the core entity nodes corresponding to the first fluctuation event, and obtain in-depth analysis results.

[0049] The association data between the target weight coefficients and the first fluctuation event is obtained. This association data is then processed using a data cleaning tool to obtain a standard association dataset. Based on this standard association dataset, the core entity nodes corresponding to the first fluctuation event are extracted to obtain an initial entity node set. This initial entity node set is then processed using a relationship analysis model to calculate the connection strength between each node, resulting in a node connection strength matrix. If the connection strength in the node connection strength matrix is ​​greater than a preset strength threshold, the corresponding node connection relationship is retained, obtaining a core node pair set. Based on this core node pair set, a relationship network structure between the core entity nodes is constructed to obtain an initial relationship network graph. This initial relationship network graph is then processed using a graph neural network model to extract its topological attributes, obtaining deep analysis results.

[0050] For example, in the process of acquiring the correlation data between the target weight coefficient and the first fluctuation event, the input sources are the previously determined target weight coefficient and the records of traffic surge fluctuation events monitored in the information flow transmission network. The correlation data is processed using a data cleaning tool, with a mean-filling algorithm used to handle missing fields, invalid records with incorrect formats removed, and a standard correlation dataset output. Based on this standard correlation dataset, the sending and receiving nodes that directly interact with the traffic surge fluctuation events are extracted as core entity nodes, thereby constructing an initial set of entity nodes.

[0051] Specifically, after obtaining the initial set of entity nodes, the mutual information algorithm is used as the relationship analysis model to process this set. The input is the historical information flow interaction frequency sequence of each node in the initial entity node set. The mutual information algorithm calculates the dependency relationship between any two node sequences and outputs a node connection strength matrix. The values ​​in this matrix represent the degree of interaction between nodes. A preset strength threshold of 85 is set. If the connection strength between two nodes in the node connection strength matrix reaches 90, which is greater than the preset strength threshold, the connection relationship between these two nodes is retained and included in the core node pair set.

[0052] In one possible implementation, based on the core node pair set obtained through the above filtering, nodes are used as vertices and the connections between nodes are used as edges to construct a relational network structure between the core entity nodes, generating an initial relational network graph. This initial relational network graph is then input into a graph convolutional network model for processing. The graph convolutional network model extracts the topological properties of the initial relational network graph by aggregating the feature information of neighboring nodes, such as calculating the degree centrality feature vector of each node. The final output depth analysis result includes the distribution feature vectors of key bottleneck nodes in the network.

[0053] Step S108: If the activity index of key nodes in the deep analysis results meets the pre-established response triggering conditions, then a dynamic triggering command for the target domain is generated.

[0054] An initial analysis set of the target domain is obtained, and the initial analysis set is classified using a random forest algorithm to obtain a first node group. Based on the first node group, the interaction frequency of each node in the first node group is extracted, and a first activity level is calculated based on the interaction frequency. A pre-established response trigger line is obtained; if the first activity level is greater than the response trigger line, the first node group is determined to be a key node group. Based on the key node group, a dynamic set corresponding to the key node group is extracted, and a first instruction code is generated using the dynamic set. The first instruction code is obtained, and the first instruction code is classified and identified using a support vector machine algorithm to obtain a second instruction code. Based on the second instruction code, the target domain identifier corresponding to the second instruction code is extracted, and a dynamic trigger instruction for the target domain is generated based on the target domain identifier.

[0055] In one possible implementation, an initial analysis set for the target domain is obtained. The target domain is a specific transaction monitoring area, and the initial analysis set contains feature data of each transaction account within that area. A random forest algorithm is used to classify the initial analysis set. The input consists of the account's transaction amount and transaction time interval features. By constructing 100 decision trees for voting, the classification results are output, and accounts with high-frequency transaction characteristics are grouped into the first node group.

[0056] For example, based on the first node group, the interaction frequency of each node is extracted. Interaction frequency refers to the number of times funds are transferred between nodes and other accounts within 24 hours. A first activity level is calculated based on the interaction frequency, with a weight of 0.8 for interaction frequency and a weight of 0.2 for the amount of each interaction. A pre-established response trigger line is obtained, which is set to an activity level value of 85. If the first activity level is greater than 85, the first node group is determined to be a key node group.

[0057] It should be noted that, based on the key node group, its corresponding dynamic set is extracted. The dynamic set contains the fund flow sequence of the key node group over the past 12 hours. A first instruction code is generated from the dynamic set, mapping the fund flow sequence to a specific encoding format, and encoding the behavior of flowing to a specific account as the first instruction code.

[0058] In one possible implementation, a first instruction code is obtained, and a support vector machine (SVM) algorithm is used to classify and identify it. The input is the feature vector of the first instruction code, which is mapped to a high-dimensional space using a kernel function to find the optimal classification hyperplane. The classification result is output, yielding a second instruction code. The second instruction code represents a specific risk level. Based on the second instruction code, the corresponding target domain identifier is extracted, such as the number of a specific trading area. Dynamic triggering instructions are generated for the target domain based on the target domain identifier, such as instructions to restrict the trading permissions of relevant accounts within that area.

[0059] Step S109: Obtain the dynamic triggering instruction, activate the response mechanism for the target domain according to the dynamic triggering instruction, and output the final decision layout information.

[0060] The trigger command is obtained and processed through text parsing to obtain the current state value. Based on the current state value, the corresponding rule base is retrieved from the pre-established target domain to determine the initial matching degree between the current state value and the rule base. If the initial matching degree is greater than a preset matching threshold, the response end corresponding to the target domain is activated to obtain the initial action set output by the response end. The initial action set is classified using a random forest algorithm to determine the classified target node tree. The initial weight values ​​of each branch node are calculated based on the target node tree to obtain an initial decision graph composed of the initial weight values. The initial decision graph is spatially partitioned using a support vector machine algorithm to obtain the final layout table.

[0061] In one possible implementation, after obtaining the trigger command, the trigger command is parsed using word segmentation technology in natural language processing algorithms to extract the current state value that represents the characteristics of the current environment.

[0062] For example, if the trigger command includes an abnormal temperature increase, the parsed current state value can be quantified as state identifier 85. Based on state identifier 85, feature vectors are compared in a pre-established target domain rule base, and the cosine similarity between the state identifier and a preset standard state in the rule base is calculated to determine the initial matching degree. If the calculated initial matching degree is 90, which is greater than the preset matching threshold 80, the trigger condition is determined to be met, and the response terminal corresponding to the target domain is activated to collect the initial action set containing multiple actions to be executed generated by the response terminal in the current state.

[0063] For example, the initial action set is classified using a random forest algorithm.

[0064] Specifically, each action feature in the initial action set is used as input data, and multiple decision trees are constructed using the random forest algorithm for ensemble learning. During classification, the algorithm splits nodes based on features such as action execution priority and resource consumption, ultimately outputting a classified target node tree. This target node tree divides the initial action set into branch nodes at different levels, such as emergency action nodes and regular action nodes, providing structured data support for subsequent decision-making.

[0065] It should be noted that, based on the target node tree, the initial weight values ​​of each branch node are further calculated. By extracting the historical execution success frequency and resource consumption ratio of each branch node, the initial weight value is calculated according to a preset weighting rule; for example, the initial weight value of an emergency action node is set to 75. These initial weight values ​​constitute the initial decision graph. Subsequently, the initial decision graph is used as input, and a support vector machine (SVM) algorithm is employed for spatial partitioning. The SVM algorithm finds the optimal hyperplane to isolate high-weight nodes from low-weight nodes in the initial decision graph in a multi-dimensional space, ultimately outputting a final layout table. This layout table clearly defines the execution order and spatial distribution of each action in the target domain.

[0066] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for analyzing science and technology event intelligence based on multi-source data fusion, characterized in that, The method includes: Acquire multi-source heterogeneous data, analyze the timeliness differences and corresponding trust level characteristics of the multi-source heterogeneous data under multiple publishing channels, and obtain a first feature set; The first feature set is processed using a machine learning model to identify the information association relationships of multiple information dimensions in the multi-source heterogeneous data and to calculate the initial weight coefficients corresponding to each information dimension. Calculate the comprehensive correlation strength value of the multiple information dimensions within the same monitoring time window based on the initial weighting coefficients, and obtain the comprehensive superposition strength value of the signal for the target domain by accumulating the comprehensive correlation strength value. If the combined intensity value of the signals is greater than the preset fluctuation judgment threshold, then it is determined that there is an abnormal fluctuation phenomenon in the target area, and the abnormal fluctuation phenomenon is recorded to obtain the first fluctuation event. A time-series prediction model is used to process the historical evolution time-series data of the first fluctuation event to determine the information arrival time difference characteristics of the first fluctuation event under the multiple release channels. The initial weight coefficients are dynamically corrected and adjusted based on the information arrival time difference characteristics to obtain the target weight coefficients; By processing the correlation data between the target weight coefficient and the first fluctuation event through a relational analysis model, the relational network structure between the core entity nodes corresponding to the first fluctuation event is extracted to obtain in-depth analysis results. If the activity index of key nodes in the deep analysis results meets the pre-established response triggering conditions, then a dynamic triggering command for the target domain is generated. Obtain the dynamic triggering instruction, initiate a response mechanism for the target domain based on the dynamic triggering instruction, and output the final decision layout information.

2. The method for analyzing science and technology events based on multi-source data fusion according to claim 1, characterized in that, The process involves acquiring multi-source heterogeneous data, analyzing the timeliness differences and corresponding trust level characteristics of the multi-source heterogeneous data across multiple publishing channels, and obtaining a first feature set, including: Acquire multi-source heterogeneous data, extract the publication timestamps of the multi-source heterogeneous data on multiple publishing ends, and obtain the first time-sensitive sequence; Based on the first timeliness sequence, calculate the absolute value of the time difference between different publishing ends to obtain the timeliness difference; Regarding the timeliness difference, if the timeliness difference is greater than a preset threshold, it is determined that the corresponding publishing end has a delay, and a set of delayed publishing ends is determined. The set of delayed publishers is classified using the random forest algorithm to obtain the trust level of each publisher. The trust level is obtained. If the trust level is lower than a preset trust threshold, the trust level is downgraded to obtain a rating score. The first feature set is obtained by concatenating the grade score and the time difference.

3. The method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, The process of using a machine learning model to process the first feature set, identifying the information relationships across multiple information dimensions in the multi-source heterogeneous data, and calculating the initial weight coefficients corresponding to each information dimension includes: Acquire heterogeneous source data, extract the dimension values ​​of the heterogeneous source data, and obtain the first feature set based on the dimension values; The cosine distance algorithm is used to process the first feature set, calculate the distance value of each feature in the first feature set, and obtain the initial similarity. The association status between features is determined based on the initial similarity. If the initial similarity is greater than the preset similarity threshold, it is determined that there is a first degree of association between the features. The decision tree model is used to receive the first degree of association, analyze the number of nodes corresponding to the first degree of association, and determine the initial weight value corresponding to each number of nodes. Obtain the initial weight value, extract the deviation contained in the initial weight value, and obtain the first compensation value based on the deviation; The initial weight value is processed using the first compensation value to obtain the target weight value.

4. The method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, The step of calculating the comprehensive correlation strength value of the multiple information dimensions within the same monitoring time window based on the initial weighting coefficients, and obtaining the comprehensive superposition strength value of the signal for the target domain by accumulating the comprehensive correlation strength values, includes: Acquire the first signal source data within the target domain, and extract the second signal source data by truncating the first signal source data according to a preset time window; By parsing the data from the second signal source, first values ​​of multiple information dimensions are obtained. Principal component analysis algorithm is used to perform dimensionality reduction processing on the first values ​​of the multiple information dimensions to obtain second values ​​of the multiple information dimensions. The first product term is obtained by multiplying the pre-established initial weight value with the second value of the multiple information dimensions, and the first correlation degree is obtained by summing the first product term. If the first correlation degree is greater than the preset correlation degree threshold, then the first correlation degree is determined to be the target correlation degree, and the first strength value corresponding to the target correlation degree is obtained; A first superposition value is obtained by summing and calculating multiple first intensity values ​​within the same monitoring period, and a comprehensive superposition intensity value of the signal for the target domain is determined based on the first superposition value. If the combined intensity value of the signals is greater than a preset intensity threshold, then the abnormal state of the target domain corresponding to the combined intensity value of the signals is determined, and an alarm command corresponding to the abnormal state is output.

5. The method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, If the combined intensity value of the signals is greater than a preset fluctuation judgment threshold, then an abnormal fluctuation phenomenon is determined to exist in the target area, and the abnormal fluctuation phenomenon is recorded to obtain a first fluctuation event, including: The initial signal value is obtained from the acquisition end of the target domain, and the first sequence group is obtained by truncating the initial signal value through a sliding window; Based on the first sequence group, the outlier score of the first sequence group is calculated using the isolated forest algorithm to obtain the signal strength value; Obtain a preset reference strength value, and calculate the absolute difference between the signal strength value and the reference strength value to obtain a first difference term; Based on the first difference term, a smoothing filter is applied to the first difference term to obtain the smoothed fluctuation amount; If the smoothed fluctuation amount is greater than the preset judgment line, then the target domain is determined to be in the first abnormal state; For the first abnormal state, the duration of the first abnormal state is extracted to obtain the abnormal state quantity; Based on the abnormal state quantities, the abnormal state quantities are classified according to preset classification rules to obtain the first event set.

6. The method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, The step of processing the historical evolution time series data of the first fluctuation event using a time series prediction model to determine the information arrival time difference characteristics of the first fluctuation event under the multiple release channels includes: Obtain the historical set of the first fluctuation event, process the historical set using data cleaning tools, and obtain standard timeline data; Based on the standard timeline data, the information flow of the first fluctuation event at multiple channels is extracted to obtain an initial sequence set; The initial sequence set is processed using a time-series prediction model to obtain the evolution points of the first fluctuation event at the multiple channel ends; By analyzing the evolution points of the multiple channel ends using comparison tools, the information arrival points of the multiple channel ends are determined; Based on the arrival point of the information, calculate the delay between different channel ends to obtain the initial difference value; If the initial difference value is greater than a preset threshold, the initial difference value is smoothed to obtain the target difference value; Based on the target difference value, the information arrival time difference characteristics of the first fluctuation event under the multiple channels are determined.

7. The method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, The step of dynamically correcting and adjusting the initial weight coefficients based on the information arrival time difference characteristics to obtain the target weight coefficients includes: When the arrival time of the first information stream is obtained, the time difference between adjacent information streams is calculated based on the arrival time to obtain the initial time difference; The initial time difference is detected as abnormal using the isolated forest algorithm. If the initial time difference is determined to be abnormal, it is removed to obtain the standard time difference. Based on the standard time difference, extract the time dimension difference value to obtain the first difference value; Obtain a preset initial weight, calculate a correction amount for the initial weight based on the first difference value, and obtain a first correction amount; The initial weight is dynamically compensated using the first correction amount to obtain a compensation value; The weight adjustment rate is calculated based on the compensation value to obtain the first adjustment rate; The initial weights are updated using the first adjustment rate to obtain the target weights.

8. The method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, The process involves processing the correlation data between the target weight coefficient and the first fluctuation event using a relational analysis model, extracting the relational network structure between the core entity nodes corresponding to the first fluctuation event, and obtaining in-depth analysis results, including: Obtain the correlation data between the target weight coefficient and the first fluctuation event, and process the correlation data using a data cleaning tool to obtain a standard correlation dataset; Based on the standard associated dataset, extract the core entity nodes corresponding to the first fluctuation event to obtain an initial set of entity nodes; The initial entity node set is processed by a relational analysis model to calculate the connection strength between each node in the initial entity node set, thereby obtaining a node connection strength matrix. If the connection strength in the node connection strength matrix is ​​greater than a preset strength threshold, the corresponding node connection relationship is retained to obtain the core node pair set. Based on the set of core node pairs, construct the relational network structure between the core entity nodes to obtain the initial relational network graph; The initial relational network graph is processed by a graph neural network model to extract its topological attributes and obtain deep analysis results.

9. A method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, If the activity index of key nodes in the deep analysis results meets the pre-established response triggering conditions, then a dynamic triggering instruction for the target domain is generated, including: Obtain the initial analysis set of the target domain, and use the random forest algorithm to classify the initial analysis set to obtain the first node group; Based on the first node group, the interaction frequency of each node in the first node group is extracted, and the first activity level is calculated based on the interaction frequency. Obtain a pre-established response trigger line; if the first activity level is greater than the response trigger line, then determine the first node group as a critical node group. Based on the key node group, extract the dynamic set corresponding to the key node group, and generate the first instruction code through the dynamic set. The first instruction code is obtained, and the first instruction code is classified and identified using a support vector machine algorithm to obtain the second instruction code; Based on the second instruction code, extract the target domain identifier corresponding to the second instruction code, and generate a dynamic trigger instruction for the target domain based on the target domain identifier.

10. A method for analyzing science and technology event intelligence based on multi-source data fusion according to claim 1, characterized in that, The process of obtaining the dynamic triggering instruction, initiating a response mechanism for the target domain based on the dynamic triggering instruction, and outputting final decision layout information includes: The trigger command is obtained, and the trigger command is processed by text parsing to obtain the current status value; Based on the current state value, the corresponding rule base is retrieved in the pre-established target domain to determine the initial matching degree between the current state value and the rule base; If the initial matching degree is greater than the preset matching threshold, then the response terminal corresponding to the target domain is activated to obtain the initial action set output by the response terminal; The initial action set is classified using the random forest algorithm to determine the target node tree after classification; The initial weight values ​​of each branch node are calculated based on the target node tree to obtain an initial decision graph composed of the initial weight values; The initial decision graph is spatially partitioned using the support vector machine algorithm to obtain the final layout table.

Images