Intelligent decision support method for frontier defense public opinion and multi-source data fusion

Abstract

The invention belongs to the field of intelligent security and protection, and relates to an intelligent decision support method for frontier public opinion and multi-source data fusion, which comprises the following steps: S1, collecting preset type public opinion data and preset type basic multi-source data related to frontier protection, and removing public opinion data and multi-source data irrelevant to frontier protection by adopting a data filtering algorithm to obtain a data filtering algorithm; primary data screening is completed; s2, performing semantic and spatial dimension alignment on the primarily screened public opinion data and multi-source data, and cleaning the aligned data to process data conflicts to obtain an associated fusion data set; s3, extracting a corresponding data resolution, collecting frequency class decision demand indexes, and performing priority ranking on the demand indexes; s4, integrating the scheduled data with the public opinion data to generate a decision support report; and S5, collecting application effect feedback of the decision support report, and optimizing data acquisition key points and fusion parameters in the associated fusion data set based on the feedback.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent security and defense, and particularly relates to a method for supporting intelligent decision-making of border defense public opinion and multi-source data fusion. BACKGROUND

[0002] In the existing border defense security technology, multi-source data such as remote sensing and unmanned aerial vehicles are relied on for monitoring, and public opinion is screened manually to assist in judgment, such as earthquakes and military-related information. However, the two are often processed in isolation, lacking semantic and spatial dimension correlation fusion, which makes it difficult to combine soft information and hard data, i.e., public opinion and images, to form a two-dimensional support, and risks misjudgment or information omission.

[0003] In addition, in the existing technology, the conversion from risk identification to decision-making requirements relies on manual experience, such as data resolution and collection frequency, lacking standardized mapping and dynamic quantification mechanisms, which leads to low matching degree of decision-making requirements and actual risks and insufficient multi-source data scheduling efficiency.

[0004] At the same time, the existing technology only focuses on the one-way process of risk identification and decision-making output, without establishing a correlation feedback mechanism between decision-making effect and front-end data collection and data fusion: neither the actual usage rate of multi-source fusion data in decision-making reports nor the landing effect of decision-making suggestions is evaluated, and data fusion parameters cannot be optimized and iterated, making it difficult to guarantee the stability and reliability of decision-making support, and without considering the resource consumption caused by decision-making, leading to high decision-making cost.

[0005] Therefore, a method for supporting intelligent decision-making of border defense public opinion and multi-source data fusion is needed to solve the above problems. SUMMARY

[0006] The application provides a method for supporting intelligent decision-making of border defense public opinion and multi-source data fusion.

[0007] A method for supporting intelligent decision-making of border defense public opinion and multi-source data fusion, comprising the following steps: S1: collecting preset type public opinion data related to border defense and preset type basic multi-source data, removing irrelevant public opinion data and multi-source data by using a data filtering algorithm, and completing data preliminary screening; S2: aligning the preliminary screened public opinion data and multi-source data in semantic and spatial dimensions, cleaning the aligned data to handle data conflicts, and obtaining a correlation fusion data set; S3: determining a risk type through the correlation fusion data set, extracting corresponding data resolution, collection frequency and other decision-making requirement indexes, and prioritizing the requirement indexes; S4: integrating the scheduled data with the public opinion data obtained in S2 by the demand index and priority scheduling multi-source data, to generate a decision support report; S5: collecting the application effect feedback of the decision support report, and optimizing the data collection focus and the correlation fusion data set based on the feedback.

[0008] As preferred, the data filtering algorithm is used in S1 to remove the public opinion data and multi-source data irrelevant to the border defense, to complete the data preliminary screening, wherein the data filtering algorithm includes a combined filtering algorithm, including keyword filtering and feature filtering two sub-processes: the keyword filtering adopts a combination of Boolean filtering and weighted filtering, retains the public opinion data containing a preset keyword combination, and retains the public opinion data with a preset theme similarity reaching a preset threshold; the feature filtering retains the multi-source data within a preset border geographic range based on geographic coordinates and data quality, and eliminates the data not meeting the preset threshold.

[0009] As preferred, the preliminary screened public opinion data and multi-source data are aligned in semantic and spatial dimensions in S2, including the following contents: S21: binding the preliminary screened public opinion data keywords with the multi-source data features, while obtaining the geographic coordinates of the multi-source data, correlating the public opinion data with the multi-source data corresponding to the geographic coordinates, establishing a public opinion and data feature dictionary, and the binding of the keywords and data features adopts a feature matching algorithm; S22: when a new public opinion type and multi-source data type appear, performing feature matching incremental training, adjusting the weight parameters based on the historical matching data through an iterative optimization method, including the new public opinion keywords and new multi-source data features in the public opinion and data feature dictionary, and updating the weight parameters in the feature matching algorithm.

[0010] As preferred, the aligned data is cleaned to handle data conflicts, including the following contents: S23: establishing a conflict judgment standard, assigning weights to the preliminary screened public opinion data and multi-source data according to the information source credibility, when the features of the two are inconsistent, calculating the credibility weighted value, if the weighted value is higher than a preset threshold, it is judged as no conflict, if it is lower than the preset threshold, it is judged as conflict; if it is judged as conflict, triggering the data supplement process of the corresponding area, re-aligning the supplement data in semantic and spatial dimensions, and processing data conflicts until the conflict-free correlation fusion data set is obtained.

[0011] As preferred, the S3 determines the risk type by associating the fused data set, extracts the corresponding data resolution, collection frequency class decision demand index, and includes the following contents: a mapping library of risk type and decision demand is established, the determined risk type is matched with the mapping library entry, the data resolution and collection frequency index are extracted, and the index is quantified based on the risk urgency; the risk urgency is divided into several levels according to the preset grading standard, and different index quantization results are corresponded.

[0012] As preferred, the S3 prioritizes the demand index, including the following contents: the border area weight is preset according to the importance of the border area, the demand priority score is calculated, the demand index is sorted according to the score, and if there are multiple risk demands in the same area, the demand index is combined and the parameters with the highest resolution and the highest collection frequency are retained.

[0013] As preferred, the S3 calls the public opinion keyword features and multi-source data anomaly features of the associated fused data set when judging the risk, and if the risk has no said features, sends a re-fusion instruction, adjusts the parameters of semantic and spatial dimension alignment again, performs secondary fusion on the data after preliminary screening and re-enters S3.

[0014] As preferred, the training of the scheduling model in the S4 includes the following contents: Based on the historical data within a preset time, including data source state, demand index, and scheduling result data, the training data is updated daily within a preset time period, and the model scheduling success rate is kept stable at more than 90%; When performing scheduling, the model calls the demand index and priority in the S3, and schedules the data source corresponding to the highest priority demand first, and if the data source is not available, the secondary data source is selected and the collection parameters are adjusted.

[0015] As preferred, the decision support report in the S4 includes three modules of risk profile, data evidence, and decision suggestion, the profile module includes the determined risk type, risk location, influence range, and risk urgency; the data evidence module includes the key data of the associated fused data set; the decision suggestion module generates targeted suggestions based on the risk type and data evidence; after the report is generated, it is encrypted and pushed to the preset border defense command system through a preset encryption algorithm, including public opinion summary, multi-source data annotation map, and data collection time; the decision suggestion module generates targeted suggestions based on the risk type and data evidence.

[0016] Preferably, in step S5, feedback on the application effect of the decision support report is collected, and data utilization rate and the rate of achievement of treatment effect are used as quantitative evaluation indicators. The data utilization rate includes the proportion of the amount of related and fused data cited in the decision support report to the total amount of related and fused dataset. If the utilization rate is lower than a preset threshold, an instruction is sent to adjust the data collection ratio. The rate of achievement of treatment effect is the proportion of the expected goals achieved after the decision recommendations are implemented. If the rate of achievement is lower than a preset threshold, an instruction is sent to update the weight parameters of the feature matching algorithm and to update the entry parameters of the risk and demand mapping library.

[0017] The present invention has the following beneficial effects: 1. By establishing a semantic and spatial alignment mechanism and conflict handling rules for public opinion and multi-source data, the two have been deeply integrated, enabling risk assessment to simultaneously possess qualitative descriptions of public opinion risk points and quantitative evidence from multi-source data. This effectively improves the comprehensiveness and accuracy of border risk identification and avoids assessment bias caused by a single information source.

[0018] 2. This method achieves automated transformation from risk type to data resolution and collection frequency by constructing a risk-decision demand mapping library and a priority quantification model, reducing manual intervention. At the same time, combined with a multi-source data dynamic scheduling mechanism, it ensures that the scheduled data is highly matched with decision-making needs, significantly improving the efficiency and accuracy of border defense decision support.

[0019] 3. This method optimizes the key points and fusion parameters of front-end data collection through a decision-making effect feedback mechanism, thereby continuously improving the adaptability of the system to border defense scenarios during long-term operation and further ensuring the stability and reliability of decision support. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a smart decision support method for fusing border defense public opinion and multi-source data according to the present invention; Figure 2 This is a schematic diagram of an intelligent decision support device for fusing border defense public opinion and multi-source data, according to the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions in the embodiments of this invention will be clearly described below in conjunction with the examples.

[0022] like Figure 1 As shown, this invention proposes a smart decision support method that integrates border defense public opinion with multi-source data, specifically including the following steps: S1, frontier public opinion and basic multi-source data acquisition, collect earthquake, military public opinion data related to the frontier, collect remote sensing image, unmanned aerial vehicle image type basic multi-source data, use data filtering algorithm to remove irrelevant public opinion data and multi-source data, complete the data preliminary screening; S2, public opinion and multi-source data association fusion, the semantic and spatial dimension alignment of the public opinion data and multi-source data in S1 is performed, the data conflict is arbitrated after cleaning to obtain the association fusion data set; S3, risk judgment and frontier decision demand extraction, based on the association fusion data set in S2, the risk type is determined, the corresponding data resolution, acquisition frequency type decision demand index is extracted, and the demand index is prioritized; S4, multi-source data dynamic scheduling and decision support information generation, according to the decision demand index and priority in S3, the multi-source data is scheduled, the data after scheduling and the public opinion risk point in S2 are integrated, and the decision support report is generated in fixed format; S5, decision feedback and data, public opinion collection optimization, collect the application effect feedback of the decision support report in S4, optimize the data collection key in S1 and the fusion parameters in S2 based on the feedback.

[0023] S2-1: public opinion-data feature binding and geographical association, including: establishing a special public opinion-data feature dictionary for the frontier, binding the keywords of the public opinion data after preliminary screening in S1 with the multi-source data features, and obtaining the geographical coordinates of the multi-source data in S1, associating the public opinion data with the multi-source data corresponding to the geographical coordinates; Wherein, the feature matching algorithm is used to realize the binding of keywords and data features, and the cosine similarity formula is used for matching degree calculation: In the formula is the feature vector of the public opinion keyword, is the feature vector of the multi-source data, is the weight coefficient of the public opinion keyword, is the weight coefficient of the multi-source data, is the vector dot product, , The module length of the weighted vector is respectively.

[0024] S2-2: dynamic alignment rule update, including: when new public opinion types or new multi-source data types not covered in S1 appear, trigger the feature matching incremental training process, include the new public opinion keywords, new multi-source data features in the public opinion-data feature dictionary, and update the 、 coefficient of the feature matching algorithm; wherein, the incremental training is based on historical matching data, and the gradient descent method is used for iterative optimization , This ensures that the new features match the existing features consistently.

[0025] S2-3: Conflict determination and supplementary data collection, including: establishing conflict determination criteria, assigning weights to the public opinion data and multi-source data after the initial screening in S1 according to the credibility of the information source, and calculating the credibility weighted value when the two characteristics are inconsistent. If the weighted value is higher than the threshold, it is determined that there is no conflict, and if it is lower than the threshold, it is determined that there is a conflict. The formula for weighted credibility is as follows: , For weighted values, official public opinion Unofficial public opinion High-resolution remote sensing Low-resolution remote sensing , The validity of the information source is represented by 1 for valid and 0 for invalid.

[0026] Furthermore, when providing data to S2 in step S1, risk type labels are marked on the public opinion data, and geographical coordinates and type labels are marked on the multi-source data. S2 calls the labels to improve alignment efficiency. If the integration degree of a certain area is lower than the set value, S2 sends an instruction to S1, and S1 strengthens the data collection frequency and public opinion monitoring range of that area.

[0027] S3-1: Risk-demand mapping library establishment and quantification, including: establishing a mapping library between risk types and decision-making needs, matching the risk types determined in S2 with the entries in the mapping library, extracting resolution and collection frequency indicators, and quantifying indicators based on risk urgency. The formula for the sampling frequency is: , For frequency, For urgency, Based on the coefficient, This is the influence coefficient.

[0028] S3-2: Ranking and merging of multiple risk requirements, including: introducing border defense area weights. Calculate the priority score of the requirements. ,according to Sorting; when there are multiple risk requirements in the same area, merge indicators and retain high-resolution, high-frequency parameters.

[0029] Furthermore, when assessing risks in step S3, the public opinion keyword features and multi-source data anomaly features of the associated fusion dataset in S2 are called; if the risk is not supported by the aforementioned features, S3 sends a refusion instruction to S2, S2 readjusts the alignment parameters, performs a second fusion of the S1 data and feeds back the new dataset to S3.

[0030] Furthermore, in step S4, when scheduling data, the demand metrics in S3 are called. The value is matched with the resolution and the frequency with the available time period of the data source; if the data source is unavailable, S4 sends an adjustment instruction to S3, and S3 lowers the corresponding indicator and recalculates. The value is then fed back to S4.

[0031] Furthermore, during step S5's feedback collection, the proportion of S2-related fusion data used in the decision support report in S4 and the processing effect are statistically analyzed; if the proportion is lower than the set value, S5 sends an instruction to S1 to adjust the data collection ratio; if the processing effect does not meet expectations, S5 sends an instruction to S2 to update the feature matching algorithm. , Send a command to S3 to update the parameters of the mapping library entry.

[0032] Example 1 This invention relates to the field of intelligent border security technology, and in particular to an intelligent decision support method that integrates public opinion on border defense with multi-source data, specifically including the following steps: S1. Acquisition of public opinion and basic multi-source data related to border defense: collect public opinion data related to earthquakes and military affairs, collect basic multi-source data such as remote sensing images and UAV images, and use data filtering algorithms to remove public opinion data and multi-source data that are not related to border defense to complete the initial screening of data. S2, Public Opinion - Multi-Source Data Association and Fusion: The public opinion data initially screened in S1 is aligned with the multi-source data in terms of semantics and spatial dimensions. The aligned data is cleaned to arbitrate data conflicts and obtain the association and fusion dataset. S3. Risk assessment and border defense decision-making requirements extraction: Based on the associated and fused dataset in S2, the risk type is determined, the corresponding data resolution and collection frequency decision-making requirements indicators are extracted, and the requirements indicators are prioritized. S4. Dynamic scheduling of multi-source data and generation of decision support information: Based on the decision demand indicators and priorities in S3, multi-source data is scheduled, and the scheduled data is integrated with the public opinion risk points in S2 to generate a decision support report in a fixed format. S5. Decision feedback and data / public opinion collection optimization: Collect application effect feedback of the decision support report in S4, and optimize the data collection focus in S1 and the fusion parameters in S2 based on the feedback.

[0033] S2-1: Public Opinion - Data Feature Binding and Geographic Association, including: establishing a border defense-specific public opinion - data feature dictionary, binding the public opinion data keywords after initial screening in S1 with multi-source data features, and simultaneously obtaining the geographic coordinates of the multi-source data in S1, associating the public opinion data with the corresponding geographic coordinates of the multi-source data; wherein, a feature matching algorithm is used to bind keywords and data features, and the matching degree is calculated using the cosine similarity formula: In the formula This is a feature vector of public opinion keywords. For multi-source data feature vectors, This refers to the keyword weighting coefficient for public opinion; for earthquake and military keywords, the value ranges from 0.6 to 0.9. The weighting coefficients for multi-source data features range from 0.7 to 0.95 for high-resolution images. For vector dot product, , These are the magnitudes of the weighted vectors, respectively. Feature similarity is measured based on the cosine similarity principle, and the basic formula is: This method only reflects the similarity of the original features; this method is designed for border defense scenarios and introduces... and Adjusting the weights of core features, earthquake-related public opinion is directly linked to rescue decisions. Using a value of 0.8-0.9, high-resolution remote sensing data, due to its rich detail, The parameter value is set to 0.85-0.95; the parameter value is based on the historical border defense public opinion data matching case training over the past 3 years. The cases are divided into a training set (70%) and a validation set (30%), and optimization is performed using gradient descent. and The final value range is determined when the accuracy of the validation set matching reaches 92% or higher.

[0034] S2-2: Dynamic alignment rule update, including: when a new public opinion type or new multi-source data type not covered in S1 appears, triggering the feature matching incremental training process, incorporating the new public opinion keywords and new multi-source data features into the public opinion-data feature dictionary, and updating the feature matching algorithm. , Coefficients; where incremental training is based on an expanded training set consisting of historical border defense data training set and newly added data. The expanded training set contains 500+ samples corresponding to the newly added public opinion types. Each sample contains public opinion keywords, matched multi-source data features, and manually labeled matching results. Iterative optimization is performed using stochastic gradient descent. , The number of iterations is set to 1000, and the learning rate is set to 0.01. After each iteration, the accuracy of the validation set is calculated. Training stops when the accuracy does not improve for 10 consecutive iterations, and the current accuracy is saved. , Coefficients; for example, when adding the public opinion type of "border geological disasters", the training set is expanded to include public opinion containing the keyword "border geological disasters" and samples with image features of "surface subsidence and landslides". After training, the coefficients of "border geological disasters" will be used. Set to 0.82, corresponding to image features Set it to 0.9.

[0035] S2-3: Conflict Determination and Supplementary Data Collection, including: establishing conflict determination criteria; assigning weights to the initial-screened public opinion data and multi-source data in S1 according to the credibility of the information sources; when the characteristics of the two are inconsistent, calculating a credibility weighted value; if the weighted value is higher than a threshold, it is determined that there is no conflict; if it is lower than the threshold, it is determined that there is a conflict; the credibility weighted value formula is as follows: , For weighted values, Official public opinion = 0.9 Unofficial public opinion = 0.4 High-resolution remote sensing = 0.85 Low-resolution remote sensing = 0.5 The validity of the information source is defined as 1 for valid and 0 for invalid, with a threshold of 0.7. The basic principle is a weighted summation based on the reliability of information sources. In border defense scenarios, official public opinion is highly authoritative. Set to a maximum of 0.9; unofficial public opinion fluctuates greatly due to the high degree of uncertainty surrounding the accuracy of information. Set to 0.4; high-resolution remote sensing can clearly present surface details. Setting it to 0.85 results in blurred details in low-resolution remote sensing. Set to 0.5; threshold 0.7 is based on historical conflict case statistics, when When the information source is consistent, the probability is over 88%, so the threshold is set to 0.7. If it is determined to be a conflict and the public opinion weighting value ( Higher than the data weighting value ( If the missing area is captured, the drone system in S1 will be triggered to recapture the missing area. The recapture area is within 10km of the geographic coordinates mentioned in the public opinion. The resolution of the recaptured image is set to ≤0.5m. After the recapture is completed, the recaptured data will be input into S2, and the semantic and spatial dimension alignment and conflict arbitration process will be re-executed until a conflict-free associated fusion dataset is obtained.

[0036] Furthermore, the data filtering algorithm in S1 includes a combined filtering algorithm, comprising two sub-processes: keyword filtering and feature filtering. Keyword filtering uses a combination of Boolean filtering and TF-IDF weighted filtering. The Boolean filtering stage retains public opinion data containing preset keyword combinations such as "border + earthquake," "border + military exercise," "border + military assembly," and "border + natural disaster." These preset keyword combinations are obtained by summarizing public opinion related to border risk events over the past five years, containing a total of 32 core keyword combinations. The TF-IDF weighted filtering stage calculates the similarity between the public opinion text and the "border risk" theme, constructs a border risk theme thesaurus, calculates the TF-IDF value of each theme word in the public opinion text, and sums them to obtain the similarity between the public opinion and the theme. The similarity threshold is set to 0.6, retaining public opinion data with a similarity ≥ 0.6. The multi-source data feature filtering stage is based on geographic coordinates and image quality. Geographic coordinate filtering retains data within a 50km range on both sides of the border. The border vector data is converted into latitude and longitude ranges through the GIS system, and multi-source data outside this range are removed. Image quality filtering calculates the ambiguity and signal-to-noise ratio of the image. The ambiguity is calculated using the Laplace variance method. Images with a variance <100 are judged as blurry, and images with a signal-to-noise ratio <20dB are judged as having excessive noise. Blurry and excessively noisy image data are removed.

[0037] S3-1: Risk-Demand Mapping Library Establishment and Quantification, including: establishing a mapping library between risk types and decision-making needs. The mapping library is stored using a relational database. Each entry contains a risk type ID, risk type name, recommended resolution range, basic collection frequency, and urgency level standard. Among them, the risk type name corresponds one-to-one with the risk type determined in S2, such as "border earthquake disaster", "abnormal border military activities", "illegal border crossing", etc. Recommended resolution range is set according to risk identification needs. For border earthquake disasters, details such as building damage and road interruptions need to be identified, with a recommended resolution of 0.5m-1m. For abnormal border military activities, equipment types and personnel numbers need to be identified, with a recommended resolution of 0.3m-0.8m. The basic data collection frequency is the default frequency when the urgency level E=3. When based on risk urgency quantification indicators, the urgency level E is divided into 1-5 levels according to the scope and severity of risk impact: "Border earthquake disasters" E=5, "abnormal border military activities" E=4, and "illegal border crossings" E=2. The data collection frequency formula is... , Frequency (times / hour), E represents urgency. The base coefficient is 0.2-0.5. Impact coefficient = 0.3-0.6; The basic model is an exponential function. Since the impact of urgency on the collection frequency increases non-linearly, high urgency risks require a significant increase in frequency to ensure real-time performance. The minimum acquisition interval in border defense scenarios is limited by device battery life and data transmission bandwidth, serving as the base frequency when urgency level E=0. Set it to 0.2-0.5 to ensure a minimum frequency of ≥0.2 times / hour; Controlling the rate of increase in frequency is crucial, as military risks change rapidly. Set at 0.5-0.6, earthquake-related risks change relatively slowly. Set it to 0.3-0.5; For example, E=5, , hour, The frequency is set at 10 times per hour to meet the real-time monitoring needs of abnormal military activities.

[0038] S3-2: Ranking and merging of multiple risk requirements, including: introducing border defense area weights. The regional weights are set based on the importance of the border defense areas. Border ports are valued for their high density of people and goods and their strategic importance. Border towns have residents, The border no-man's-land is sparsely populated. Calculate the priority score for requirements. , Higher values ​​indicate higher priority. Demands are sorted in descending order of value; when multiple risk demands exist in the same area, resolution and frequency indicators are combined, with resolution taking the highest value among all demands and frequency taking the highest value among all demands; for example, a border crossing simultaneously experiences earthquake disasters: E=5, f=3 times / hour, k=0.9, and illegal border crossings: E=2, f=2 times / hour, k=0.9, calculate... , "Earthquake disaster" has a higher priority; when merging indicators, the resolution is 0.8m of the 1m resolution for "earthquake disaster" and the 0.8m resolution for "illegal border crossing", and the frequency is 3 times per hour.

[0039] Furthermore, the multi-source data scheduling in step S4 specifically adopts a reinforcement learning-based scheduling model. The scheduling model aims to "meet the demand indicators while minimizing resource consumption." The state space includes the current state of each data source, data transmission bandwidth, and remaining device battery life. The action space includes selecting the data source, setting data acquisition parameters, determining the transmission period, and the reward function. , For the degree of matching of demand indicators, The percentage of resource consumption below a threshold is used for evaluation. For example, 1 point is awarded for device battery life consumption below 50%, and 0 points are awarded for consumption above 80%; 1 point is awarded for bandwidth usage below 60%, and 0 points are awarded for consumption above 90%. The scheduling model is trained using historical data from the past year, including data on data source status, demand metrics, and scheduling results. The training period is 30 days, with daily updates to ensure a stable scheduling success rate above 90%. During scheduling execution, the model calls upon the demand metrics in S3... Value, priority scheduling The highest required data source is selected. If the data source is unavailable, the second-best data source is selected and the acquisition parameters are adjusted. For example, if the required resolution is 0.5m, a drone is preferred. If a drone is unavailable, a high-resolution remote sensing is selected, and the acquisition frequency is increased from 10 times / hour to 12 times / hour to compensate for the insufficient resolution.

[0040] Furthermore, the decision support report generation in step S4 uses a fixed template, which includes three modules: risk overview, data evidence, and decision recommendations. The risk overview module is filled in with the risk type, risk location, scope of impact, and urgency level E determined in S3. The data evidence module inserts key data from the associated and fused dataset in S2, including public opinion summaries, multi-source data annotation maps, and data collection time. The decision recommendations module is generated based on the risk type and data evidence. For "border earthquake disaster," recommendations include the number of rescue teams deployed, rescue routes, and required rescue equipment. For "abnormal border military activities," recommendations include increasing the number of patrol police, adjusting monitoring points, and setting early warning levels. After the report is generated, it is pushed to the border defense command system through an encrypted interface. The data is encrypted using the national cryptographic SM4 algorithm during the push process to ensure transmission security.

[0041] Furthermore, the decision-making effect feedback in step S5 adopts quantitative evaluation indicators, including data utilization rate and the rate of achievement of treatment effects; the data utilization rate is the proportion of the amount of S2-related fusion data cited in the decision support report to the total amount of S2 output data, calculated by the following formula: , For the amount of referenced data, Given the total data volume, the utilization threshold is set to 80%. The data utilization was deemed sufficient; the compliance rate was the proportion of decisions made after they were implemented and the expected goals were achieved. The expected goals for the "border earthquake disaster" were "the evacuation of affected people within 24 hours" and "road closures repaired and reopened within 48 hours." One point was awarded for each goal achieved, and zero points for failure to achieve them. The compliance rate was... The threshold for achieving the target is set at 85%. The action was deemed effective; when optimizing based on feedback, if S5 sends a command to S1 to adjust the data acquisition ratio and reduce the acquisition volume of low-usage data types. For example, if the drone imagery usage rate is only 60%, the drone acquisition frequency will be reduced from 10 times / hour to 8 times / hour. S5 sends a command to S2 to update the feature matching algorithm. , For example, if earthquake disaster response falls short of expectations because the "road crack" feature in the images is not effectively identified, the coefficient for the "road crack" feature will be... The resolution was increased from 0.8 to 0.88; at the same time, S5 sent a command to S3 to update the parameters of the risk-demand mapping library entries. For example, the recommended resolution for "earthquake disaster" was reduced from 1m to 0.8m to improve the ability to identify details.

[0042] Example 2 This invention relates to the field of intelligent border security technology, and in particular to an intelligent decision support method that integrates public opinion on border defense with multi-source data, specifically including the following steps: S1. Acquisition of public opinion and basic multi-source data related to border defense: collect public opinion data related to earthquakes and military affairs, collect basic multi-source data such as remote sensing images and UAV images, and use data filtering algorithms to remove public opinion data and multi-source data that are not related to border defense to complete the initial screening of data. S2, Public Opinion - Multi-Source Data Association and Fusion: The public opinion data initially screened in S1 is aligned with the multi-source data in terms of semantics and spatial dimensions. The aligned data is cleaned to arbitrate data conflicts and obtain the association and fusion dataset. S3. Risk assessment and border defense decision-making requirements extraction: Based on the associated and fused dataset in S2, the risk type is determined, the corresponding data resolution and collection frequency decision-making requirements indicators are extracted, and the requirements indicators are prioritized. S4. Dynamic scheduling of multi-source data and generation of decision support information: Based on the decision demand indicators and priorities in S3, multi-source data is scheduled, and the scheduled data is integrated with the public opinion risk points in S2 to generate a decision support report in a fixed format. S5. Decision feedback and data / public opinion collection optimization: Collect application effect feedback of the decision support report in S4, and optimize the data collection focus in S1 and the fusion parameters in S2 based on the feedback.

[0043] This embodiment is applied to the scenario of "border earthquake disaster emergency response". In this scenario, after an earthquake occurs in the border area, it is necessary to quickly integrate earthquake-related public opinion and multi-source image data to support rescue decision-making and disaster monitoring. The following is a detailed description of the technical implementation process of each step: In step S1, the border defense public opinion collection focuses on the theme of "border earthquake." The collection channels include the disaster reporting section of the official website of the Ministry of Emergency Management, the official Weibo API of provincial earthquake bureaus, the official WeChat accounts of border area emergency management bureaus, and earthquake news reporting interfaces of authoritative media. The collection cycle is set for 72 hours after the earthquake, with a collection frequency of 30 minutes / time for the first 24 hours, 1 hour / time for 24-48 hours, and 2 hours / time for 48-72 hours, ensuring the real-time nature of public opinion during the high-urgency phase. The collected content includes the public opinion title, release time, publishing entity, core content, and the name of the border area mentioned. Basic multi-source data, excluding remote sensing imagery, is also collected. In addition to UAV imagery, ground sensor data deployed by border defense forces were added, including vibration sensors to monitor the intensity of aftershocks, tilt sensors to monitor landslide risks, and temperature and humidity sensors to assist in assessing environmental conditions in disaster-stricken areas. Remote sensing imagery was acquired using Gaofen-2 and Gaofen-6 satellites, with acquisition frequency during satellite transit periods. UAV imagery was acquired using a swarm of multi-rotor UAVs deployed by border defense forces, with an initial acquisition range of 30km around the earthquake epicenter, which was subsequently adjusted based on public opinion. Ground sensor data was acquired every 5 minutes, with vibration sensors recording amplitude and frequency, tilt sensors recording tilt angles, and temperature and humidity sensors recording real-time temperature and humidity.

[0044] When the data filtering algorithm in step S1 is executed, the public opinion filtering first uses Boolean filtering to retain public opinion containing keyword combinations such as "border + earthquake", "border + aftershock", "border + disaster", "border + rescue", "border + road interruption", and "border + house damage", while removing general earthquake public opinion that does not contain information about border areas. Secondly, it uses TF-IDF weighted filtering to construct a border earthquake disaster topic thesaurus, calculates the similarity between public opinion text and the topic, sets the similarity threshold to 0.65, and retains public opinion with a similarity ≥ 0.65. For example, if a public opinion contains "A 3.2 magnitude aftershock occurred near a border port, and some houses developed cracks", the TF-IDF weighted filtering will retain the relevant information. - The IDF similarity is calculated to be 0.72, so it is retained. When filtering multi-source data, remote sensing and UAV images are filtered based on geographic coordinates. The latitude and longitude of the earthquake epicenter are converted into a vector range through the GIS system. Images within 20km of this range are retained, and images outside the range are removed. Image quality filtering uses the Laplace variance method to calculate ambiguity. Images with a variance ≥150 are judged as clear and are retained. When filtering ground sensor data, vibration sensors retain data with amplitude >0.1g, tilt sensors retain data with tilt angle >5°, and temperature and humidity sensors retain data with temperature <0℃ or >35℃ and humidity >85%.

[0045] Step S2: First, obtain the public opinion data and multi-source data after the initial screening in S1. Preprocess the public opinion data, including text segmentation, keyword extraction, and geographic entity recognition. A rule-based geographic entity recognition algorithm is used to extract the name of the border area from the core content of the public opinion, such as "a certain border town", and convert it into latitude and longitude coordinates through the Gaode Map API. Preprocess the multi-source data. Perform radiometric correction, geometric correction, and geographic registration on remote sensing and UAV imagery. Perform 3σ criterion outlier removal on ground sensor data, removing data that exceeds the mean ± 3 times the standard deviation. Standardize the data by converting vibration amplitude and tilt angle into standardized values ​​in the 0-1 range.

[0046] S2-1: When binding public opinion data features with geographic association, first update the border defense-specific public opinion data feature dictionary, adding feature mappings related to "border earthquakes," including image features of "surface cracks, building damage, and road interruption" and sensor features of "vibration sensor amplitude > 0.1g" for "earthquakes," "vibration sensor amplitude > 0.05g" for "aftershocks," "landslides" for "tilt sensor tilt angle > 10°" and "changes in surface vegetation cover" for "landslides," and "building damage" for "building roof collapse and wall cracks" for "building damage." Feature binding uses the aforementioned feature matching algorithm to calculate the public opinion keyword feature vector. With multi-source data feature vectors The cosine similarity, for example, if a certain public opinion keyword is "damage to houses in a border town". This is the vector representation of the keyword. Vector representation of the "building roof collapse" feature in UAV imagery, keywords for building damage. Set to 0.85, roof collapse characteristics Set it to 0.92, calculate If the similarity is greater than 0.8, a match is determined. When geographic association is performed, the latitude and longitude coordinates N40°28′, E80°15′ converted from public opinion are imported into the GIS system and matched with multi-source data within 10km around the coordinates, including drone images, vibration sensor data, and tilt sensor data collected in the area, to form a geographic association data group.

[0047] S2-2: Dynamic alignment rule updates are triggered once in this scenario. 12 hours after the earthquake, a new public opinion type, "border earthquake triggers landslide," emerges. This type was not included in the initial dictionary. At this time, the incremental training process is triggered, collecting 500 samples of this type of public opinion. These, along with 1500 historical "border earthquake" samples, form an expanded training set of 1600 samples and a validation set of 400 samples. Iterative optimization is performed using stochastic gradient descent. , The iteration count was 1000, the learning rate was 0.01, and after training, the keyword "border earthquake triggered landslide" was... Set to 0.88, the "mountain displacement" image feature Set to 0.93, for the characteristic of "tilt sensor tilt angle > 15°". Set to 0.90, the updated dictionary will be used for subsequent public opinion-data alignment.

[0048] S2-3: During conflict assessment and supplementary mining, a public opinion event originated from the official Weibo account of a provincial earthquake bureau. The event, posted at 14:00, stated, "A 4.0 magnitude aftershock occurred in a border town, suspected to have triggered a small-scale landslide." This was official public opinion. , The corresponding multi-source data is drone imagery acquired at 14:05, which is high-resolution. , Because the imagery did not cover the suspected landslide area; tilt sensor data from 14:00-14:05 showed a tilt angle of 8°. Not applicable; sensor data alone is judged as "no landslide risk"; calculate the confidence weighted value. The system determined no conflict, but the drone imagery did not cover the key area. At this point, a drone re-shooting command was triggered. The re-shooting area was the "mountain on the west side of a border town" mentioned in the public opinion, with a latitude and longitude range of N40°30′-N40°32′ and E80°18′-E80°20′. Two drones were dispatched to simultaneously re-shoot, with a re-shooting resolution of 0.3m and a flight altitude of 100m. After the re-shooting was completed, the acquired imagery showed that there were traces of a small-scale landslide on the mountain. The re-shooting imagery was input into S2, and semantic alignment was re-performed. The cosine similarity between the public opinion keyword "mountain landslide" and the imagery feature "landslide traces" was 0.91, completing the association fusion and obtaining an association fusion dataset containing public opinion, re-shooting imagery, and sensor data.

[0049] Step S3: First, risk assessment is performed based on the associated fusion dataset in S2, extracting public opinion keywords, image features, and sensor data from the fusion data; using a risk type matching algorithm, based on rule reasoning, the extracted features are matched with a preset risk type rule base. The rule for "border earthquake secondary disasters" in the rule base is "containing earthquake and aftershock keywords + landslide image features + building damage image features + vibration amplitude > 0.1g". The fusion data meets this rule, and the risk type is determined to be "border earthquake secondary disasters".

[0050] S3-1: When calling the risk-demand mapping library, match the "Border Earthquake Secondary Disasters" entry. The recommended resolution range for this entry is 0.3m-0.8m. It is necessary to identify the landslide area and details of building damage. The basic data collection frequency (E=3) is 4 times / hour. When determining the risk urgency E, based on the impact range and degree of harm, refer to the urgency classification standard. If E=5 is determined, the impact range is >100m², and large-scale personnel evacuation is required. The data collection frequency is calculated using the formula. This scenario depicts a secondary disaster following an earthquake. Take 0.4, Substituting 0.45, we get... The sampling frequency is 4 times per hour; the initial decision-making requirements are determined as follows: resolution 0.3m, sampling frequency 4 times per hour, and geographical range of a border town, latitude and longitude N40°28′-N40°33′, E80°15′-E80°22′.

[0051] S3-2: When prioritizing demands, a certain border town has a regional weight due to its inhabitants. Calculate priority score Meanwhile, the border area also has a demand for "aftershock monitoring at a certain border port" with E=3, f=2 times / hour, and k=0.9. The requirement for "landslide monitoring in a border no-man's-land" is E=2, f=1 times / hour, k=0.3. ;according to Based on the priority ranking, the demand for "landslide + house damage in a border town" has the highest priority and will be allocated resources first.

[0052] Step S4: Multi-source data scheduling invokes a reinforcement learning-based scheduling model. The model inputs the demand metrics from S3, with a resolution of 0.3m and a frequency of 4 times / hour. The current data source status is as follows: The drone swarm currently has 3 drones available, with drone 1 having 35 minutes of remaining flight time, drone 2 having 30 minutes, and drone 3 having 25 minutes. The next remote sensing imagery overpass time is 16:00. All ground sensors are available. The bandwidth remaining is 70%. During the scheduling model decision-making process, because only drones can meet the required resolution of 0.3m, and the frequency is 4 times / hour requiring data collection every 15 minutes, a drone rotation scheduling scheme is adopted: Drone 1 is scheduled from 14:30 to 14:45, Drone 2 from 14:45 to 15:00, Drone 3 from 15:00 to 15:15, and Drone 1 is rescheduled after recharging from 15:15 to 15:30. Ground sensor data is collected every 5 minutes and transmitted synchronously. Remote sensing imagery is added to the scheduling data after collection at 16:00.

[0053] When generating the decision support report in step S4, the "Border Earthquake Disaster" template is used. The risk overview module is filled with the risk type "Secondary Disaster of Border Earthquake (Landslide + House Damage in a Border Township)", risk location (latitude and longitude N40°30′, E80°19′), affected area (landslide area 500m², 3 houses damaged, affecting 12 residents), urgency level E=5, and current aftershock situation (1 aftershock of magnitude 3.2, amplitude 0.12g). The data verification module inserts landslide images taken by drone at 14:30, house damage images at 14:45, and images from 14:00 onwards. 14:30 Vibration sensor amplitude curve, 16:00 Remote sensing imagery; Decision suggestion module generates: 1. Rescue team: Dispatch a 10-person rescue team, carrying demolition equipment and first aid kits, to the damaged housing area in a border town; 2. Evacuation plan: Organize 12 affected residents to evacuate to a temporary resettlement point in the town, avoiding landslide risk areas and road cracks along the evacuation route; 3. Monitoring focus: Monitor the landslide area with drones every 15 minutes and collect tilt data every 5 minutes to be vigilant for landslide expansion; 4. Aftershock warning: Continuously monitor vibration sensors and issue a warning when aftershocks exceed magnitude 3.0. After the report is generated, it is pushed to the border rescue command terminal via an SM4 encrypted interface and simultaneously synchronized to the provincial emergency management platform.

[0054] Step S5: The feedback collection period for decision-making effectiveness is 24 hours after the release of the decision recommendation. The collected data includes: data utilization rate. The decision report references S2 correlated and fused data; S2 outputs total data and calculates... The percentage of cases below the threshold (80%) indicates insufficient data utilization; the rate of compliance with treatment objectives is also low. Of the four objectives in the decision-making recommendations, the "personnel evacuation" objective was achieved: 12 people evacuated within 24 hours; the evacuation of 12 people was completed within 10 hours. The "rescue team deployment" objective, the "landslide monitoring" objective, and the "aftershock early warning" objective were also achieved. All four objectives were met. The decision was deemed effective.

[0055] Based on feedback optimization, S5 sends instructions to S1 to adjust the data collection focus: Because the utilization rate of public opinion data is low (5 / 12≈41.7%), the analysis indicates that the low-utilization public opinion is related to widespread aftershock reports. Therefore, the keywords for public opinion collection are adjusted, adding more precise keywords such as "a border town + specific village," "a border town + relocation," and "a border town + resettlement site" to reduce the collection of widespread aftershock public opinion; the utilization rate of UAV imagery is 53.3% (8 / 15), so the collection scope is adjusted to focus on three core areas: damaged houses, landslides, and road cracks, reducing the ineffective collection area; the utilization rate of ground sensor data is 46.7%, so the frequency of temperature and humidity sensor collection is reduced from 5 minutes / time to 10 minutes / time, while retaining the frequency of vibration and tilt sensors; S5 sends instructions to S2 to update the feature matching algorithm. The coefficient, due to the low citation of the "road crack" feature in the report and insufficient accuracy in identification, will be used to determine the accuracy of the "road crack" image feature. The matching priority was increased from 0.85 to 0.92; S5 sent an instruction to S3 to update the basic collection frequency of the "Border Earthquake Secondary Disasters" entry in the risk-demand mapping library, from 4 times / hour to 3 times / hour, to ensure efficient use of resources.

[0056] Example 3 This invention relates to the field of intelligent border security technology, and in particular to an intelligent decision support method that integrates public opinion on border defense with multi-source data, specifically including the following steps: S1. Acquisition of public opinion and basic multi-source data related to border defense: collect public opinion data related to earthquakes and military affairs, collect basic multi-source data such as remote sensing images and UAV images, and use data filtering algorithms to remove public opinion data and multi-source data that are not related to border defense to complete the initial screening of data. S2, Public Opinion - Multi-Source Data Association and Fusion: The public opinion data initially screened in S1 is aligned with the multi-source data in terms of semantics and spatial dimensions. The aligned data is cleaned to arbitrate data conflicts and obtain the association and fusion dataset. S3. Risk assessment and border defense decision-making requirements extraction: Based on the associated and fused dataset in S2, the risk type is determined, the corresponding data resolution and collection frequency decision-making requirements indicators are extracted, and the requirements indicators are prioritized. S4. Dynamic scheduling of multi-source data and generation of decision support information: Based on the decision demand indicators and priorities in S3, multi-source data is scheduled, and the scheduled data is integrated with the public opinion risk points in S2 to generate a decision support report in a fixed format. S5. Decision feedback and data / public opinion collection optimization: Collect application effect feedback of the decision support report in S4, and optimize the data collection focus in S1 and the fusion parameters in S2 based on the feedback.

[0057] This embodiment is applied to the scenario of "abnormal monitoring of border military activities". In this scenario, the border intelligence department detects public opinion related to military activities in the overseas border area. It needs to combine multi-source data to identify the type, scale and location of the activities to support the deployment and monitoring decisions of border police forces. The following is a detailed description of the technical implementation process of each step: In step S1, the channels for collecting public opinion on border defense include international military news platforms, such as Jane's Defence Weekly's border military reports, military activity news from authoritative overseas media, military public opinion reports from the internal information system of border defense intelligence departments, and compliant information mentioning border military activities on social media platforms. The collection period is set to within 48 hours after the first appearance of public opinion, with a collection frequency of 15 minutes / time for the first 12 hours, 30 minutes / time for 12-24 hours, and 1 hour / time for 24-48 hours, to ensure the real-time nature of public opinion in highly dynamic scenarios. The collected content includes the title of public opinion, the time of publication, the publishing entity, the core content, the type of military activity mentioned (such as military exercises, equipment assembly, and enhanced patrols), the type of weapons and equipment involved (such as tanks, armored vehicles, and helicopters), and the border section mentioned.

[0058] In addition to remote sensing imagery and UAV imagery, the basic multi-source data now includes border defense radar data and infrared UAV imagery. The remote sensing imagery uses Gaofen-1 and Gaofen-3 satellites, and is collected during satellite transit periods. The UAV imagery includes visible light UAVs and infrared UAVs, deployed at border defense stations within 10km on both sides of a certain border line. The initial collection range is 20km around the border line section mentioned in public opinion. The border defense radar data comes from ground-based early warning radars deployed on the border, and is collected every 2 minutes, recording the number of targets, average speed, and target cluster area. The infrared UAV imagery is collected every hour during the day and every 30 minutes at night. Due to the greater concealment of military activities at night, the frequency needs to be increased.

[0059] The data filtering algorithm in step S1 is as follows: First, Boolean filtering is used to retain public opinion containing keyword combinations such as "border + military exercise", "border + equipment assembly", "border + tank", "border + armored vehicle", "border + helicopter", and "border line + military", while removing public opinion that does not mention specific military equipment or border line sections. Second, TF-IDF weighted filtering is used to construct an abnormal topic thesaurus of border military activities, calculate the similarity between public opinion text and topic, and set the similarity threshold to 0.7. Public opinion with a similarity of ≥0.7 is retained. For example, a public opinion containing "Foreign armored forces were found to be assembling near a 150-kilometer section of a border line, suspected of preparing for an exercise" is retained. The TF-IDF similarity was 0.81 and was retained. When filtering multi-source data, remote sensing, UAV, and infrared imagery were filtered based on geographic coordinates. The "150-kilometer section of a certain border" mentioned in the public opinion was converted into a latitude and longitude range (N41°10′-N41°20′, E81°30′-E81°40′) through the GIS system, and the imagery within 15km of this range was retained. In the image quality filtering, the blur threshold for visible light and infrared UAV images was set to Laplace variance ≥200. Radar data filtering retained data with ≥5 targets, average speed ≥10km / h, and cluster area ≥1000m², while removing scattered small target data.

[0060] Step S2: First, preprocess the public opinion data after the initial screening in S1, including text segmentation, military entity identification, and preliminary judgment of public opinion credibility; preprocess the multi-source data, perform radiometric correction, geometric correction, and ground feature classification on remote sensing images, perform image stabilization, thermal signal enhancement, and target labeling on UAV and infrared images, and perform target clustering on radar data using the DBSCAN clustering algorithm to group targets with a distance of <100m into one cluster; remove false targets by removing non-military targets with a speed >100km / h based on a speed threshold.

[0061] In step S2, when performing semantic and spatial dimension alignment, and executing S2-1 public opinion-data feature binding and geographic association, the border defense-specific public opinion-data feature dictionary is updated, adding feature mappings related to "border military activities": "military exercises" correspond to image features of "equipment movement trajectory, firing marks, and temporary tents" and the feature of "radar target cluster movement speed 10-30km / h"; "equipment assembly" corresponds to image features of "tank and armored vehicle clusters, and helicopter take-off and landing points" and the feature of "radar target cluster area > 5000m²"; "tanks" correspond to visible light image features of "track marks and turret shape" and infrared features of "high-temperature thermal signal"; "armored vehicles" correspond to image features of "wheeled chassis and armored shell" and infrared features of "medium-temperature thermal signal". Feature binding uses the aforementioned cosine similarity formula. For example, if a public opinion keyword is "foreign armored forces assembling along a 150km section of a border line",... This is the keyword vector. The feature vector of the "cluster of 10 armored vehicles" in the UAV imagery. Set it to 0.9. Set it to 0.95, and calculate. When determining a match, the latitude and longitude range of the "150-kilometer section of a certain border" mentioned in the public opinion is matched with the geographical coordinates of the UAV images and radar data. Two UAV images, three sets of radar data, and one infrared image within 15km of this range are retained to form a related data group.

[0062] The dynamic alignment rule update in step S2-2 is triggered in this scenario. When a new type of public opinion emerges, namely "border + helicopter take-off and landing," this type was not included in the initial dictionary. 400 samples of this type of public opinion are collected, and these, along with 1600 historical samples of "border military activities," form an extended training set (1600 samples in the training set and 400 samples in the validation set). Iterative optimization is then performed using stochastic gradient descent. , The algorithm iterated 800 times with a learning rate of 0.02. After training, the keyword "helicopter takeoff and landing" was... Set to 0.92, the "helicopter shape" image feature Set to 0.96, the "radar target takeoff and landing speed" characteristic Set to 0.91; the updated dictionary will be used for subsequent alignment.

[0063] S2-3: During conflict assessment and supplementary data collection, a public opinion report from an internal border intelligence department stated that "helicopter take-off and landing activities were detected overseas along a 150-kilometer section of the border." The corresponding multi-source data was a high-resolution UAV image from 10:00 AM. , No helicopter detected; radar data from 10:00 to 10:05: 2 targets, speed 3 km / h. Not applicable; the location is suspected to be a helicopter takeoff or landing. Infrared imagery from 22:00 the previous day showed no helicopter thermal signal. ; Calculate the confidence weighted value The system determined no conflict, but the drone imagery failed to capture the target. An infrared drone re-capture command was triggered, with the re-capture area within 5km of the radar target location (latitude and longitude N41°15′, E81°35′), a re-capture time of 10:15, and a re-capture resolution of 0.5m. The re-captured imagery showed two helicopters parked in an open area. After inputting the re-captured imagery into S2 and realigning it, the cosine similarity between the public opinion "helicopter takeoff and landing" and the imagery "helicopter thermal signal" features was 0.94, completing the association fusion and obtaining an association fusion dataset containing public opinion, re-captured infrared imagery, and radar data.

[0064] Step S3: Based on the S2 associated fusion dataset, conduct risk assessment, extract public opinion keywords, image features, and radar data from the fusion data; through the risk type matching algorithm, match the "abnormal border military activities" rule in the preset rule base, and determine the risk type as "abnormal border military activities".

[0065] S3-1: When calling the risk-demand mapping library, match the "abnormal border military activities" entry. The recommended resolution range is 0.2m-0.5m, and the basic collection frequency (E=3) is 6 times / hour. The risk urgency level E is determined based on equipment scale and activity nature, referring to the grading standard, and is set to E=4. The collection frequency is calculated using the formula... Military scene Take 0.5, Take 0.55, substitute it into the equation, and we get... The sampling frequency was 5 times per hour; the decision-making requirements were determined as follows: resolution 0.2m, nighttime resolution 0.5m, sampling frequency 5 times per hour during the day and 8 times per hour at night, and the geographical range was 15km outside a 150km section of a certain border line.

[0066] S3-2: When prioritizing demands, a 150-kilometer section of a certain border line is close to a border crossing, and the area weight is... Calculate priority score Other concurrent needs include: increased patrols along an 80-kilometer section of the border, with E=2, f=2 times / hour, and k=0.7. Monitoring in a border no-man's-land: E=1, f=1 times / hour, k=0.3. ;according to The priority is given to the mobilization of armored forces and helicopters, with priority given to the deployment of drones and radar resources.

[0067] Step S4: Multi-source data scheduling invokes the reinforcement learning scheduling model. Input requirements and data source status: 4 visible light UAVs are available, 2 infrared UAVs are available, radar is available, the next remote sensing transit is at 14:00, and bandwidth remains at 80%. Scheduling plan: During the day, 1 visible light UAV is scheduled every 12 minutes, with 4 UAVs rotating to cover different angles of the assembly area; at night, 1 infrared UAV is scheduled every 7.5 minutes, with 2 UAVs rotating to improve the monitoring of covert activities; radar data is collected every 2 minutes, and the changes in the number and speed of targets are transmitted synchronously; remote sensing images are collected at 14:00 and 19:00 and then added to the scheduling data to verify changes in equipment scale.

[0068] Step S4, the decision support report generation uses the "Border Military Activities" template. The risk overview module includes the risk type "Abnormal Border Military Activities," location, equipment scale, activity status, and urgency level (E=4). The data supporting module inserts 10:15 infrared UAV imagery, 10:30 visible light UAV imagery, a radar target cluster change curve from 10:00 to 10:30, and 14:00 remote sensing imagery. The decision recommendations module generates: 1. Police Deployment: Add two patrol teams within a 150km section of the border, equipped with infrared night vision devices, increasing patrol frequency from 2 hours / time to 1 hour / time; 2. Enhanced Surveillance: Implement 24-hour rotating UAV surveillance, focusing on tracking the movement direction of armored vehicles and helicopters, and using radar to monitor target speed changes; 3. Warning Level: Set to "Yellow Warning," updating the risk assessment daily; 4. Intelligence Sharing: Synchronize data to adjacent border posts and the superior command center. The report is pushed to the border command center via an encrypted dedicated line, and a paper version is also generated and sent to the intelligence department.

[0069] Step S5: The feedback collection period for decision-making effectiveness is 48 hours. The collected data includes: data utilization rate. The report cites fused data, including 6 public opinion reports, 24 UAV images, 144 radar data points, and 2 remote sensing images. The S2 output total data includes 10 public opinion reports, 30 UAV images, 240 radar data points, and 3 remote sensing images. Calculations are then performed. The percentage was below the threshold of 80%, mainly due to low radar data utilization (144 / 240 = 60%); the rate of compliance with treatment effectiveness standards was also low. All four objectives—"police deployment," "enhanced surveillance," "early warning system setup," and "intelligence sharing"—were achieved, with no equipment crossing the border or exercises escalating. .

[0070] When optimizing based on feedback, S5 sends instructions to S1: The low radar data utilization rate is due to some data being of stationary targets. Adjust the radar acquisition frequency, reducing it from 2 minutes / time when the target is stationary to 5 minutes / time, and restoring it to 2 minutes / time when the target is moving; Add keywords "armored vehicle model" and "helicopter model" to public opinion collection to improve the matching degree between public opinion and data; S5 sends a command to S2 to update the feature matching algorithm. The coefficient, due to insufficient identification of "armored vehicle model", will be reduced to the keywords "tank" and "armored vehicle". The value was increased from 0.9 to 0.93. S5 sends a command to S3 to update the nighttime collection frequency of the "abnormal border military activities" entry in the mapping database, reducing it from 8 times / hour to 6 times / hour.

[0071] An embodiment of the intelligent decision support device for fusing border defense public opinion and multi-source data, as described in this invention, can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is part of the intelligent decision support device for fusing border defense public opinion and multi-source data according to the present invention. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of that data processing device, which will not be elaborated further. The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0072] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0073] This application also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the method or approach as described in the foregoing embodiments.

[0074] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0075] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0076] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for supporting intelligent decision-making by fusing border public opinion and multi-source data, characterized in that, It comprises the following steps: S1: Collecting the preset type of public opinion data related to border defense and the preset type of basic multi-source data, removing the public opinion data and multi-source data irrelevant to border defense by using a data filtering algorithm, and completing the initial screening of data; S2: Aligning the screened public opinion data and multi-source data in semantic and spatial dimensions, cleaning the aligned data to handle data conflicts, and obtaining a correlation fusion data set; S3: Determining the risk type through the correlation fusion data set, extracting the corresponding data resolution and collection frequency decision demand index, and prioritizing the demand index; S4: Scheduling multi-source data through the demand index and priority, integrating the scheduled data with the public opinion data obtained in S2, and generating a decision support report; S5: Collecting the application effect feedback of the decision support report, and optimizing the data collection focus and fusion parameters in the correlation fusion data set based on the feedback. 2.The border defense public opinion and multi-source data fusion intelligent decision support method according to claim 1, characterized in that, In S1, the data filtering algorithm is used to remove the public opinion data and multi-source data irrelevant to border defense, and complete the initial screening of data, wherein the data filtering algorithm includes a combined filtering algorithm, including two sub-processes of keyword filtering and feature filtering: the keyword filtering adopts a combination of Boolean filtering and weighted filtering, retains public opinion data containing a preset keyword combination, and retains public opinion data with a similarity to a preset theme reaching a preset threshold; the feature filtering retains multi-source data within a preset border defense geographic range based on geographic coordinates and data quality, and eliminates data with data quality not meeting a preset threshold. 3.The border defense public opinion and multi-source data fusion intelligent decision support method of claim 1, wherein, In S2, the screened public opinion data and multi-source data are aligned in semantic and spatial dimensions, including the following contents: S21: Bind the keywords of the screened public opinion data with the features of the multi-source data, and obtain the geographic coordinates of the multi-source data, associate the public opinion data with the multi-source data corresponding to the geographic coordinates, establish a public opinion and data feature dictionary, and the binding of the keywords and data features adopts a feature matching algorithm; S22: When new public opinion types and multi-source data types appear, perform incremental training of feature matching, adjust the weight parameters based on historical matching data through iterative optimization, include new public opinion keywords and new multi-source data features in the public opinion and data feature dictionary, and update the weight parameters in the feature matching algorithm.

4. The border defense public opinion and multi-source data fusion intelligent decision support method according to claim 1, characterized in that, The aligned data is cleaned to handle data conflicts, including the following contents: S23: Establish a conflict judgment standard, assign weights to the screened public opinion data and multi-source data according to the information source credibility, calculate the credibility weighted value when the features of the two are inconsistent, and if the weighted value is higher than the preset threshold, it is judged as no conflict, and if it is lower than the preset threshold, it is judged as conflict; If it is judged as conflict, trigger the data supplement process of the corresponding area, re-align the supplement data in semantic and spatial dimensions and handle data conflicts until a conflict-free correlation fusion data set is obtained.

5. The method of claim 1, wherein, The S3 determines a risk type through the associated fusion data set, extracts corresponding data resolution and collection frequency decision demand indicators, and includes the following contents: a mapping library of risk types and decision demands is established, the determined risk type is matched with the mapping library entries, data resolution and collection frequency indicators are extracted, and the indicators are quantified based on risk urgency; the risk urgency is divided into several levels according to a preset grading standard, and different indicator quantification results are corresponded. 6.The method of border defense public opinion and multi-source data fusion of wisdom decision support according to claim 1, characterized in that, The S3 prioritizes the demand indicators, including the following contents: a border defense area weight is preset according to the importance of the border defense area, a demand priority score is calculated, the demand indicators are sorted according to the score, and if there are multiple risk demands in the same area, the demand indicators are combined and the parameters with the highest resolution and the highest collection frequency are retained. 7.The method of claim 1, wherein, The S4 dispatches multi-source data through the demand indicators and priority, including the following contents: a scheduling model that meets the decision demand indicators and has the lowest resource consumption is included, the scheduling model determines a scheduling scheme in combination with the current state of the data source, data transmission bandwidth and device endurance; If the target data source is unavailable, an adjustment instruction is sent, the corresponding decision demand indicator is reduced, the priority score is recalculated, and then feedback is performed. 8.The method of border defense public opinion and multi-source data fusion of wisdom decision support according to claim 7, characterized in that, The training of the scheduling model includes the following contents: Based on historical data within a preset time, including data source state, demand indicator, and scheduling result data, the training data is updated daily within a preset time period, and the model scheduling success rate is kept stable at more than 90%; When performing scheduling, the model calls the demand indicators and priority in the S3, and preferentially schedules the data source corresponding to the highest priority demand, and if the data source is unavailable, a suboptimal data source is selected and the collection parameters are adjusted. 9.The method of claim 7, wherein, The decision support report in the S4 includes three modules of risk profile, data evidence, and decision suggestion, the profile module includes the determined risk type, risk location, influence range, and risk urgency; The data evidence module includes key data of the associated fusion data set; The decision suggestion module generates targeted suggestions based on the risk type and data evidence; After the report is generated, it is encrypted and pushed to the preset border defense command system through a preset encryption algorithm, including public opinion summary, multi-source data annotation graph, and data collection time; The decision suggestion module generates targeted suggestions based on the risk type and data evidence.

10. The method of claim 1, wherein, The S5 collects the application effect feedback of the decision support report, and uses data usage rate and treatment effect compliance rate as quantitative evaluation indicators; the data usage rate includes the proportion of the associated fusion data amount cited in the decision support report to the total data amount of the associated fusion data set, if the usage rate is lower than a preset threshold, an instruction is sent to adjust the data collection proportion; the treatment effect compliance rate is the proportion of the expected target achieved after the decision suggestion is implemented, if the compliance rate is lower than a preset threshold, an instruction is sent to update the weight parameters of the feature matching algorithm, and an instruction is sent to update the entry parameters of the risk and demand mapping library.

Images