A situation assessment model construction method and system for fusing multi-source heterogeneous data

Abstract

The application provides a situation assessment model construction method and system for fusing multi-source heterogeneous data, and belongs to the technical field of situation assessment model construction. The method comprises the following steps: generating a standardized feature vector through a multi-source heterogeneous situation data normalization processing unit; deploying a multi-modal situation element fusion hub to generate a basic situation feature matrix, building an initial situation reasoning prototype based on the matrix, collecting operation data, and generating an optimization scheme through a deep deterministic policy gradient optimization algorithm; forming a final model cluster through multi-dimensional consistency verification, encapsulating and constructing a collaborative communication link using a distributed architecture, and continuously optimizing the evaluation model through a full-cycle perception and dynamic iteration mechanism. The application solves the problem of difficult fusion of multi-source heterogeneous data, improves the utilization efficiency of situation elements, enhances the model evaluation accuracy and operation efficiency, ensures the model to adapt to data feature changes and scene migration, and effectively improves the accuracy, real-time performance and adaptability of situation assessment.

Description

Technical Field

[0001] This invention belongs to the field of situation assessment model construction technology, specifically a method and system for constructing a situation assessment model that integrates multi-source heterogeneous data. Background Technology

[0002] In various complex scenarios of modern society, situation assessment serves as a core support for accurate decision-making, risk prevention and control, and efficient management, and has been widely applied in many key areas such as transportation and network operation and maintenance. With the rapid iteration of technologies such as the Internet of Things, big data, and artificial intelligence, the data sources for situation awareness are showing a diversified and heterogeneous development trend. The massive amounts of data generated by various sensing devices differ significantly in format, dimension, and accuracy, posing a huge challenge to the effective integration and value mining of data.

[0003] Current situation assessment solutions generally have significant limitations: on the one hand, the integration of multi-source heterogeneous data lacks efficient and unified processing logic, the compatibility of data from different sources is poor, and problems such as data redundancy and conflicts are prone to occur, resulting in low data utilization and difficulty in fully leveraging the collaborative value of multi-source data; on the other hand, the adaptability and dynamic adjustment capabilities of assessment models are insufficient, making it difficult to adapt to the dynamic changes in scenarios, and it is difficult to balance assessment accuracy and operational efficiency, failing to meet the real-time and accuracy requirements of situation assessment in complex scenarios.

[0004] Furthermore, existing assessment systems lack a full-cycle optimization mechanism. After deployment, models struggle to dynamically iterate based on changes in data characteristics and scenario shifts, leading to performance degradation over long periods and further hindering the reliability of situation assessments. Therefore, overcoming existing technological bottlenecks to achieve efficient integration and in-depth utilization of multi-source heterogeneous data, and constructing a highly adaptable, accurate, and sustainably optimized situation assessment system, has become a crucial technical challenge in this field and a key driver for the intelligent and precise development of situation assessment technology. Summary of the Invention

[0005] This invention provides a method and system for constructing a situation assessment model that integrates multi-source heterogeneous data, in order to address the deficiencies in the existing technology.

[0006] On the one hand, the present invention provides a method for constructing a situation assessment model that integrates multi-source heterogeneous data, including:

[0007] The multi-source heterogeneous situation data normalization processing unit performs anomaly cleaning and heterogeneous feature alignment on the multi-source heterogeneous situation data collected from the target situation scene, generating a standardized feature vector.

[0008] Deploy a multimodal situational element fusion center, input standardized feature vectors into the multimodal situational element fusion center, and generate a basic situational feature matrix.

[0009] An initial situation inference prototype is built based on the basic situation feature matrix, and the entire process of model inference of the initial situation inference prototype is collected. The entire process of the operational data is input into the multimodal situation element fusion center, and a personalized model parameter tuning scheme is generated through a deep deterministic strategy gradient optimization algorithm.

[0010] Multi-dimensional consistency verification is carried out on the basic situation feature matrix and the personalized model parameter optimization scheme. Conflicting features are eliminated and missing related features are supplemented to form the final situation inference model cluster.

[0011] A distributed service-oriented architecture is adopted to containerize the final situation inference model cluster, construct a collaborative communication link within the final situation inference model cluster, and generate a situation assessment model that integrates multi-source heterogeneous data.

[0012] The situation assessment model, which integrates multi-source heterogeneous data, is continuously optimized through a full-cycle situation awareness and dynamic iteration mechanism, and is adapted in real time to the feature changes and situation scenario migrations of multi-source heterogeneous situation data.

[0013] According to the present invention, a method for constructing a situation assessment model integrating multi-source heterogeneous data is provided. The multi-source heterogeneous situation data includes environmental perception data, target attribute data, behavioral trajectory data, and interaction link data. Environmental perception data includes meteorological parameters, geographic information, and electromagnetic environment indicators. Target attribute data includes target type, state parameters, and identification. Behavioral trajectory data includes movement path, temporal characteristics, and interaction frequency. Interaction link data includes data transmission latency, protocol compatibility, and interface connectivity.

[0014] According to the present invention, a method for constructing a situation assessment model that integrates multi-source heterogeneous data includes the following process for generating standardized feature vectors:

[0015] Construct a situational data corpus with a labeling system. The situational data corpus includes industry-standard situational terms, scenario-specific data descriptions, and historical situational sample data. The labeling system includes situational keywords, attribute indicator words, and scenario identifier words.

[0016] The Min-Max normalization algorithm was used to normalize the original multi-source heterogeneous situational data to obtain normalized data.

[0017] Based on CNN combined with Transformer model, normalized data is feature encoded to extract global situation feature vector and capture temporal correlation information.

[0018] A situational domain label library is constructed, and situational scenarios are classified based on a heterogeneous feature alignment algorithm. The feature information of different structural data is analyzed by combining feature encoding results to form a preliminary feature list.

[0019] The initial feature list is deduplicated, completed, and validated for consistency to generate standardized feature vectors.

[0020] According to the present invention, a method for constructing a situation assessment model that integrates multi-source heterogeneous data includes the following process for deploying a multi-modal situation element fusion hub:

[0021] The core of the fusion strategy scheduling is constructed, which includes a situation element metadata management module, a data-situation element mapping rule base, a fusion scheduling strategy unit, and an algorithm execution engine.

[0022] The status quo element metadata management module stores the attribute information of preset status quo elements. The attribute information includes the element function description, data usage parameters, interface specifications, dependencies, and adaptation scenario tags.

[0023] Based on attribute information, dynamic association rules between multi-source heterogeneous situational data and situational elements are established in the data-situation element mapping rule base.

[0024] The fusion scheduling strategy unit has preset scheduling algorithms based on data priority, fusion degree, and scenario matching degree, which are used to automatically select the appropriate scheduling algorithm according to the situation and scenario.

[0025] The algorithm execution engine has pre-set heterogeneous feature alignment algorithm, deep deterministic strategy gradient optimization algorithm and feature similarity matching algorithm, and provides algorithm call interface and parameter configuration interface.

[0026] A real-time monitoring module is configured for the multimodal situational element fusion hub. The real-time monitoring module is used to collect hub operation status data in real time, including data processing rate, element matching accuracy and resource utilization rate.

[0027] The central operating status data is fed back to the algorithm execution engine for anomaly analysis, enabling anomaly alarms and adaptive optimization.

[0028] According to the present invention, a method for constructing a situation assessment model that integrates multi-source heterogeneous data includes the following process for generating a basic situation feature matrix:

[0029] The multimodal situational element fusion center receives standardized feature vectors and extracts core data features through a feature similarity matching algorithm.

[0030] Based on the core features of the data, the data-situation element mapping rule base is queried to select candidate situation elements that meet the requirements of situation awareness, satisfy the assessment accuracy constraints, and are suitable for the characteristics of the scenario.

[0031] Data compatibility verification is performed on candidate situation elements. Combined with the scheduling algorithm selected by the fusion scheduling strategy unit, the matching degree between the data requirements of candidate situation elements and the available resources of the system is analyzed, and elements that cannot be met by resources are eliminated.

[0032] The Softmax algorithm is used to calculate the weights of situation elements. Candidate situation elements that pass the verification are grouped according to situation correlation, and a basic situation feature matrix is ​​constructed, which includes element name, feature description, evaluation parameters, related elements, data requirements and adaptation scenarios.

[0033] The integrity of the basic situation feature matrix is ​​verified. If any elements are missing, the data processing stage is returned to supplement multi-source heterogeneous situation data information until the matrix meets the coverage requirements of the basic situation assessment.

[0034] According to the present invention, a method for constructing a situation assessment model that integrates multi-source heterogeneous data includes the following steps: Building an initial situation inference prototype and collecting data from the entire model inference process.

[0035] Based on the element dependencies in the basic situation feature matrix, an element call chain diagram is constructed, and situation assessment elements are deployed using model orchestration technology.

[0036] Based on the feature call chain diagram, configure the communication protocol and data serialization format between features, and deploy a real-time data acquisition gateway.

[0037] The initial situation inference prototype was built by starting the situation assessment elements and the real-time data acquisition gateway, and the situation scenario was simulated for trial operation to verify the feasibility of basic interactions between elements.

[0038] The real-time data acquisition gateway collects the entire process data of the model inference prototype according to preset rules, and stores it in a distributed database after cleaning and formatting. The entire process data of the model inference includes data processing accuracy, model evaluation latency, resource utilization, and situation identification coverage.

[0039] According to the present invention, a method for constructing a situation assessment model that integrates multi-source heterogeneous data includes the following process for generating a personalized model parameter tuning scheme through a deep deterministic policy gradient optimization algorithm:

[0040] Training data sets are constructed by extracting the entire process of model inference from a distributed database, and a reinforcement learning objective function is set with the goal of balancing model evaluation accuracy and resource utilization.

[0041] Define the state space, action space, and reward function for reinforcement learning. The state space is the feature vector of the data running throughout the entire model inference process. The action space includes operations such as feature splitting, feature merging, model replacement, parameter reallocation, and interface optimization and adjustment.

[0042] A decision model is constructed using a deep deterministic strategy gradient optimization algorithm, and the model is trained using a training data set to optimize the decision strategy.

[0043] The trained decision-making model is used to analyze the model state corresponding to the current basic situation feature matrix, and to identify existing evaluation bottlenecks, data conflicts and scenario adaptation defects.

[0044] Based on the identification results, a personalized model parameter optimization plan is generated, including the identifier of the element to be adjusted, the type of adjustment operation, the configuration of adjustment parameters, suggestions for synchronous adjustment of related elements, and the expected evaluation indicators after optimization.

[0045] According to the present invention, a method for constructing a situation assessment model that integrates multi-source heterogeneous data includes the following process for forming a final situation inference model cluster:

[0046] Construct a multi-dimensional consistency verification index system, including element consistency index, evaluation matching index, association compatibility index, resource balance index and scenario adaptability index, and set the qualified threshold for each index.

[0047] Based on the element consistency index, the consistency of model elements after the personalized model parameter tuning scheme is adjusted is verified.

[0048] Based on the evaluation matching index, the evaluation parameters of the model in the personalized model parameter tuning scheme are compared with the situational requirement index.

[0049] Based on the correlation compatibility index, the correlation between the optimized elements is analyzed.

[0050] Based on the resource balance index, the balance of resource utilization distribution of each element after optimization is evaluated.

[0051] Based on the scenario adaptability index, the cross-scenario adaptability of the optimized model set is verified.

[0052] Calculate the multi-source data fusion degree index, and perform a final verification on the model set that has passed the multi-dimensional consistency check based on the fusion degree index to form the final situation inference model cluster.

[0053] The present invention provides a method for constructing a situation assessment model that integrates multi-source heterogeneous data. The process of continuous optimization through a full-cycle situation awareness and dynamic iteration mechanism includes:

[0054] Construct a real-time monitoring system covering model instances and element link levels, and collect operational indicators and data adaptation indicators of the situation assessment model that integrate multi-source heterogeneous data.

[0055] Deploy an adaptive iterative engine to connect with the real-time monitoring system and the situational requirements management platform to identify abnormal indicator data and explicit and implicit scenario changes.

[0056] Based on the anomaly identification results and scene change information, parameter allocation and model adaptation adjustment strategies are generated by combining the historical optimization case library, and then implemented after feasibility verification.

[0057] Verify the optimization effect. If it does not meet expectations, re-optimize the strategy and save the optimization process, strategy and effect data into the historical optimization case library.

[0058] On the other hand, the present invention also provides a situation assessment model construction system that integrates multi-source heterogeneous data, comprising:

[0059] The data processing unit is used to perform anomaly cleaning and heterogeneous feature alignment operations on multi-source heterogeneous situational data to generate standardized feature vectors.

[0060] The fusion scheduling unit is used to deploy a multimodal situational element fusion hub to achieve the matching and scheduling of situational elements;

[0061] The model building unit is used to build an initial situational reasoning prototype and generate personalized model parameter tuning schemes.

[0062] The verification and optimization unit is used to perform multi-dimensional consistency verification and continuous optimization of the model;

[0063] The deployment and execution unit is used to implement containerized encapsulation and deployment of the model.

[0064] This invention provides a method and system for constructing a situation assessment model that integrates multi-source heterogeneous data. It standardizes situation data from different structures and sources through a multi-source heterogeneous situation data normalization processing unit, solving the problem of difficulty in integrating multi-source heterogeneous data. It automatically matches and dynamically schedules situation elements through a multi-modal situation element fusion hub, improving the utilization efficiency of situation elements. It automatically optimizes model parameters through a deep deterministic strategy gradient optimization algorithm, improving the model's assessment accuracy and operational efficiency. It continuously optimizes the model through a full-cycle situation awareness and dynamic iteration mechanism, ensuring the model can adapt to changes in data characteristics and scene migrations in real time. This method effectively improves the accuracy, real-time performance, and adaptability of situation assessment, and is applicable to various situation assessment scenarios such as battlefield situation, traffic situation, and network situation. Attached Figure Description

[0065] The invention will now be further described with reference to the accompanying drawings.

[0066] Figure 1 This is a flowchart illustrating a method for constructing a situation assessment model that integrates multi-source heterogeneous data in this invention.

[0067] Figure 2 This is a schematic diagram of the structure of a situation assessment model construction system that integrates multi-source heterogeneous data in this invention. Detailed Implementation

[0068] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0069] like Figures 1 to 2 As shown in the embodiment of the present invention, a method and system for constructing a situation assessment model that integrates multi-source heterogeneous data are provided. The executing entity can be a method for constructing a situation assessment model that integrates multi-source heterogeneous data, the method including:

[0070] The multi-source heterogeneous situation data normalization processing unit performs anomaly cleaning and heterogeneous feature alignment on the multi-source heterogeneous situation data collected from the target situation scene, generating a standardized feature vector.

[0071] Multi-source heterogeneous situational data is collected synchronously by multiple types of sensing devices deployed in the target situational scenario. Among them, environmental sensing data is collected by meteorological sensors, geographic information acquisition terminals, and electromagnetic spectrum monitoring devices respectively, with the acquisition frequency set to once per second.

[0072] Target attribute data is jointly acquired by radar detection equipment, photoelectric recognition equipment, and an identity authentication system. The radar detection equipment has a scanning cycle of 0.5 seconds, and the photoelectric recognition equipment has a frame rate of 30 frames per second. Behavioral trajectory data is collected by a GPS positioning module, an inertial navigation device, and a data interaction recording system. The positioning data is updated twice per second.

[0073] Interactive link data is collected by network traffic monitoring devices, protocol parsing gateways, and interface status monitoring programs, with a data sampling interval of 100 milliseconds.

[0074] All collected raw data are first filtered through edge computing nodes to remove invalid data that is clearly beyond the physical reasonable range, and then transmitted to the multi-source heterogeneous situational data normalization processing unit.

[0075] The anomaly cleaning process is implemented using the 3σ criterion combined with the Isolation Forest algorithm. First, the mean and standard deviation of the numerical data are calculated. Data that exceed the mean ± 3 times the standard deviation are marked as preliminary outliers. Then, the Isolation Forest algorithm is used to score all data for anomalies. When the anomaly score of a data is greater than 0.7, it is determined to be a final outlier. Outliers are replaced by linear interpolation of data from adjacent time points.

[0076] The heterogeneous feature alignment process adopts differentiated processing methods for data with different structures. For structured data, field values ​​are directly extracted and converted into floating-point numbers. For unstructured text data, word segmentation and stop word removal are performed first, and then the data is converted into a 128-dimensional feature vector through a word embedding model. For semi-structured log data, key fields are parsed first, and then timestamps, event types, and parameter values ​​are converted into corresponding numerical features. Finally, all data features with different structures are uniformly mapped into a 256-dimensional feature space.

[0077] The process of generating standardized feature vectors includes:

[0078] Construct a situational data corpus with a labeling system. The situational data corpus includes industry-standard situational terms, scenario-specific data descriptions, and historical situational sample data. The labeling system includes situational keywords, attribute indicator words, and scenario identifier words.

[0079] The situational data corpus is built incrementally. The initial corpus contains 100,000 annotated historical situational data entries, and the corpus is updated every time 1,000 new situational data entries are added.

[0080] In the labeling system, situation keywords are used to identify the core elements of situational events, including target type, event type, and location. Attribute indicator terms are used to quantify the characteristics of situational elements, including speed, distance, intensity, and frequency. Scene identifier terms are used to distinguish different application scenarios, including battlefield situation, traffic situation, and network situation.

[0081] The Min-Max normalization algorithm is used to normalize the original multi-source heterogeneous situational data, resulting in normalized data. The normalization formula is as follows:

[0082]

[0083] in, For normalized situational data, This is the original multi-source heterogeneous situational data. For the minimum value of the corresponding data sample, This represents the maximum value of the corresponding data sample. During the normalization process, the minimum and maximum values ​​of each data category are calculated separately. The sample set is selected from data of the same type collected within the last 30 days. When the data type changes, the corresponding minimum and maximum values ​​are automatically recalculated.

[0084] This model combines a CNN and a Transformer to encode features from normalized data, extracting global situational feature vectors and capturing temporal correlation information. The CNN part employs a three-layer convolutional neural network with kernel sizes of 3, 5, and 7, each with a stride of 1. Each convolutional layer is followed by a max-pooling layer with a pooling window size of 2 and a stride of 2, used to extract local spatial features. The Transformer part uses a six-layer encoder structure, with each encoder containing a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer has 8 attention heads, and the feedforward neural network layer has a hidden layer dimension of 512, used to capture long-range temporal correlation information. The model has an input dimension of 256 and an output dimension of 128. It is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training iterations. Training is stopped early when the validation set loss no longer decreases after 10 consecutive iterations.

[0085] A situational domain label library is constructed, and situational scenarios are classified based on a heterogeneous feature alignment algorithm. The feature information of different structural data is analyzed by combining feature encoding results to form a preliminary feature list.

[0086] The situation domain tag library adopts a hierarchical structure, divided into three levels: first-level tags, second-level tags, and third-level tags. First-level tags include four categories: environmental situation, target situation, behavioral situation, and interaction situation. Second-level tags are further subdivided based on first-level tags. For example, under target situation, it is divided into three categories: air targets, ground targets, and maritime targets. Third-level tags correspond to specific situation elements. For example, under air targets, it is divided into three categories: fighter jets, helicopters, and drones.

[0087] The heterogeneous feature alignment algorithm maps data features to corresponding label categories by calculating the cosine similarity between features of different data structures and label features in the label library. A successful match is determined when the cosine similarity is greater than 0.8.

[0088] The initial feature list undergoes deduplication, completion, and consistency checks to generate standardized feature vectors. Deduplication is achieved by calculating the cosine similarity between feature vectors; if the cosine similarity between two feature vectors is greater than 0.95, they are considered duplicate features, one is retained, and the other is deleted. Completion involves using the average of other similar features from the same scene at the same time for features with a missing rate below 30%, while features with a missing rate above 30% are directly discarded. Consistency checks verify that feature values ​​conform to the range of their corresponding labels; for example, the range of target velocity features cannot exceed physical limits. Feature values ​​that do not conform to the range are corrected or discarded. The final standardized feature vector is a 128-dimensional floating-point vector, with each dimension corresponding to a specific situational feature.

[0089] A multimodal situational awareness fusion center is deployed, and standardized feature vectors are input into the center to generate a basic situational awareness feature matrix. The multimodal situational awareness fusion center is deployed on a cloud server cluster, employing a distributed architecture design that supports horizontal scaling and can automatically adjust the number of server nodes according to the amount of data processed.

[0090] The process of deploying a multimodal situational awareness fusion hub includes building a fusion strategy scheduling core, which comprises a situational awareness metadata management module, a data-situational awareness mapping rule base, a fusion scheduling strategy unit, and an algorithm execution engine. The situational awareness metadata management module uses a relational database to store the attribute information of preset situational awareness elements. The database table structure includes fields such as element ID, element name, function description, data usage parameters, interface specifications, dependencies, and adaptation scenario tags. Data usage parameters include CPU utilization, memory usage, and disk I / O speed, while interface specifications include interface address, request method, parameter format, and return value format.

[0091] Based on attribute information, dynamic association rules are established between multi-source heterogeneous situational data and situational elements in the data-situational element mapping rule base. These dynamic association rules are generated using a decision tree algorithm, taking the features of the multi-source heterogeneous situational data as input and the corresponding situational elements as output. The decision tree model is updated every 1000 new data-element matching samples. The matching weights of the association rules are dynamically adjusted based on the reliability and timeliness of the data. Data reliability is determined by the accuracy of the acquisition equipment and its historical accuracy, while timeliness is determined by the difference between the data acquisition time and the current time; a smaller difference indicates higher timeliness and a larger weight.

[0092] The fusion scheduling strategy unit pre-sets three scheduling algorithms: one based on data priority, one based on fusion degree, and one based on scenario matching degree. These algorithms are used to automatically select the appropriate scheduling algorithm based on the situation and scenario. When a high-priority target exists in the situation, the data priority-based scheduling algorithm is automatically selected to prioritize processing the data source corresponding to the high-priority target. When it is necessary to improve the fusion effect of multi-source data, the fusion degree-based scheduling algorithm is automatically selected to prioritize processing data source data with a high fusion degree. When the situation and scenario change, the scenario matching degree-based scheduling algorithm is automatically selected to prioritize processing data source data with a high matching degree to the current scenario. The switching time between scheduling algorithms does not exceed 1 second, ensuring the real-time performance of data processing.

[0093] The algorithm execution engine includes pre-defined heterogeneous feature alignment, deep deterministic gradient optimization, and feature similarity matching algorithms, providing both an algorithm call interface and a parameter configuration interface. The algorithm call interface is designed in a RESTful style and supports HTTP and HTTPS protocols. The parameter configuration interface is a web page, allowing users to modify algorithm parameters online; changes take effect immediately without requiring an engine restart. The algorithm execution engine employs a multi-threaded concurrent processing mode, capable of handling multiple algorithm tasks simultaneously. Task scheduling is implemented using a priority queue, with higher-priority tasks executed first.

[0094] A real-time monitoring module is configured for the multimodal situational awareness fusion hub. This module collects hub operational status data in real time, including data processing rate, feature matching accuracy, and resource utilization. The real-time monitoring module collects data once per second, and the collected data is stored in a time-series database for 90 days. The hub's operational status data is fed back to the algorithm execution engine for anomaly analysis, enabling anomaly alerts and adaptive optimization. Anomaly analysis uses a threshold-based method. When the data processing rate is below 1000 records / second, the feature matching accuracy is below 90%, CPU utilization is above 80%, or memory utilization is above 85%, an anomaly alert mechanism is triggered, sending alert information to maintenance personnel via SMS, email, and system pop-ups. The adaptive optimization process automatically adjusts algorithm parameters and resource allocation based on the hub's operational status data. When the data processing rate is too low, the number of threads in the algorithm execution engine is increased. When resource utilization is too high, the resource quota for low-priority tasks is reduced.

[0095] The process of generating the basic situation feature matrix includes the multimodal situation element fusion center receiving standardized feature vectors and extracting core features of the data through a feature similarity matching algorithm. The feature similarity matching algorithm uses cosine similarity to calculate the similarity between the standardized feature vectors and the preset core feature template. The preset core feature template contains core features of 100 common situation scenarios, and features with a similarity greater than 0.7 are extracted as core features of the data.

[0096] Based on the core characteristics of the data, a data-situation element mapping rule base is queried to select candidate situation elements that meet situational awareness requirements, satisfy assessment accuracy constraints, and are suitable for specific scenarios. Situational awareness requirements include the functional coverage of the elements; assessment accuracy constraints include the error range and response time of the elements; and scenario suitability features include the types of scenarios and environmental conditions to which the elements are applicable. The selection process first identifies situational elements suitable for the current scenario based on scenario characteristics, then selects elements that meet the accuracy requirements based on assessment accuracy constraints, and finally selects elements that cover all core functions based on situational awareness requirements.

[0097] Candidate situation elements undergo data compatibility verification. Combined with the scheduling algorithm selected by the fusion scheduling strategy unit, the matching degree between the data requirements of the candidate situation elements and the available system resources is analyzed, and elements whose resource requirements cannot be met are eliminated. Data requirements include the CPU, memory, disk, and network resources required for element operation. Available system resources are obtained in real time by the resource monitoring system of the cloud server cluster. When the resource requirements of an element exceed 80% of the available system resources, it is determined that the resources cannot be met, and the element is eliminated.

[0098] The Softmax algorithm is used to calculate the weights of situational features. The formula for the feature weights is as follows:

[0099]

[0100] in, For the first The weights of each situational element, For the first Feature scores of each element This represents the total number of situational elements. The feature score of an element is determined by its functional importance, data reliability, and assessment accuracy. Functional importance is determined by domain experts, with scores ranging from 1 to 10. Data reliability is determined by the historical accuracy of the data source corresponding to the element. Assessment accuracy is determined by the error range of the element; the smaller the error, the higher the accuracy and the higher the score.

[0101] Candidate situation elements that pass the verification are grouped according to situational correlation, and a basic situational feature matrix is ​​constructed, including element name, feature description, evaluation parameters, related elements, data requirements, and adaptation scenarios. Situational correlation is determined by calculating the mutual information value between elements; elements with a mutual information value greater than 0.5 are grouped together. The basic situational feature matrix is ​​stored in the form of a two-dimensional array, with rows corresponding to different situational element groups and columns corresponding to element attribute information. The matrix has a dimension of [missing information]. ,in This represents the number of situation element groups.

[0102] The basic situation feature matrix undergoes integrity verification. If any elements are missing, the process returns to the data processing stage to supplement multi-source heterogeneous situation data until the matrix meets the coverage requirements for basic situation assessment. Integrity verification is achieved by checking whether the matrix covers all core situation functions, which include four categories: target detection, status assessment, trend prediction, and risk warning. Each function requires at least two corresponding situation elements. When a function is missing, a data supplementation request is automatically sent to the data acquisition system, specifying the data source type and data range to be supplemented. After supplementation, element selection and matrix reconstruction are performed again.

[0103] An initial situation inference prototype is built based on the basic situation feature matrix, and the entire process of model inference of the initial situation inference prototype is collected. The entire process of the operational data is input into the multimodal situation element fusion center, and a personalized model parameter tuning scheme is generated through a deep deterministic strategy gradient optimization algorithm.

[0104] The process of building the initial situation inference prototype and collecting data on the entire model inference process includes constructing an element call chain graph based on the element dependencies in the basic situation feature matrix, and deploying situation assessment elements using model orchestration technology. The element call chain graph is represented as a directed acyclic graph (DAG), where nodes correspond to situation elements and edges correspond to the call relationships between elements. The graph construction process first determines the root node, i.e., the element without any prerequisite dependencies, and then adds subsequent nodes sequentially according to the dependencies, ensuring that there are no circular dependencies in the graph. The model orchestration technology is implemented using Kubernetes, encapsulating each situation element as an independent Pod, and managing the deployment and lifecycle of the Pods through a Deployment controller.

[0105] Based on the feature call chain diagram, configure the communication protocol and data serialization format between features, and deploy a real-time data acquisition gateway. The communication protocol between features uses gRPC, and the data serialization format uses Protocol Buffers. Compared to traditional HTTP and JSON protocols, gRPC and Protocol Buffers offer higher transmission efficiency and lower latency. The real-time data acquisition gateway is deployed on the edge nodes of the Kubernetes cluster, using Envoy as the data broker, responsible for collecting interaction data and operational status data between all features.

[0106] An initial situational reasoning prototype was built by integrating situational assessment elements and a real-time data acquisition gateway. Simulated situational scenarios were tested to verify the feasibility of basic interactions between elements. These scenarios were simulated by replaying historical situational data, selecting 10 typical scenarios, each with a 30-minute replay duration at 1x speed. The verification criteria for basic interaction feasibility included an interface call success rate of no less than 99% between elements, data transmission latency of no more than 100 milliseconds, and no data loss or out-of-order delivery.

[0107] The real-time data acquisition gateway collects the entire inference process data of the initial situation inference prototype model according to preset rules. This data includes data processing accuracy, model evaluation latency, resource utilization, and situation identification coverage. Data is collected once per second. The collected data is first cleaned and formatted to remove invalid and duplicate data, converted to a unified JSON format, and then stored in a distributed database. The distributed database uses ClickHouse, which supports high-concurrency writes and fast queries, meeting the storage and analysis needs of massive amounts of operational data.

[0108] The process of generating personalized model parameter tuning schemes through deep deterministic policy gradient optimization algorithms includes extracting the entire process of model inference from a distributed database to construct a training data set, and setting a reinforcement learning objective function with the goals of model evaluation accuracy and resource utilization.

[0109] The training data set is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model training, the validation set is used to tune model parameters, and the test set is used to evaluate model performance. The reinforcement learning objective function is... ,in The objective function value, and These are the weighting coefficients, and , To improve the accuracy of the model evaluation, To optimize resource utilization, the weighting coefficients are adjusted based on actual business needs, prioritizing assessment accuracy. Take 0.7, Take 0.3. When prioritizing resource utilization, Take 0.3, Take 0.7.

[0110] Define the state space, action space, and reward function for reinforcement learning. The state space is a feature vector of the data running throughout the model inference process, with a dimension of 16, including data processing accuracy, model evaluation latency, CPU utilization, memory utilization, disk I / O rate, network bandwidth, interface call success rate, data transmission latency, situational awareness coverage, target detection accuracy, state evaluation error, trend prediction accuracy, risk warning accuracy, element response time, data throughput, and system stability.

[0111] The action space includes element splitting, element merging, model replacement, parameter reallocation, and interface optimization adjustment operations. Element splitting is to split an element containing multiple sub-functions into multiple independent sub-elements. Element merging is to merge multiple closely related sub-elements into a comprehensive element. Model replacement is to replace the currently used model with a better performing model. Parameter reallocation is to adjust the resource quotas and algorithm parameters of each element. Interface optimization adjustment is to optimize the interface design and data transmission methods between elements.

[0112] Reinforcement learning reward function formula:

[0113]

[0114] in, For the overall reward value, Normalized weight coefficients and satisfying , To improve the accuracy of situation assessment, For model evaluation latency, This represents the system resource consumption value. The weighting coefficient is set to... , , The accuracy of situation assessment ranges from 0 to 1, the model assessment latency is in seconds, and the system resource consumption is a weighted average of CPU utilization, memory utilization, and network bandwidth utilization.

[0115] A decision model is constructed using a deep deterministic strategy gradient optimization algorithm, and the model is trained using a training data set to optimize the decision strategy.

[0116] The decision model consists of an actor network and a critic network. The actor network outputs the action policy, and the critic network evaluates the value of the action policy. Both the actor and critic networks use 3-layer fully connected neural networks with a hidden layer dimension of 256 and the ReLU activation function. The model's learning rate is set to 0.0001, the discount factor to 0.99, the soft update coefficient to 0.001, the experience replay buffer size to 100,000, and the batch size to 64. During training, the model is evaluated every 1000 steps. The model is considered converged when the average reward value on the test set no longer increases after 10 consecutive evaluations.

[0117] The trained decision-making model is used to analyze the model state corresponding to the current basic situation feature matrix, identifying evaluation bottlenecks, data conflicts, and scenario adaptation defects. Evaluation bottlenecks are defined as model evaluation latency exceeding 2 seconds or data processing rate below 500 data points / second. Data conflicts are defined as the difference in the same situation parameter output by different elements exceeding 10%. Scenario adaptation defects are defined as situation recognition coverage below 80% or target detection accuracy below 90%.

[0118] Based on the identification results, a personalized model parameter tuning plan is generated, including the identifier of the element to be adjusted, the type of adjustment operation, the configuration of adjustment parameters, suggestions for synchronous adjustment of related elements, and the expected evaluation indicators after optimization. The identifier of the element to be adjusted is the unique ID of the element; the type of adjustment operation corresponds to the operation in the action space; the configuration of adjustment parameters includes resource quotas, algorithm parameters, interface parameters, etc.; the suggestions for synchronous adjustment of related elements list the related elements that need to be adjusted synchronously and the adjustment methods; and the expected evaluation indicators after optimization include model evaluation accuracy, evaluation latency, resource utilization, etc.

[0119] A multi-dimensional consistency verification was conducted on the basic situation feature matrix and the personalized model parameter optimization scheme. Conflicting features were eliminated, and missing related features were supplemented to form the final situation inference model cluster. The process included constructing a multi-dimensional consistency verification index system, including element consistency index, evaluation matching degree index, correlation compatibility index, resource balance index, and scenario adaptability index, and setting qualification thresholds for each index. The qualification thresholds for element consistency index and evaluation matching degree index are 95%, 90%, 100%, 0.2, and 85% respectively.

[0120] Based on the element consistency index, the consistency of model elements after the personalized model parameter tuning scheme is verified. The element consistency index is achieved by calculating the overlap of element functions before and after adjustment; the higher the overlap, the better the consistency. Based on the evaluation matching index, the evaluation parameters of the model in the personalized model parameter tuning scheme are compared with the situational requirement indicators. The evaluation matching index is achieved by calculating the ratio of the difference between the evaluation parameters and the requirement indicators to the requirement indicators; the smaller the ratio, the higher the matching degree. Based on the association compatibility index, the association relationships between optimized elements are analyzed to check for circular dependencies, missing dependencies, and interface incompatibility issues, ensuring that the association relationships between all elements are correct. Based on the resource balance index, the resource utilization distribution of each element after optimization is evaluated to avoid situations where some elements have excessive resources while others have insufficient resources. Based on the scenario adaptability index, the cross-scenario adaptability of the optimized model set is verified. The model performance is tested by simulating different situational scenarios to ensure that the model can achieve the expected evaluation results in various scenarios.

[0121] The formula for calculating the multi-source data fusion degree index is as follows:

[0122]

[0123] in, To improve the fusion efficiency of multi-source data, For the first The weight of each data source, For the first The similarity of features between data sources and situational elements. This represents the total number of data sources. The weight of each data source is determined based on its reliability and importance, and feature similarity is calculated using cosine similarity. The multi-source data fusion degree ranges from 0 to 1; a higher fusion degree indicates better utilization of the multi-source data.

[0124] The model set that passed the multi-dimensional consistency check was then subjected to a final verification based on the fusion degree index, forming the final situation inference model cluster. The final verification standard is that the multi-source data fusion degree is greater than 0.8, and all multi-dimensional consistency check indicators reach the qualified threshold. The final situation inference model cluster contains multiple interrelated situation inference models, each corresponding to a specific situation assessment task. The models interact with each other and fuse results through a collaborative communication link.

[0125] A distributed service-oriented architecture is adopted to containerize the final situation inference model cluster, constructing collaborative communication links within the final situation inference model cluster to generate a situation assessment model that integrates multi-source heterogeneous data. The distributed service-oriented architecture consists of four layers: a data access layer, a model inference layer, a result fusion layer, and an application presentation layer. The data access layer is responsible for receiving multi-source heterogeneous situation data; the model inference layer is responsible for running the situation inference model; the result fusion layer is responsible for integrating the inference results of multiple models; and the application presentation layer is responsible for displaying the situation assessment results to the user. Each layer adopts a microservice architecture design, and services communicate with each other via RESTful APIs.

[0126] Containerization is implemented using Docker technology. An independent Docker image is built for each situational inference model, containing the operating system, dependency libraries, model files, and configuration files required for model operation. The Docker images are built in layers: the base image is Ubuntu 20.04, the upper layers sequentially install the Python environment, deep learning framework, and model dependency libraries, and finally add the model files and configuration files. Once the images are built, they are pushed to a private image repository for use by the container orchestration platform.

[0127] The container orchestration platform uses Kubernetes to automate the deployment, elastic scaling, and fault healing of containers for the situational reasoning model. The deployment strategy employs a rolling update approach to ensure uninterrupted service during updates. Elastic scaling rules automatically adjust the number of container instances based on CPU utilization and request volume. When CPU utilization exceeds 70% or request volume exceeds 80% of the current instance's processing capacity, the platform automatically scales up by adding container instances. When CPU utilization is below 30% and request volume is below 50% of the current instance's processing capacity, the platform automatically scales down by reducing the number of container instances. The health check mechanism uses HTTP GET requests, sending a health check request to the container every 10 seconds. If three consecutive requests fail, the container is considered faulty, and it is automatically restarted or replaced.

[0128] The collaborative communication link is implemented using service mesh technology. Each model container is attached with an independent proxy component via the Sidecar proxy, responsible for handling inter-model communication, traffic control, load balancing, and security authentication. The service mesh is implemented using Istio, supporting dynamic routing, circuit breaking, rate limiting, and other functions, and can automatically adjust traffic allocation based on the model's running status. Communication is encrypted using the TLS 1.3 protocol to ensure data transmission security.

[0129] After generating a situation assessment model that integrates heterogeneous data from multiple sources, comprehensive integration testing is conducted. This testing includes functional testing, performance testing, compatibility testing, and security testing. Functional testing verifies whether the model can correctly complete all situation assessment tasks; performance testing verifies whether the model's assessment latency, throughput, and resource utilization meet the requirements; compatibility testing verifies whether the model can operate normally in different hardware and software environments; and security testing verifies whether the model has any security vulnerabilities or data leakage risks. Only after all tests are passed can the model be officially deployed.

[0130] The situation assessment model, which integrates multi-source heterogeneous data, is continuously optimized through a full-cycle situational awareness and dynamic iteration mechanism. This allows for real-time adaptation to changes in the characteristics of multi-source heterogeneous situational data and shifts in situational scenarios. The process includes building a real-time monitoring system covering model instances and element link levels, and collecting operational and data adaptation metrics of the situation assessment model. Operational metrics are quantitative indicators of the operational status of the situation assessment model itself, including model evaluation latency, resource utilization, data processing rate, situational identification accuracy, and model operational stability. Data adaptation metrics are quantitative indicators of the degree of adaptation and matching between the model and the multi-source heterogeneous situational data, including multi-source data fusion degree, heterogeneous data feature matching rate, data cleaning accuracy, data transmission adaptation rate, and cross-data source adaptation stability. The real-time monitoring system is implemented using Prometheus + Grafana. Prometheus is responsible for collecting and storing monitoring data, while Grafana is responsible for visualizing the monitoring data. The monitoring data is collected once per second and retained for 180 days.

[0131] An adaptive iterative engine is deployed, connecting to the real-time monitoring system and the situational awareness management platform to identify anomalies in indicator data and explicit and implicit scenario changes. The adaptive iterative engine employs a rule-based engine combined with machine learning algorithms. The rule-based engine identifies known anomalies in indicator data and explicit scenario changes, while the machine learning algorithms identify unknown anomalies and implicit scenario changes. Explicit scenario changes are detected through scenario change requests submitted by users via the situational awareness management platform. Implicit scenario changes are detected by analyzing the distribution changes of multi-source heterogeneous situational awareness data; when the KL divergence of the data distribution is greater than 0.5, it is determined that an implicit scenario change has occurred.

[0132] Based on anomaly identification results and scenario change information, parameter allocation and model adaptation adjustment strategies are generated using a historical optimization case library. These strategies are then executed after feasibility verification. The historical optimization case library stores relevant information for all historical optimization processes, including optimization background, anomaly type, scenario change details, adjustment strategies, and optimization effects. When generating adjustment strategies, similar cases are first retrieved from the historical optimization case library. Based on the adjustment strategies of these similar cases and the current system state, a preliminary adjustment strategy is generated. This preliminary adjustment strategy then undergoes feasibility verification, including technical feasibility, resource feasibility, and business impact. Execution only occurs after successful verification.

[0133] Verify the optimization effect. If it does not meet expectations, re-optimize the strategy and store the optimization process, strategy, and effect data in the historical optimization case library. Optimization effectiveness is verified by comparing operational metrics and data adaptation metrics before and after optimization. Optimization is considered successful when the improvement in all metrics reaches the expected target. If it does not meet expectations, re-analyze the cause of the problem, adjust the optimization strategy, and execute the optimization process again until the expected target is achieved. After each optimization, all data from the optimization process is stored in the historical optimization case library for use in generating subsequent optimization strategies.

[0134] In summary, this embodiment provides a method for constructing a situation assessment model that integrates multi-source heterogeneous data. It achieves standardized processing of situation data with different structures and sources through a multi-source heterogeneous situation data normalization processing unit, solving the problem of difficulty in integrating multi-source heterogeneous data. It improves the utilization efficiency of situation elements by automatically matching and dynamically scheduling situation elements through a multi-modal situation element fusion hub. It enhances the model's assessment accuracy and operational efficiency by automatically tuning model parameters through a deep deterministic strategy gradient optimization algorithm. Finally, it ensures continuous model optimization through a full-cycle situation awareness and dynamic iteration mechanism, guaranteeing that the model can adapt to changes in data characteristics and scene migrations in real time. This method effectively improves the accuracy, real-time performance, and adaptability of situation assessment and is applicable to various situation assessment scenarios such as battlefield situation, traffic situation, and network situation.

[0135] Based on the same general inventive concept, this invention also protects a situation assessment model construction system that integrates multi-source heterogeneous data. The situation assessment model construction system that integrates multi-source heterogeneous data provided by this invention will be described below. The situation assessment model construction system that integrates multi-source heterogeneous data described below can be referred to in correspondence with the situation assessment model construction method and system that integrates multi-source heterogeneous data described above.

[0136] A situation assessment model construction system that integrates multi-source heterogeneous data includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The processor includes a data processing unit, a fusion scheduling unit, a model construction unit, a verification and optimization unit, and a deployment and operation unit.

[0137] The data processing unit is used to perform anomaly cleaning and heterogeneous feature alignment operations on multi-source heterogeneous situational data to generate standardized feature vectors.

[0138] The fusion scheduling unit is used to deploy a multimodal situational element fusion hub to achieve the matching and scheduling of situational elements.

[0139] The model building unit is used to build an initial situational reasoning prototype and generate personalized model parameter tuning schemes.

[0140] The verification and optimization unit is used to perform multi-dimensional consistency verification and continuous optimization of the model.

[0141] The deployment and execution unit is used to implement containerized encapsulation and deployment of the model.

[0142] The memory is used to store computer programs, situation data, model parameters, historical cases, and other data, while the processor is used to execute the computer programs to implement the various steps of the situation assessment model construction method described above.

[0143] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a situation assessment model that integrates multi-source heterogeneous data, characterized in that, include: Through the multi-source heterogeneous situational data normalization processing unit, anomaly cleaning and heterogeneous feature alignment are performed on the multi-source heterogeneous situational data collected from the target situational scene to generate standardized feature vectors. Deploy a multimodal situational element fusion center, input the standardized feature vector into the multimodal situational element fusion center, and generate a basic situational feature matrix; Based on the aforementioned basic situation feature matrix, an initial situation inference prototype is built, and the entire process operation data of the initial situation inference prototype is collected. The entire process operation data is input into the multimodal situation element fusion center, and a personalized model parameter tuning scheme is generated through a deep deterministic strategy gradient optimization algorithm. A multi-dimensional consistency check is performed on the basic situation feature matrix and the personalized model parameter optimization scheme to eliminate conflicting features and supplement missing related features, forming the final situation inference model cluster. A distributed service architecture is adopted to implement the containerized encapsulation of the final situation inference model cluster, construct the collaborative communication link within the final situation inference model cluster, and generate a situation assessment model that integrates multi-source heterogeneous data. The situation assessment model, which integrates multi-source heterogeneous data, is continuously optimized through a full-cycle situation awareness and dynamic iteration mechanism, and is adapted in real time to the feature changes and situation scenario migrations of multi-source heterogeneous situation data.

2. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 1, characterized in that, The multi-source heterogeneous situational data includes environmental perception data, target attribute data, behavioral trajectory data, and interaction link data; the environmental perception data includes meteorological parameters, geographic information, and electromagnetic environment indicators; the target attribute data includes target type, state parameters, and identity identifier; the behavioral trajectory data includes movement path, temporal characteristics, and interaction frequency; and the interaction link data includes data transmission latency, protocol compatibility rate, and interface connectivity.

3. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 1, characterized in that, The process of generating standardized feature vectors includes: Construct a situational data corpus with a labeling system. The situational data corpus includes industry-standard situational terms, scenario-specific data descriptions, and historical situational sample data. The labeling system includes situational keywords, attribute indicator words, and scenario identifier words. The Min-Max normalization algorithm was used to normalize the original multi-source heterogeneous situational data to obtain normalized data. Based on CNN combined with Transformer model, feature encoding is performed on normalized data to extract global situation feature vector and capture temporal correlation information; Construct a situation domain label library, classify situation scenarios based on heterogeneous feature alignment algorithm, and analyze the feature information of different structured data by combining feature encoding results to form a preliminary feature list; The preliminary feature list is deduplicated, completed, and validated for consistency to generate a standardized feature vector.

4. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 1, characterized in that, The process of deploying a multimodal situational awareness fusion hub includes: A fusion strategy scheduling core is constructed, which includes a situation element metadata management module, a data-situation element mapping rule base, a fusion scheduling strategy unit, and an algorithm execution engine. The situation element metadata management module stores the attribute information of preset situation elements, including element function description, data usage parameters, interface specifications, dependencies, and adaptation scenario tags. Based on the attribute information, dynamic association rules between multi-source heterogeneous situational data and situational elements are established in the data-situation element mapping rule base; The fusion scheduling strategy unit presets a scheduling algorithm based on data priority, a scheduling algorithm based on fusion degree, and a scheduling algorithm based on scenario matching degree, which are used to automatically select the appropriate scheduling algorithm according to the situation and scenario. The algorithm execution engine pre-sets a heterogeneous feature alignment algorithm, a deep deterministic strategy gradient optimization algorithm, and a feature similarity matching algorithm, and provides an algorithm call interface and a parameter configuration interface; A real-time monitoring module is configured for the multimodal situational element fusion hub. The real-time monitoring module is used to collect hub operation status data in real time. The hub operation status data includes data processing rate, element matching accuracy and resource utilization rate. The central operating status data is fed back to the algorithm execution engine for anomaly analysis, enabling anomaly alarms and adaptive optimization.

5. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 4, characterized in that, The process of generating the basic situation feature matrix includes: The multimodal situational element fusion center receives the standardized feature vector and extracts the core features of the data through a feature similarity matching algorithm; Based on the core features of the data, the data-situation element mapping rule base is queried to select candidate situation elements that meet the situation awareness requirements, satisfy the assessment accuracy constraints, and are suitable for the scene characteristics. The candidate situation elements are subjected to data adaptability verification. Combined with the scheduling algorithm selected by the fusion scheduling strategy unit, the matching degree between the data requirements of the candidate situation elements and the available resources of the system is analyzed, and elements that cannot be met by resources are eliminated. The Softmax algorithm is used to calculate the weights of situation elements. Candidate situation elements that pass the verification are grouped according to situation correlation, and a basic situation feature matrix is ​​constructed, including element name, feature description, evaluation parameters, related elements, data requirements and adaptation scenarios. The basic situation feature matrix is ​​subjected to integrity verification. If any elements are missing, the data processing stage is returned to supplement multi-source heterogeneous situation data information until the matrix meets the basic situation assessment coverage requirements.

6. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 1, characterized in that, The process of building an initial situational reasoning prototype and collecting data from the entire model reasoning process includes: Based on the element dependencies in the basic situation feature matrix, an element call chain diagram is constructed, and situation assessment elements are deployed using model orchestration technology. Based on the element call chain diagram, configure the communication protocol and data serialization format between elements, and deploy a real-time data acquisition gateway; The initial situational reasoning prototype was built by starting the situational assessment elements and the real-time data acquisition gateway, and the situational scenario was simulated for trial operation to verify the basic interaction feasibility between the elements. The real-time data acquisition gateway collects the full-process operation data of the model inference prototype according to preset rules, and stores it in a distributed database after cleaning and formatting. The full-process operation data of the model inference includes data processing accuracy, model evaluation latency, resource utilization, and situation identification coverage.

7. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 6, characterized in that, The process of generating personalized model parameter tuning schemes using deep deterministic gradient optimization algorithms includes: The training data set is constructed by extracting the entire process data of model inference from the distributed database, and a reinforcement learning objective function is set with the model evaluation accuracy and resource utilization as the goals. Define the state space, action space, and reward function for reinforcement learning. The state space is the feature vector of the data running through the entire inference process of the model. The action space includes element splitting, element merging, model replacement, parameter reallocation, and interface optimization and adjustment operations. A decision model is constructed using a deep deterministic strategy gradient optimization algorithm, and the model is trained using a training data set to optimize the decision strategy. The trained decision model is used to analyze the model state corresponding to the current basic situation feature matrix, and to identify existing evaluation bottlenecks, data conflicts and scenario adaptation defects. Based on the identification results, a personalized model parameter optimization plan is generated, including the identifier of the element to be adjusted, the type of adjustment operation, the configuration of adjustment parameters, suggestions for synchronous adjustment of related elements, and the expected evaluation indicators after optimization.

8. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 1, characterized in that, The process of forming the final situational reasoning model cluster includes: Construct a multi-dimensional consistency verification indicator system, including element consistency indicators, evaluation matching degree indicators, association compatibility indicators, resource balance indicators, and scenario adaptability indicators, and set qualified thresholds for each indicator; Based on the aforementioned element consistency index, the consistency of model elements after the personalized model parameter tuning scheme is adjusted is verified. Based on the aforementioned evaluation matching index, compare the evaluation parameters of the model with the situational requirement index in the personalized model parameter tuning scheme. Based on the aforementioned correlation compatibility index, the correlation relationships between the optimized elements are analyzed. Based on the aforementioned resource balance index, the resource utilization distribution balance of each element after optimization is evaluated. Based on the aforementioned scenario adaptability metrics, the cross-scenario adaptability of the optimized model set is verified. Calculate the multi-source data fusion degree index, and perform a final verification on the model set that has passed the multi-dimensional consistency check based on the fusion degree index to form the final situation inference model cluster.

9. The method for constructing a situation assessment model that integrates multi-source heterogeneous data according to claim 1, characterized in that, The process of continuous optimization through full-cycle situational awareness and dynamic iteration mechanisms includes: Construct a real-time monitoring system covering model instances and element link levels, and collect operational indicators and data adaptation indicators of situation assessment models that integrate multi-source heterogeneous data; Deploy an adaptive iterative engine to connect with the real-time monitoring system and the situational requirements management platform to identify abnormal indicator data and explicit and implicit scenario changes; Based on the anomaly identification results and scene change information, parameter allocation and model adaptation adjustment strategies are generated by combining the historical optimization case library and then implemented after feasibility verification. Verify the optimization effect. If it does not meet expectations, re-optimize the strategy and save the optimization process, strategy and effect data into the historical optimization case library.

10. A situation assessment model construction system integrating multi-source heterogeneous data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a situation assessment model construction method that integrates multi-source heterogeneous data as described in any one of claims 1 to 9.

Images