Action model learning system and method based on embedded ai

Abstract

This invention belongs to the fields of artificial intelligence, embedded systems, and operational confrontation command technology. It discloses an operational model learning system and method based on embedded AI. The system includes: an information collection module, which collects historical operational information and performs data preprocessing, the historical operational information including operational environment data, confrontation situation information, and decision data; and a feature extraction and model building module, which uses deep learning algorithms to extract features from the preprocessed historical operational information, obtaining dynamic operational change features and enemy operational pattern features. These features are then matched with corresponding decision data results to form a training set, which is used to train an embedded AI model to generate an operational model. This invention can process and learn dynamic data from operations in real time, thereby continuously optimizing the operational model and improving command and decision-making capabilities.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, embedded systems, and action confrontation command technology. More specifically, this invention relates to an action model learning system and method based on embedded AI. Background Technology

[0002] In modern combat operations, operational information is increasing exponentially. How to extract useful data from this massive amount of information and make decisions applicable to the operation has become a new challenge.

[0003] In view of this, the present invention proposes an action model learning system and method based on embedded AI to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an action model learning system based on embedded AI, comprising:

[0005] The information collection module collects historical action information and performs data preprocessing. The historical action information includes action environment data, confrontation situation information, and decision data.

[0006] The feature extraction and model building module uses deep learning algorithms to extract features from preprocessed historical action information, obtaining dynamic action change features and enemy action pattern features. It matches the dynamic action change features and enemy action pattern features with the corresponding decision data results to form a training set. The training set is used to train the embedded AI model to generate an action model.

[0007] The real-time decision-making module collects real-time action information, analyzes and infers the real-time action information through action models to generate real-time decision data, evaluates the generated real-time decision data, and obtains decision evaluation results.

[0008] The online learning module uses real-time action information and decision evaluation results to learn the action model online and gradually improve it.

[0009] The adaptive feedback module automatically adjusts the learning strategy and records decision data for each action scenario. It combines each decision data with action feedback to form a decision log, and uses the decision log to iteratively optimize the action model.

[0010] Furthermore, the method for collecting historical action information and performing data preprocessing includes:

[0011] Historical action information is collected through acquisition devices. The historical action information includes action environment data, confrontation situation information and decision data. An action information database is established based on the historical action information.

[0012] The collected historical action information is preprocessed by cleaning, denoising, and standardizing the data to obtain preprocessed historical action information.

[0013] Furthermore, the action information database includes a static information database and a dynamic information database; wherein, the static information in the static information database refers to data used for control or reference during operation and will not change over a long period of time, and the dynamic information in the dynamic information database refers to data that changes during operation, data that needs to be input and output during operation, and data that needs to be changed.

[0014] Furthermore, the method of extracting features from preprocessed historical action information using deep learning algorithms to obtain dynamic action change features and enemy action pattern features, matching these features with corresponding decision data results to form a training set, and using the training set to train an embedded AI model to generate an action model includes:

[0015] The information collected during the operation is preprocessed through cleaning, fusion, and aggregation to form a data warehouse suitable for data mining;

[0016] The preprocessed action information is used to form a data warehouse suitable for data mining. The data warehouse is then used for data mining.

[0017] Convolutional neural networks are used to mine patterns in data warehouses, and the mined patterns are evaluated to form a model library.

[0018] Further knowledge extraction is performed on the model library to form a knowledge base suitable for action decision-making, which can be used by commanders for decision-making.

[0019] A training sample is formed by combining the dynamic change characteristics of actions within the same time period, the characteristics of enemy action patterns, and the corresponding decision data results. A dataset is formed based on several training samples, and the dataset is divided into a training set and a test set according to the proportion.

[0020] An embedded AI model is built based on a deep neural network. The embedded AI model is trained using a training set to obtain a trained action model.

[0021] Furthermore, the knowledge extraction method includes:

[0022] Data preprocessing removes noisy and duplicate data from the model library, handles missing values, and performs standardization, normalization, and other transformation operations on the data according to knowledge extraction requirements to unify the data format.

[0023] Pattern recognition and analysis employs the Apriori algorithm to analyze data in the model library to uncover the relationships between different data items, and the K-Means algorithm to cluster similar data objects into a class to obtain the natural grouping structure in the data. For existing classification models, the feature weights and decision boundaries are analyzed to obtain the impact of different features on the classification results.

[0024] Knowledge extraction involves extracting understandable rule knowledge from the relationships between different data items and the impact of different features on classification results to identify features that have a significant impact on the model results. As part of the knowledge, clustering results are summarized to describe the characteristics and commonalities of each category.

[0025] By comparing with known correct knowledge or test data, the accuracy of the extracted knowledge is evaluated, and from a business perspective, it is determined whether the knowledge has practical application value and can help solve practical problems. The extracted knowledge is checked for duplication or redundancy, and redundant knowledge is removed.

[0026] Knowledge entry involves transforming evaluated and filtered knowledge according to the knowledge base's storage format, storing the formatted knowledge in the knowledge base, and selecting a suitable database management system or knowledge graph platform for storage.

[0027] Furthermore, the method of collecting real-time action information, analyzing and reasoning about the real-time action information through an action model to generate real-time decision data, and evaluating the generated real-time decision data to obtain decision evaluation results includes:

[0028] Real-time action information is collected using data acquisition devices, including real-time action environment data, real-time confrontation situation information, and real-time decision data.

[0029] Input real-time action information into the action model, and the action model can output real-time decision data;

[0030] The generated real-time decision data is evaluated based on feasibility and effectiveness to obtain decision evaluation results.

[0031] Furthermore, the method of gradually improving the action model through online learning based on real-time action information and decision evaluation results includes:

[0032] Based on real-time data and decision evaluation results, the action model learns online, automatically adjusts the parameters of the action model, and optimizes its decision data;

[0033] By repeatedly training and updating the action model, the action model can be gradually improved so that it can better adapt to changes in action.

[0034] Furthermore, the action model iterative method based on embedded AI, which involves automatically adjusting the learning strategy and recording decision data for each action scenario, forming a decision log with each decision data and action feedback, and using the decision log to iteratively optimize the action model, includes:

[0035] Acquire action information from different action scenarios, and automatically adjust the learning strategy using the action information from different action scenarios to obtain decision data;

[0036] Each decision data point and action feedback information are used to form a decision log, which is then used to iteratively optimize the action model.

[0037] Action model learning methods based on embedded AI include:

[0038] S1. Collect historical action information and perform data preprocessing, wherein the historical action information includes action environment data, confrontation situation information and decision data;

[0039] S2. Use deep learning algorithms to extract features from preprocessed historical action information to obtain action dynamic change features and enemy action pattern features. Match the action dynamic change features and enemy action pattern features with the corresponding decision data results to form a training set. Use the training set to train the embedded AI model to generate an action model.

[0040] S3. Collect real-time action information, analyze and reason about the real-time action information through action models to generate real-time decision data, evaluate the generated real-time decision data, and obtain decision evaluation results.

[0041] S4. Based on real-time action information and decision evaluation results, conduct online learning of the action model to gradually improve the action model;

[0042] S5. Automatically adjust the learning strategy and record each decision data using different action scenarios. Combine each decision data with action feedback to form a decision log, and use the decision log to iteratively optimize the action model.

[0043] The technical effects and advantages of the action model learning system and method based on embedded AI of this invention are as follows:

[0044] This invention leverages the rapid data processing capabilities of an embedded AI system to generate command decisions in real time, thereby improving command efficiency. Through online learning and model updates, the action model can adapt to changes in the action environment in a timely manner, ensuring the effectiveness of decisions. By employing advanced technologies such as deep learning, the intelligence level of the action model is enhanced, enabling it to automatically extract useful information from complex data. Through repeated iterations and optimizations, the decision-making accuracy of the action model is continuously improved, effectively reducing the risk of misjudgment and decision-making errors. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0046] Figure 2 This is a schematic diagram of the action model learning process of the present invention;

[0047] Figure 3 This is a situational information block diagram of the present invention;

[0048] Figure 4 This is a management action information situation diagram of the present invention;

[0049] Figure 5 This is a control block diagram of the embedded AI command system of the present invention;

[0050] Figure 6 This is a schematic diagram of the convolutional neural network of the present invention;

[0051] Figure 7 This is a schematic diagram of the structure of the action model learning system based on embedded AI of the present invention;

[0052] Figure 8 This is a flowchart illustrating the action model learning method based on embedded AI of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1

[0055] Please see Figures 1 to 7 As shown, the action model learning system based on embedded AI described in this embodiment includes:

[0056] The information collection module collects historical action information and performs data preprocessing. The historical action information includes action environment data, confrontation situation information, and decision data.

[0057] The feature extraction and model building module uses deep learning algorithms to extract features from preprocessed historical action information, obtaining dynamic action change features and enemy action pattern features. It matches the dynamic action change features and enemy action pattern features with the corresponding decision data results to form a training set. The training set is used to train the embedded AI model to generate an action model.

[0058] The real-time decision-making module collects real-time action information, analyzes and infers the real-time action information through action models to generate real-time decision data, evaluates the generated real-time decision data, and obtains decision evaluation results.

[0059] The online learning module uses real-time action information and decision evaluation results to learn the action model online and gradually improve it.

[0060] The adaptive feedback module automatically adjusts the learning strategy and records decision data for each action scenario. It combines each decision data with action feedback to form a decision log, and uses the decision log to iteratively optimize the action model.

[0061] Furthermore, the method for collecting historical action information and performing data preprocessing includes:

[0062] Historical action information is collected through data acquisition devices (such as sensors, radar, sonar, drones, satellites, etc.). This historical action information includes action environment data, confrontation situation information, and decision-making data. An action information database is established based on this historical action information, and the action information situation is as follows: Figure 3 As shown;

[0063] Data preprocessing, such as cleaning, denoising, and standardizing, is performed on the collected historical action information to obtain preprocessed historical action information, which can ensure the validity and accuracy of the historical action information.

[0064] It should be noted that the operational environment data includes, but is not limited to: topographic data, such as mountains, plains, rivers, and lakes; meteorological data, such as temperature, humidity, wind speed, and precipitation; hydrological data, such as river flow, water depth, and tides; geological data, such as soil type and geological structure; population distribution, such as residential areas and population density; transportation networks, such as roads, bridges, railways, and airports; infrastructure, such as electricity, communications, and water supply; economic resources, such as industrial areas, agricultural areas, and mineral distribution; electromagnetic spectrum, such as radio frequencies and radar signals; communication networks, such as communication base stations and satellite communications; interference sources, such as electronic jamming equipment and electromagnetic interference sources; countermeasures deployment, such as the troop deployment and position locations of both sides; mission actions, such as the action routes and action plans of both sides; and operational dynamics, such as real-time status and operational progress.

[0065] Information on the confrontation situation includes, but is not limited to: the deployment of forces, weapon configuration, reaction capability, mobility capability, electromagnetic capability, intelligence gathering capability, overall operational capability, fortifications, logistical support, communications and command, the strength of friendly forces, the operational missions of friendly forces, and possible routes for friendly support operations.

[0066] Historical decision-making data includes, but is not limited to: attack, defense, etc.

[0067] Specifically, action information is collected through various sensors (such as radar and sonar) to establish an action information database, and the action information situation is as follows: Figure 3 As shown; information management at the network center is the foundation for acquiring, storing, processing, and using operational situation data. Establishing and maintaining an effective operational information network is crucial for real-time and accurate understanding of operational situation. Managing situation information involves providing an integrated data model, storing and accessing data through PC services, and ensuring that specified content can be displayed, interacted with, and plotted in a PC environment. The PC-managed operational situation data diagram is shown below. Figure 4 As shown.

[0068] Furthermore, the method of extracting features from preprocessed historical action information using deep learning algorithms to obtain dynamic action change features and enemy action pattern features, matching these features with corresponding decision data results to form a training set, and using the training set to train an embedded AI model to generate an action model includes:

[0069] The information collected during the operation is preprocessed through cleaning, fusion, and aggregation to form a data warehouse suitable for data mining;

[0070] The preprocessed action information is used to form a data warehouse suitable for data mining. The data warehouse is then used for data mining.

[0071] Convolutional neural networks are used to mine patterns in data warehouses, and the mined patterns are evaluated to form a model library.

[0072] Further knowledge extraction is performed on the model library to form a knowledge base suitable for action decision-making, which is used by commanders. Information management, perception, knowledge creation, and decision-making are interconnected, forming a complete action command process. The control block diagram of the embedded AI command system is shown below. Figure 5 As shown, specifically, information management and perception correspond to the information collection module, knowledge creation corresponds to the feature extraction and model building module, and decision-making corresponds to the real-time decision-making module;

[0073] A training sample is formed by combining the dynamic change characteristics of actions within the same time period, the characteristics of enemy action patterns, and the corresponding decision data results. A dataset is formed based on several training samples, and the dataset is divided into a training set and a test set according to the proportion.

[0074] An embedded AI model is built based on a deep neural network. The embedded AI model is trained using a training set to obtain a trained action model.

[0075] It should be noted that a typical neural network includes an input layer, hidden layers, and an output layer; a neural network with more than two hidden layers is called a deep neural network. Both typical and deep neural networks are fully connected; those that are not fully connected but only maintain partial connections are called convolutional neural networks. A schematic diagram of a convolutional neural network is shown below. Figure 6 As shown; the core of a convolutional neural network is the convolutional layer, which extracts feature information through convolution operations. Generally, after the convolution operation, non-linear activation is used to simulate complex inputs; common non-linear activation functions include: Sigmoid, ReLU, LeakyReLU, etc.; the formulas for the ReLU function and its derivative are shown below;

[0076] ,

[0077] ,

[0078] The drawback of the ReLU activation function is that the derivative is 0 in the negative region. During model training, if the input falls into the negative region (…), the model may fail to activate. The network will also be unable to continue updating weights. The LeakyReLU function can solve the neuron death problem encountered by ReLU, and its calculation formula is shown below;

[0079] ,

[0080] The core idea of ​​the LeakyReLU function is that when When the function has a positive gradient of 0.1, the ParametricReLU function is based on LeakyReLU, but replaces the positive gradient of 0.1 with random hyperparameters. This allows the neuron to select the best gradient in the negative region; the calculation formula is shown below.

[0081] ,

[0082] like Figure 6 As shown, in addition to the input layer, output layer, and convolutional layer, there is also a pooling layer. The role of the pooling layer is to extract significant features from the data and ignore detailed features, thereby reducing the parameters and computation in the network and controlling overfitting. The output layer performs the final processing on the features obtained after convolution and pooling, completes the task, and outputs the result.

[0083] Specifically, this invention uses the TensorFlow open-source framework. First, the model parameters are initialized by randomly assigning values. For simpler models, the initial values ​​of each parameter are set to 0. Next, the allocated training data samples and their expected outputs are read. Then, the inference model is executed on the training data, producing an output for each training sample. The loss is then calculated based on the training data and the expected output. Finally, the model parameters are continuously adjusted. Under the constraint of a given loss function, the values ​​of each parameter are improved through numerous training steps to minimize the loss. The gradient descent algorithm tf.gradients provided by TensorFlow is used, and the above process is repeated through numerous loops. After training, the model is evaluated. The model is used to infer from different test sets containing the same expected output information, and the loss on these test sets is evaluated to understand the model's recognition ability outside the training set.

[0084] Once the accuracy of the model is satisfactory, the model is exported and the model directory file is imported into the embedded system. The embedded system uses the model to collect action information through sensors, inputs it into the model for verification, and the model automatically judges and classifies the information to obtain control commands (i.e., decision data).

[0085] Furthermore, the knowledge extraction method includes:

[0086] Data preprocessing involves removing noisy and duplicate data from the model library and handling missing values. Specifically, this includes selecting appropriate methods based on data characteristics, such as deleting missing value records or filling in the mean or mode. According to knowledge extraction requirements, data is standardized and normalized to unify the data format, such as unifying text data to lowercase for easier subsequent processing.

[0087] Pattern recognition and analysis employs the Apriori algorithm to analyze data in the model library to uncover the relationships between different data items, and the K-Means algorithm to cluster similar data objects into a class to obtain the natural grouping structure in the data. For existing classification models, the feature weights and decision boundaries are analyzed to obtain the impact of different features on the classification results.

[0088] Knowledge extraction involves extracting understandable rule knowledge from the relationships between different data items and the impact of different features on classification results to identify features that have a significant impact on the model results. As part of the knowledge, clustering results are summarized to describe the characteristics and commonalities of each category.

[0089] The accuracy of the extracted knowledge is evaluated by comparing it with known correct knowledge or test data. For example, the prediction accuracy of the extracted rules on the test dataset is checked. From a business perspective, it is judged whether the knowledge has practical application value and can help solve practical problems. The duplicates or redundancies among the extracted knowledge are checked and redundant knowledge is removed. If two rules are found to express essentially the same meaning, the one that is more concise or representative is retained.

[0090] Knowledge entry involves transforming evaluated and filtered knowledge according to the knowledge base's storage format, such as converting rule knowledge into XML or JSON format, and storing the formatted knowledge in the knowledge base. A suitable database management system or knowledge graph platform is selected for storage to facilitate knowledge retrieval, updating, and sharing.

[0091] Furthermore, the action information database includes a static information database and a dynamic information database; wherein, the static information in the static information database refers to data used for control or reference during operation and will not change over a long period of time, and the dynamic information in the dynamic information database refers to data that changes during operation, data that needs to be input and output during operation, and data that needs to be changed.

[0092] It should be noted that the operational information database is divided into static information databases and dynamic information databases. Static information refers to data used for control or reference during operation, which remains unchanged over a long period. Its characteristics include: low data update frequency, long-term immutability; primarily used for query operations, with less writing and updating; high data consistency requirements, typically not requiring frequent synchronization; simple storage structure design for easy and fast querying; and technically, it can be implemented using relational databases suitable for structured data, such as MySQL and PostgreSQL, and NoSQL databases suitable for semi-structured or unstructured data, such as MongoDB. Static information in the operational environment mainly includes: the parameters of the equipment of both sides, the terrain of the operational area, etc.

[0093] Dynamic information refers to data that changes during operation, data that needs input and output during operation, and data that needs to be changed. Its characteristics are: high data update frequency and strong real-time performance; support for a large number of concurrent read and write operations; requirement for fast response and low latency; and support for horizontal scaling to cope with high load. Technically, it can be implemented through in-memory databases suitable for low latency requirements, such as Redis and Memcached; time-series databases suitable for event sequence data, such as InfluxDB; and high-concurrency databases suitable for high concurrency and large-scale data, such as Cassandra. Dynamic information is mainly reflected in: weather conditions, action mobilization, equipment losses, personnel numbers, confrontation situation, energy consumption, command, control, communication, strategic judgment, decision-making, and planning of decision-makers.

[0094] Furthermore, the method of collecting real-time action information, analyzing and reasoning about the real-time action information through an action model to generate real-time decision data, and evaluating the generated real-time decision data to obtain decision evaluation results includes:

[0095] Real-time action information is collected using data acquisition devices, including real-time action environment data, real-time confrontation situation information, and real-time decision data.

[0096] Input real-time action information into the action model, and the action model can output real-time decision data;

[0097] The generated real-time decision data is evaluated based on feasibility and effectiveness to obtain decision evaluation results.

[0098] It should be noted that the specific implementation process is as follows: Real-time action information can be collected through a multi-source sensor network to obtain high-precision, low-latency action dynamic data; optical / infrared satellites can provide panoramic images of the action area; drone swarms achieve accurate indication of moving targets; ground reconnaissance vehicles are equipped with relevant sensors to detect signals; individual soldier wearable devices integrate biosensors to transmit soldiers' vital signs in real time; data fusion is used to align heterogeneous sensor data; Kalman filtering is used to eliminate GPS positioning errors; time synchronization equipment ensures time synchronization of multi-source data (error <1ms), and outputs a JSON-formatted structure containing coordinates, speed, equipment type, etc. The system generates a situational awareness map; the action model analysis and reasoning can be divided into two steps: situational understanding and decision generation. Situational understanding can be based on equipment identification, Bayesian network threat assessment, and spatiotemporal pattern analysis (using LSTM to predict troop movement trajectories) using YOLOv7; decision generation can generate action plans through Monte Carlo search to achieve 450 action simulations per second; the output is a set of decision plans in XML format containing operational instructions such as resource allocation and route planning; decision evaluation and optimization can calculate the expected action loss ratio through the Lanchester equation, predict material consumption based on the SEAS model, simulate communication network load, update model parameters through Federated Learning, and maintain 15% reserves to cope with more than 80% of contingencies; the output is a decision score (0-100 points, with confidence interval markings).

[0099] Furthermore, the method of gradually improving the action model through online learning based on real-time action information and decision evaluation results includes:

[0100] Based on real-time data and decision evaluation results, the action model learns online, automatically adjusts the parameters of the action model, and optimizes its decision data;

[0101] By repeatedly training and updating the action model, the action model can be gradually improved so that it can better adapt to changes in action.

[0102] It should be noted that the specific implementation process is as follows: online learning is carried out through incremental learning algorithms, such as online random forest algorithm: 50 decision trees are updated for each batch of data; online training of deep Q network (DQN): policy network parameters are updated every 60 seconds; federated learning architecture: edge nodes (such as drones) are trained locally, and the central node aggregates the model; online knowledge distillation is carried out through teacher-student model mechanism; teacher model: 128-layer ResNet (98% accuracy) deployed in the command center, MobileNetV3 (92% accuracy) that can run on student edge devices, and model compression is performed every 24 hours (maintaining <5% accuracy loss).

[0103] Furthermore, the method of automatically adjusting the learning strategy and recording decision data for each action scenario, forming a decision log with each decision data and action feedback, and using the decision log to iteratively optimize the action model includes:

[0104] Acquire action information from different action scenarios, and automatically adjust the learning strategy using the action information from different action scenarios to obtain decision data;

[0105] Each decision data point and action feedback information are used to form a decision log, which is then used to iteratively optimize the action model.

[0106] It should be noted that the specific implementation process is as follows: The system has a built-in "action environment parser" that can automatically identify the characteristic dimensions of the current action scenario: terrain features: geographical features such as mountain assault, urban lane operations, and plains maneuver; confrontation intensity: action levels such as low-intensity standoff and full-scale operation; equipment composition: equipment combinations such as drone swarm confrontation and electronic operation dominance; time pressure: time-sensitive requirements such as second-level response air defense interception and hour-level action planning; each identified scenario feature will activate the corresponding "learning strategy package": enabling a rapid exploration strategy using reinforcement learning, focusing on maneuver route optimization; consumption action scenario: using Bayesian optimization methods to finely calculate the supply consumption ratio; starting the transfer learning module to call historical similar equipment response cases; constructing a three-layer decision log architecture: the original decision layer, which records a complete snapshot of the input data during model inference, saves all candidate decision schemes and scores, and marks the weight basis for the final selection decision; the execution tracking layer, which records the action unit... The system includes: a meta-level understanding of instructions (confirmation / questioning feedback), verification of command transmission delay and completeness, and recording of fine-tuning actions during actual execution; an effect feedback layer, a visual evidence chain for damage assessment (comparison with drone aerial photography), sensor credibility correction in electronic operations environments, and indirect effect indicators (such as sudden changes in communication volume); iterative optimization based on decision logs, enabling "microscopic" review of key turning point decisions: reconstructing the information completeness at the time of decision-making, quantifying the potential benefit distribution of each alternative, and comparing the judgments of human experts under the same situation; cross-scenario decision pattern recognition: discovering similar flanking awareness vulnerabilities in mountain and urban operations, identifying common supply line vulnerabilities in conflicts of different scales, and establishing a matching rule base for equipment combinations and optimal action decisions; model parameter adjustment: strengthening the training of neural network branches that repeatedly exhibit judgment biases, and adjusting the weight coefficients of outdated action rules in the decision tree; and knowledge graph updates: adding special terrain passability parameters.

[0107] In this embodiment, the rapid data processing capabilities of the embedded AI system enable real-time generation of command decisions, improving command efficiency. Through online learning and model updates, the action model can adapt to changes in the action environment in a timely manner, ensuring the effectiveness of decisions. By adopting advanced technologies such as deep learning, the intelligence level of the action model is improved, enabling it to automatically extract useful information from complex data. Through repeated iterations and optimizations, the decision accuracy of the action model is continuously improved, effectively reducing the risk of misjudgment and decision-making errors.

[0108] Example 2

[0109] Please see Figure 8 As shown in this embodiment, the action model learning method based on embedded AI includes the following steps:

[0110] S1. Collect historical action information and perform data preprocessing, wherein the historical action information includes action environment data, confrontation situation information and decision data;

[0111] S2. Use deep learning algorithms to extract features from preprocessed historical action information to obtain action dynamic change features and enemy action pattern features. Match the action dynamic change features and enemy action pattern features with the corresponding decision data results to form a training set. Use the training set to train the embedded AI model to generate an action model.

[0112] S3. Collect real-time action information, analyze and reason about the real-time action information through action models to generate real-time decision data, evaluate the generated real-time decision data, and obtain decision evaluation results.

[0113] S4. Based on real-time action information and decision evaluation results, conduct online learning of the action model to gradually improve the action model;

[0114] S5. Automatically adjust the learning strategy and record each decision data using different action scenarios. Combine each decision data with action feedback to form a decision log, and use the decision log to iteratively optimize the action model.

[0115] In this embodiment, the present invention can generate command decisions in real time through the rapid data processing capabilities of the embedded AI system, thereby improving command efficiency; furthermore, through repeated iteration and optimization, the decision accuracy of the action model is continuously improved, effectively reducing the risk of misjudgment and decision-making errors.

[0116] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0117] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0119] In conclusion, 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 spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An action model learning system based on embedded AI, characterized in that, include: The information collection module collects historical action information and performs data preprocessing. The historical action information includes action environment data, confrontation situation information, and decision data. An action information database is established based on historical action information. The action information database includes a static information database and a dynamic information database. The static information in the static information database includes equipment parameters of both sides and terrain of the action area. The dynamic information in the dynamic information database includes weather conditions, action mobilization, equipment losses, personnel numbers, action situation, and command and control communication information of decision-makers. The feature extraction and model building module uses deep learning algorithms to extract features from preprocessed historical action information, obtaining dynamic action change features and enemy action pattern features. These features are then matched with corresponding decision data to form a training set. The feature extraction and model building module constructs the action model in the following ways: The information collected during the operation is preprocessed by cleaning, merging, and aggregating to form a data warehouse; Convolutional neural networks are used to mine patterns in data warehouses, and the mined patterns are evaluated to form a model library. Knowledge extraction is performed on the model library to form a knowledge base. The knowledge extraction includes: using the Apriori algorithm to analyze the data in the model library to mine the relationship between different data items, and using the K-Means algorithm to cluster similar data objects into one class to obtain the natural grouping structure in the data. An embedded AI model is built based on a deep neural network, and the embedded AI model is trained using a training set to generate an action model. The knowledge extraction also includes: extracting rule knowledge from the relationships between different data items, identifying features that have a significant impact on the model results; summarizing the clustering results, describing the characteristics and commonalities of each category; and storing the evaluated and filtered knowledge in the knowledge base after converting it according to the knowledge base's storage format. The real-time decision-making module collects real-time action information, analyzes and infers the real-time action information through action models to generate real-time decision data, evaluates the generated real-time decision data, and obtains decision evaluation results. The online learning module uses real-time action information and decision evaluation results to learn the action model online and gradually improve it. The adaptive feedback module automatically adjusts the learning strategy and records each decision data using different action scenarios. It combines each decision data with action feedback to form a decision log, and uses the decision log to iteratively optimize the action model. The adaptive feedback module constructs a three-layer decision log architecture, including: The original decision layer records a complete snapshot of the input data during model inference, all candidate decision schemes and their scores, and the weighting basis for the final selection decision; The execution tracking layer records the action unit's understanding of the instructions, verifies the delay and completeness of command transmission, and records the fine-tuning of actions during actual execution. The effect feedback layer records the visual evidence chain of damage assessment, sensor credibility correction, and indirect effect indicators.

2. The action model learning system based on embedded AI according to claim 1, characterized in that, The methods for collecting historical action information and performing data preprocessing include: Historical action information is collected through data acquisition devices; The collected historical action information is preprocessed by cleaning, denoising, and standardizing the data to obtain preprocessed historical action information.

3. The action model learning system based on embedded AI according to claim 2, characterized in that, The static information in the static information database refers to data that is used for control or reference during operation and will not change over a long period of time. The dynamic information in the dynamic information database refers to data that changes during operation, data that needs to be input and output during operation, and data that needs to be changed.

4. The action model learning system based on embedded AI according to claim 1, characterized in that, The method of collecting real-time action information, analyzing and reasoning about the real-time action information through action models to generate real-time decision data, evaluating the generated real-time decision data, and obtaining decision evaluation results includes: Real-time action information is collected using data acquisition devices, including real-time action environment data, real-time confrontation situation information, and real-time decision data. Input real-time action information into the action model, and the action model can output real-time decision data; The generated real-time decision data is evaluated based on feasibility and effectiveness to obtain decision evaluation results.

5. The action model learning system based on embedded AI according to claim 1, characterized in that, The method of gradually improving the action model through online learning based on real-time action information and decision evaluation results includes: Based on real-time data and decision evaluation results, the action model learns online, automatically adjusts the parameters of the action model, and optimizes its decision data; By repeatedly training and updating the action model, the action model can be gradually improved so that it can better adapt to changes in action.

6. The action model learning system based on embedded AI according to claim 4, characterized in that, The method of automatically adjusting the learning strategy and recording decision data for each action scenario, forming a decision log with each decision data and action feedback, and using the decision log to iteratively optimize the action model includes: Acquire action information from different action scenarios, and automatically adjust the learning strategy using the action information from different action scenarios to obtain decision data; Each decision data point and action feedback information are used to form a decision log, which is then used to iteratively optimize the action model.

7. An action model learning method based on embedded AI, applied to the action model learning system based on embedded AI as described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Collect historical action information and perform data preprocessing, wherein the historical action information includes action environment data, confrontation situation information and decision data; S2. Use deep learning algorithms to extract features from preprocessed historical action information to obtain dynamic action change features and enemy action pattern features. Match the dynamic action change features and enemy action pattern features with the corresponding decision data results to form a training set. Through data warehouse construction, model library formation, and knowledge extraction, combine the Apriori algorithm and K-Means algorithm to extract knowledge. Use the training set to train the embedded AI model to generate an action model. S3. Collect real-time action information, analyze and reason about the real-time action information through action models to generate real-time decision data, evaluate the generated real-time decision data, and obtain decision evaluation results. S4. Based on real-time action information and decision evaluation results, conduct online learning of the action model to gradually improve the action model; S5. Automatically adjust the learning strategy and record each decision data using different action scenarios. Form a decision log by combining each decision data with action feedback through a three-layer decision log architecture. Use the decision log to iteratively optimize the action model.

Images