Intelligent police service simulation combat training system based on ai large model

Abstract

This invention belongs to the field of artificial intelligence and deep learning, specifically relating to a smart policing simulation training system based on a large AI model. The system includes a cross-modal data parsing device, a spatiotemporal causal knowledge graph construction device, a large model-driven game-theoretic adversarial generator, a trainee behavior capture device, and a dual-track evaluation and feedback module. It constructs a dynamically evolving knowledge hub by parsing multi-source heterogeneous data across modalities, and generates emergent real-world adversarial scenarios using a large model and multi-agent reinforcement learning. It captures and quantifies multi-dimensional interaction data of trainees in real time and performs pattern matching with a Pareto-optimal policy set, achieving a dual-track comprehensive evaluation of both process and result. Through a data-driven closed-loop iterative mechanism, this invention accurately recreates complex policing relationships and eliminates the simulation gap, improving the realism and relevance of exercises and providing core technological support for building a modern smart policing training system.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and deep learning, specifically involving a smart policing simulation training system based on a large AI model. Background Technology

[0002] With the deep integration of artificial intelligence and big data technologies, smart policing simulation training has become an important means for improving the practical skills and emergency decision-making capabilities of police officers in the modern public security system. This training method, by constructing a digital simulation environment, aims to break through the limitations of traditional teaching in terms of time, space, and cost, achieving a deep simulation of diverse police situations. Especially in complex law enforcement environments, the system's ability to perceive the police situation, the logical consistency of scenario generation, and the interactive depth of adversarial drills determine the effectiveness of translating training results into practical application.

[0003] Large-scale model-driven intelligent policing simulation training systems offer an advanced solution for enhancing the randomness of scenario generation and the interactive capabilities of intelligent agents. These systems utilize generative artificial intelligence to deconstruct features from massive amounts of police case data, attempting to simulate the adaptive behavioral logic of suspects, bystanders, and other roles in real law enforcement processes. Their core objective is to construct a highly dynamic and strategically adversarial virtual combat space, enabling trainees to conduct information analysis and tactical execution in the ever-evolving context of police situations.

[0004] Existing systems mostly employ rule-driven or script-oriented static simulation methods, making it difficult to generate high-dimensional adversarial scenarios with emergent and asymmetric characteristics. This results in a simulation gap between the training environment and real-world scenarios. Traditional technologies for utilizing police incident data are largely limited to simple text-level structuring, lacking a deep understanding of the spatiotemporal causal relationships and behavioral evolution patterns within multimodal environmental data. This makes it difficult to reproduce the information fog and decision-making pressure of real-world scenarios. Furthermore, existing evaluation systems tend to focus on a single assessment of the outcome, failing to accurately quantify trainees' information retrieval efficiency, risk assessment completeness, and cognitive logic in the decision-making process. This hinders the formation of a closed-loop and personalized capability enhancement path. Summary of the Invention

[0005] The purpose of this invention is to provide a smart policing simulation training system based on a large artificial intelligence model, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart policing simulation training system based on a large-scale artificial intelligence model includes a cross-modal data parsing device, a spatiotemporal causal knowledge graph construction device, a large-scale model-driven game-theoretic generator, a trainee behavior capture device, and a dual-track evaluation and feedback module, wherein: The cross-modal data parsing device is used for deep feature extraction and semantic association of multi-source heterogeneous police data. This device receives historical police case files, audio and video data from law enforcement recorders, urban geographic information system data, and social demographic data. Through built-in parsing algorithms, the device transforms text files into structured event sequences, audio data into sentiment-annotated text instructions, video data into behavior streams containing target trajectories and action semantics, and geographic information data into digital twin environmental spatial features with topological relationships.

[0007] The spatiotemporal causal knowledge graph construction device, connected to the cross-modal data parsing device, is used to construct a dynamically evolving police knowledge hub. This device maps the parsed heterogeneous data to a unified semantic space, encapsulating not only the relationships between basic police elements such as people, events, places, objects, and organizations, but also introducing fine-grained timestamp tags and event causal chains. By calculating the probability of behavioral patterns within a specific geographical area over a preset time period, the device reveals non-obvious social operating rules and probabilistically models the subsequent chain reactions that law enforcement actions may trigger, forming a dynamic graph foundation that supports logical deduction.

[0008] The large-scale model-driven adversarial generator, connected to a spatiotemporal causal knowledge graph construction device, is used to generate emergent real-world adversarial scenarios. This generator comprises a central generation hub with cognitive reasoning capabilities and multiple multi-agent reinforcement learning units with independent decision-making abilities. Based on causal logic within the knowledge graph, the central generation hub extracts key elements from the trainees' ability profiles to generate highly logically consistent training backgrounds and personality settings for non-player characters. Subsequently, the multi-agent reinforcement learning units take over the behavioral decisions of suspects, bystanders, and victims. Each agent shares the same adversarial reward function, dynamically generating optimal counter-surveillance and anti-law enforcement strategies by real-time perception of environmental changes, trainee behavior, and legal boundaries.

[0009] The trainee behavior capture device is used to acquire and quantify multidimensional interaction data of trainees in real time throughout the training process. The data captured by this device includes, but is not limited to, information retrieval keywords entered by the trainee on the police terminal, patrol routes in virtual space, semantics of communication phrases when interacting with non-player characters, and tactical action sequences when responding to emergencies. The device transforms the captured raw behavioral data into standardized semantic feature vectors, providing refined input for subsequent evaluation.

[0010] The dual-track evaluation and feedback module is connected to both the large-model-driven game adversarial generator and the trainee behavior capture device, and is used for comprehensive evaluation of the execution process and results. This module utilizes process mining technology to perform pattern matching and bias analysis between the trainee's real-time behavior vectors and the high-dimensional Pareto optimal strategy set pre-derived by the game adversarial generator. Evaluation dimensions include information retrieval efficiency, comprehensiveness of risk assessment, compliance of tactical choices, and timeliness of decision-making. Based on the evaluation results, this module generates personalized capability gap profiles and feeds back correction parameters to the large-model-driven game adversarial generator to dynamically adjust the complexity and adversarial intensity of subsequent training scenarios, forming a closed-loop iterative path of evaluation, generation, training, and re-evaluation.

[0011] Preferably, the cross-modal data parsing device includes a video semantic deconstruction unit. This unit is used to track human skeleton nodes in law enforcement recorder videos in real time, analyze the changes in skeleton displacement velocity and angle, identify potential violent resistance to law enforcement or dangerous object extraction actions, and convert them into semantic tags with risk level labels, which are then stored in a spatiotemporal causal knowledge graph.

[0012] Furthermore, the spatiotemporal causal knowledge graph construction device includes a causal conflict detection unit. When newly input alarm data conflicts with the logical chain in the existing knowledge graph, this unit is used to activate the conflict judgment logic. By comparing the credibility weights of multi-source intelligence, it calculates whether the conflict point belongs to a strategic deviation under special circumstances, updates the causal probability distribution in the graph, and ensures the freshness of the knowledge center.

[0013] Furthermore, the large-model-driven adversarial generator includes a cooperative game engine. This engine coordinates the interaction between the suspect agent and the bystander agents. When the trainee takes coercive measures against the suspect, the cooperative game engine, based on a herd mentality model, drives the bystander agents to exhibit specific levels of agitation or obstructive behavior, simulating public opinion pressure and physical interference in a real policing environment, thus increasing the asymmetric adversarial characteristics of the training.

[0014] Furthermore, the trainee behavior capture device includes a voice emotion analysis unit. This unit is used to analyze the trainee's tone, speech rate, and keyword selection when conversing with a non-player character. By calculating emotional valence and arousal, it assesses the trainee's psychological stability and verbal control when facing high-pressure situations, and uses this as an important reference indicator for tactical compliance assessment.

[0015] Furthermore, the dual-track evaluation and feedback module includes a Pareto optimal policy evolution unit. This unit utilizes a game theory model to generate multiple non-dominant policy sets for the current scenario before training begins. These policy sets represent optimal balance points across multiple dimensions, including compliance, efficiency, and security. During evaluation, the system calculates the Euclidean distance of the trainee's actual decision path in the policy space relative to the Pareto front, thereby quantifying the trainee's decision quality.

[0016] Furthermore, upon receiving a trainee's query for specific suspect information, the large-model-driven game-playing adversarial generator updates the suspect agent's perception state in real time. The suspect agent utilizes geographical information and social relationship data from a spatiotemporal causal knowledge graph to autonomously search for the optimal escape path within a specific range, or to interfere with the trainee's subsequent judgment by creating false information.

[0017] Furthermore, the system also includes a data feedback device. This device is used to automatically anonymize the trainees' successful decision sequences and typical failure cases in each training session, transforming them into logical nodes with specific weights, and then injecting them back into the spatiotemporal causal knowledge graph construction device. Through this self-evolutionary mechanism, the system can continuously learn and grasp the latest trends in crime evolution, ensuring that the training scenarios are always synchronized with real-world needs.

[0018] Furthermore, when processing demographic data, the spatiotemporal causal knowledge graph construction device transforms it into a group behavior density function within a specific time period. When the training scenario is set in a specific dense area, the system dynamically adjusts the cardinality and interaction frequency of non-player characters based on this density function, thus restoring the information stickiness of a real urban environment.

[0019] Furthermore, the dual-track evaluation and feedback module introduces a cognitive computing model to analyze the logical connections between trainees when processing conflicting information from multiple sources. If a trainee fails to identify maliciously altered simulated information within a preset time, the system will mark its redundancy or deficiency in a specific cognitive dimension and, in the next round of generator parameter adjustment, specifically increase the frequency of such ambiguous information to enhance its reactive decision-making ability.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves in-depth value mining of massive heterogeneous police data by constructing a cross-modal data parsing device and a spatiotemporal causal knowledge graph. The system is no longer limited to simple text structuring but can accurately reconstruct the complex relationships and dynamic evolution patterns of people, events, places, and objects in the police environment. This design, based on spatiotemporal causality, provides solid knowledge support for the subsequent generation of high-fidelity, logically consistent training scenarios.

[0021] 2. This invention, through a large-model-driven game adversarial generator, completely breaks through the limitations of traditional scripted and linear training systems. Utilizing the deep integration of large-scale artificial intelligence models and multi-agent reinforcement learning, the system can proactively generate complex scenarios with emergent characteristics. Non-player characters possess adaptive adversarial capabilities, adjusting their strategies in real time based on the trainee's behavior. This qualitative leap from pre-set to emergent allows the training intensity and complexity to infinitely approximate real-world combat, eliminating the simulation gap between training and actual combat.

[0022] 3. The dual-track evaluation mechanism of process and results introduced in this invention achieves transparency and personalization in the evaluation system. Through process mining and comparison with Pareto optimal strategy sets, the system can delve from macro-level result-oriented analysis to micro-level cognitive logic analysis, accurately quantifying trainees' performance in key areas such as information processing, risk assessment, and tactical decision-making. It can reveal trainees' skill gaps and provide them with evidence-based, personalized skill enhancement paths, improving the quality and relevance of practical exercises.

[0023] 4. This invention possesses powerful self-evolution and continuous iteration capabilities. The system's core knowledge graph and game generator constitute a closed-loop data flow system. By continuously absorbing new training data and real-world cases, the system can constantly update its adversarial strategy library and causal knowledge base. This data-driven self-preservation mechanism ensures that the system can synchronize with the latest criminal methods and law enforcement standards in real time, maintaining the forward-looking nature and vitality of the training program, and providing core technological support for building a modern, intelligent smart policing training system. Attached Figure Description

[0024] Figure 1 Yes, this is a schematic diagram of the overall technical solution architecture; Figure 2 This is a schematic diagram of the core principle framework of the large model-driven game adversarial generator in this invention. Figure 3 This is a flowchart illustrating the main stages of cross-modal data parsing and spatiotemporal causal knowledge graph construction in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the trainee behavior capture and dual-track evaluation module in this invention; Figure 5 This is a schematic diagram of the core principle framework of the collaborative game logic of the multi-agent reinforcement learning unit in this invention; Figure 6 This is a flowchart outlining the main stages of the training data feedback and system evolution closed loop in this invention. Detailed Implementation

[0025] Example 1: Please refer to the appendix Figure 1 To be continued Figure 6To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0026] A smart policing simulation training system based on a large-scale artificial intelligence model includes a cross-modal data parsing device, a spatiotemporal causal knowledge graph construction device, a large-scale model-driven game adversarial generator, a trainee behavior capture device, and a dual-track evaluation and feedback module. The cross-modal data parsing device, the spatiotemporal causal knowledge graph construction device, the large-scale model-driven game adversarial generator, the trainee behavior capture device, and the dual-track evaluation and feedback module achieve physical data connection and logical interaction through a high-speed industrial bus or a high-bandwidth, low-latency local area network.

[0027] The cross-modal data parsing device is used for deep feature extraction and semantic association of multi-source heterogeneous police data. The device includes a text parsing unit, an audio processing unit, a video semantic deconstruction unit, and a spatial feature conversion unit. The text parsing unit, deployed on a high-performance server, is configured to receive and process massive amounts of historical police case files. It integrates a pre-trained natural language understanding model, using sub-logic such as word segmentation, named entity recognition, and dependency parsing to transform unstructured legal documents, records, and case reports into structured event sequences with clear temporal characteristics. The audio processing unit integrates a speech recognition engine and sentiment analysis operators, specifically designed to parse audio streams from audio and video data captured by law enforcement recorders. It not only converts speech into text commands with role identifiers but also adds sentiment labels to the text commands by calculating fundamental frequency, energy distribution, and speech rate fluctuations in acoustic features, quantifying the intensity of psychological confrontation during law enforcement. The video semantic deconstruction unit is equipped with a deep learning-based human pose estimation model for real-time tracking of human skeleton nodes in law enforcement recorder videos. By analyzing changes in skeleton displacement velocity and angle, it identifies potential violent resistance to law enforcement or actions involving the removal of dangerous items, and converts these into semantic tags with risk level annotations. The spatial feature conversion unit is connected to the urban geographic information system and is responsible for converting static geographic information data such as geographic coordinates, building layouts, and public facility distributions into spatial features of a digital twin environment with topological relationships, providing a high-precision physical foundation for subsequent virtual training environments.

[0028] The spatiotemporal causal knowledge graph construction device, connected to the cross-modal data parsing device, is used to construct a dynamically evolving police knowledge hub. The device includes a semantic mapping unit, a causal chain generation unit, a conflict detection unit, and a behavior probability modeling unit. The semantic mapping unit maps parsed heterogeneous data to a unified semantic space, encapsulating the relationships between basic police incident elements such as people, events, places, objects, and organizations, and introducing fine-grained timestamp tags to form dynamic knowledge expressions in the form of quadruplets or quintuples. The causal chain generation unit constructs cause-and-effect relationships between events by mining logical loops in historical cases, forming a graph foundation that supports logical deduction. The conflict detection unit, upon receiving new police incident data, if the new data conflicts with the logical chains in the existing knowledge graph, initiates conflict analysis logic. By comparing the credibility weights of multi-source intelligence, it calculates whether the conflict point belongs to a strategic deviation under special circumstances and dynamically updates the causal probability distribution in the graph. The behavior probability modeling unit is responsible for calculating the probability of behavior patterns within a specific geographical area within a preset time period based on historical data, revealing the non-obvious social operation rules, and transforming demographic data into a group behavior density function within a specific time period, dynamically adjusting the base number and interaction frequency of non-player characters in the training scenario.

[0029] The large-scale model-driven game-playing adversarial generator, connected to a spatiotemporal causal knowledge graph construction device, is used to generate emergent real-world adversarial scenarios. The generator includes a central generation hub, a multi-agent reinforcement learning unit, a collaborative game engine, and a perception state update unit. The central generation hub, based on causal logic in the knowledge graph, extracts key elements from the trainee's ability profile and calls upon a large-scale pre-trained language model to generate a highly logically consistent training background and non-player character personality settings, including the suspect's motives, the victim's psychological state, and the herd mentality of onlookers. The multi-agent reinforcement learning unit manages the behavioral decisions of the suspect, onlookers, and victim, with each agent possessing independent observation, decision-making, and action capabilities within the virtual space. The collaborative game engine is configured to coordinate the interaction between the suspect agent and the onlooker agents. It internally stores a herd mentality model; when the trainee takes coercive measures against the suspect, the engine drives the onlooker agents to exhibit specific levels of agitation or obstruction, simulating public opinion pressure and physical interference in a real police environment. Upon receiving a trainee's query for information about a specific suspect, the perception state update unit updates the perception state of the suspect's intelligent agent in real time, enabling it to autonomously find the optimal escape path within a specific range using geographical information and social relationship data in the knowledge graph, or to interfere with the trainee's subsequent judgment by creating false information.

[0030] The trainee behavior capture device is used to acquire and quantify multi-dimensional interaction data of trainees in real time throughout the training process. The device includes an instruction acquisition unit, a trajectory tracking unit, a speech semantic analysis unit, and a tactical action recognition unit. The instruction acquisition unit captures information retrieval keywords, resource scheduling requests, and police feedback text input by the trainee through a connection to a police terminal interface. The trajectory tracking unit uses virtual space positioning technology to record the trainee's patrol movement routes and search blind spots within the simulated scenario. The speech semantic analysis unit includes a voice sentiment analysis subunit, used to analyze the trainee's tone, speech rate, and keyword selection when conversing with a non-player character. By calculating emotional valence and arousal, it assesses the trainee's psychological stability and speech control under high-pressure police situations. The tactical action recognition unit captures the trainee's tactical action sequences when responding to emergencies through motion capture components or visual analysis, and converts them into standardized semantic feature vectors, providing fine-grained input for evaluation.

[0031] The dual-track evaluation and feedback module is connected to both the large-model-driven game adversarial generator and the trainee behavior capture device, and is used for comprehensive evaluation of the execution process and results. The dual-track evaluation and feedback module includes a Pareto optimal strategy evolution unit, a process mining and analysis unit, a cognitive computing unit, and a closed-loop feedback unit. The Pareto optimal strategy evolution unit uses a game theory model to generate multiple non-dominant strategy sets for the current scenario before training begins. These strategy sets represent the optimal balance point across multiple dimensions such as compliance, efficiency, and security. The process mining and analysis unit performs pattern matching between the trainee's real-time behavior vector and the Pareto optimal strategy sets, calculating the Euclidean distance of the trainee's actual decision path relative to the Pareto front in the strategy space, thereby quantifying the trainee's decision quality. The cognitive computing unit analyzes the trainee's logical connections when processing multi-source contradictory intelligence. If the trainee fails to identify maliciously tampered simulated intelligence within a preset time, the system will mark their deficiency in a specific cognitive dimension. The closed-loop feedback unit generates a personalized capability weakness profile based on the evaluation results and feeds back the correction parameters to the large model-driven game adversarial generator to dynamically adjust the complexity and adversarial intensity of the next training scenario.

[0032] Furthermore, the system also includes a data feedback device. This device is connected between the dual-track evaluation and feedback module and the spatiotemporal causal knowledge graph construction device, and is used to automatically anonymize the trainees' successful decision sequences and typical failure cases in each training session. The data feedback device integrates a logic extraction operator, which can transform these cases into logical nodes with specific weights and inject them back into the spatiotemporal causal knowledge graph construction device, enabling the system's self-evolution and ensuring that the training scenario is always synchronized with the latest trends in crime evolution.

[0033] In the specific implementation process, the video semantic deconstruction unit in the cross-modal data parsing device adopts a human keypoint extraction algorithm based on residual networks. After receiving the video frame stream from the law enforcement recorder, the unit first performs preprocessing logic, including noise reduction, contrast enhancement, and size normalization. Subsequently, the unit calculates the human body heatmap in each frame image through a deep convolutional neural network and determines the key point positions of the human skeleton based on the peak coordinates of the heatmap. The key point position information is input into a spatiotemporal graph convolutional network, and the behavioral intention of the human body is evaluated by calculating the rate of change of Euclidean distance and angular offset between adjacent key points. When the calculated behavioral intention matches the preset risk action template (such as weapon holding posture, attack preparation posture) more than a preset threshold, the unit immediately generates a risk label with a timestamp and pushes the label to the spatiotemporal causal knowledge graph construction device in real time.

[0034] In one specific embodiment, the video semantic deconstruction unit performs the following logic: Input: The input is a set of consecutive image frames. ,in Indicates the frame index. and These are the height and width of the image, respectively (e.g., ).

[0035] Preprocessing: Normalize the size of the input image to... (For example, ).

[0036] Heatmap generation: The preprocessed image is input into a deep convolutional neural network with an encoder-decoder structure. This network contains 5 downsampling convolutional layers (encoder) and 5 upsampling convolutional layers (decoder). The network output is... A heat map , Indicates the number of key points on the human body (e.g., Each heatmap The pixel value represents the presence of the first pixel at that location. The probability of each key point.

[0037] ; in, These are the coordinates of the downsampled feature map.

[0038] Key point coordinate extraction: By solving for the location of the maximum value in the heatmap, the coordinates of the first key point are determined. Normalized coordinates of key points in the image .

[0039] ; Graph structure construction: extracting Key points as shown in the diagram nodes .side Based on predefined human skeletal structures, such as head-shoulder, shoulder-elbow, and elbow-wrist connections, an adjacency matrix is ​​constructed. .

[0040] Spatiotemporal graph convolution: converts continuous... Frames (e.g.) coordinate sequence The input is fed into a spatiotemporal graph convolutional network. This network contains multiple spatiotemporal graph convolutional layers, each updating node features by aggregating information from spatial and temporal neighborhoods. After passing through a fully connected layer, the output is a behavior intent classification vector. ,in For the number of predefined risk action categories (e.g., (This includes carrying weapons, preparing to attack, and walking normally).

[0041] ; when The probability value of the corresponding weapon category If so, a risk label with a timestamp is generated.

[0042] The spatiotemporal causal knowledge graph construction device employs a distributed graph database-based storage architecture when processing massive amounts of heterogeneous data. The semantic mapping unit utilizes ontology modeling technology to define a core ontology set for the policing domain, including personnel, event, geographic location, and item ontologies. During the data access phase, this unit uses text matching and knowledge fusion algorithms to eliminate semantic ambiguity for the same entity across different data sources. The causal chain generation unit uses a correlation strength calculation logic based on point mutual information to analyze the co-occurrence probability of different police incident elements over time. If the co-occurrence frequency of two events within a preset time window is significantly higher than the random distribution probability and they have a clear temporal sequence relationship, then they are preliminarily determined to have a causal relationship. The conflict detection unit maintains a credibility evaluation matrix and assigns dynamic weights to data from different sources. For example, data from official law enforcement files is set to a first preset value, while data from demographic inferences is set to a second preset value, with the first preset value being greater than the second. When new data causes logical conflicts, this unit calculates the comprehensive reliability score on the conflict path, retains the one with the highest score, and corrects the damaged causal probability to ensure the logical rigor of the graph.

[0043] The causal chain generation unit calculates event pairs. The strength of causal association is quantified by the point mutual information value, and the specific formula is as follows: ; in, It is an event The probability of it appearing in the entire historical case database. It is an event The probability of occurrence It is an event After it occurs within the preset time window Events within (e.g., 10 minutes) The joint probability of occurrence. If (For example, Then establish a rule by point to The causal edges, with attached weights. .

[0044] The conflict detection unit maintains a credibility evaluation matrix. When a new data source During input, it is related to a certain causal edge. Credibility contribution value Calculated by the following formula: ; in, It is a data source The preset basic weights (e.g., 0.9 for law enforcement files and 0.6 for social statistics). This is the matching coefficient for the logical consistency between the current data and the causal edge (1 for a match, 0 for a conflict). Causal edge Overall credibility The updated formula is: ; in, It refers to the number of times it has been updated in history. This represents the number of data sources updated this time. If... Below the preset threshold (For example, 0.3), then the causal edge is marked as a state to be checked, and its probability distribution is as follows: It will be resampled.

[0045] The central generation hub in the large-scale model-driven game adversarial generator employs a prompting engineering-based guidance mechanism, transforming the trainee's historical ability score data into input prompts for the large-scale model. Based on the instructions in the prompts, the large-scale model retrieves relevant background elements from a spatiotemporal causal knowledge graph. The multi-agent reinforcement learning unit uses an asynchronous dominant actor-critic algorithm, where each non-player agent maintains an independent policy network and value network. The cooperative game engine guides collaboration between agents by introducing a global reward function, configured as an increasing function of the trainee's failure probability. For example, in a simulated street check scenario, the suspect agent is designed to escape, while the bystander agents are designed to create blind spots. Through shared environmental observation information, they dynamically and collaboratively generate a counter-surveillance strategy. When the trainee retrieves street surveillance footage through a police terminal, this query behavior, as a change in environmental state, is input into the suspect agent's policy network in real time, prompting it to immediately execute a preset path-switching action, utilizing complex alleyways in the geographic information system for evasion.

[0046] Each agent in the multi-agent reinforcement learning unit The asynchronous advantage actor-critic algorithm is used, and its state space... and action space The definition is as follows: state : Includes the agent's own perceived state (such as location) Orientation Blood volume ), Environmental map information and the feature vector of the trainee's (opponent's) most recent behavioral trajectory. .

[0047] action : This is a discrete action space, including {forward, backward, left turn, right turn, attack, hide, shout, make noise}.

[0048] Critics Network Used to assess the value of the current state, actor network Used to output the action probability distribution. The loss functions of the two networks are as follows: ; ; in, It is the first The actual cumulative reward for each step It is the network of critics' opinions on the state. Its predictive value.

[0049] ; in, It is the dominant function. It is policy entropy. It is the entropy regularization coefficient, used to encourage exploration.

[0050] The global reward function introduced by the collaborative game engine Defined as an increasing function of the probability of law enforcement failure by trainees, its expression is: ; in, It represents the probability that the suspect's AI agent will successfully escape. It is an effectiveness index of the interference caused by onlookers to the trainees. and It is the balancing weight coefficient.

[0051] The dialogue semantic analysis unit of the trainee behavior capture device utilizes an emotion classification algorithm combining Mel-frequency cepstral coefficients and support vector machines. After capturing the trainee's speech signal in real time, this unit extracts its feature vectors and maps them to a two-dimensional emotion coordinate space (i.e., emotion valence and arousal space). If the trainee's speech exhibits extremely high arousal and low valence, the system determines that the trainee is in a high-pressure or out-of-control state, and this evaluation result is immediately sent to the dual-track evaluation and feedback module. Simultaneously, the trajectory tracking unit evaluates the spatial search efficiency by calculating the overlap area between the trainee's coordinate trajectory in the virtual space and the theoretically optimal search path.

[0052] The extraction process of Mel frequency cepstral coefficients in the discourse semantic analysis unit is as follows: For the original speech signal After pre-emphasis, framing, and windowing, a fast Fourier transform is performed to obtain the spectrum.

[0053] Spectrum through One (e.g.) The Mel filter bank is used to calculate the logarithmic energy of each filter and obtain the Mel spectrum.

[0054] Performing a discrete cosine transform on the Mel spectrum yields the following results: Dimensions (e.g.) MFCC eigenvectors .

[0055] Continuous The MFCC feature vectors of the frames are concatenated into a feature sequence. The input is fed into a two-layer Long Short-Term Memory (LSTM) network. This LSTM network outputs a two-dimensional vector. Mapped to the emotional coordinate space: ; in, Indicates affective valence (positive-negative). Indicates arousal level (calm-excited). It is the hidden state of the last time step of the LSTM.

[0056] The dual-track assessment and feedback module calculates the emotional valence hedging index in real time. Defined as: ; in, and It is the emotional valence and arousal of the trainees. and These are the corresponding numerical values ​​for the suspect. The higher the value, the more stable the trainee's emotions are and the stronger their emotional control over the suspect.

[0057] When executing the evaluation logic, the dual-track evaluation and feedback module employs a Pareto optimal strategy evolution unit that uses a multi-objective evolutionary algorithm to find the optimal balance between efficiency and risk avoidance under the hard constraint of ensuring law enforcement compliance. The process mining and analysis unit uses Petri nets or heuristic mining models to detect the consistency of the trainee's action sequences. If a trainee fails to report information as required before a high-risk arrest, the evaluation unit will detect a significant deviation of the logical path from the Pareto front. The cognitive computing unit further analyzes this deviation to determine whether it is due to a lack of understanding of specific policing regulations or decision-making confusion under multi-source intelligence interference. Finally, based on these in-depth analyses, the closed-loop feedback unit adjusts the environmental parameters of the next generator round, such as increasing the proportion of interfering information or increasing the strategy complexity of non-player characters, forcing trainees to conduct targeted training in their weak areas.

[0058] The multi-objective evolutionary algorithm for the Pareto optimal policy evolutionary unit is as follows: Objective function: for any decision path Define three objective functions that need to be minimized: Risk function: ,in Is The probability of civilian casualties occurring at any time.

[0059] Time cost: This refers to the total time from receiving the alarm to completing the response.

[0060] Cost of violation: ,in These are the weights for different types of violations. It is a violation indicator function.

[0061] Pareto dominance: for two decision options and If satisfied Then it is called Pareto is superior All solutions that are not dominated by any other solution constitute the Pareto optimal policy set. .

[0062] Competency Gap Index: This index represents the trainee's actual decision-making path. Mapping to the target space yields a vector. Then the trainee's ability gap index. Defined as the point to the Pareto front. The minimum Euclidean distance: The data feedback device employs an incremental learning mechanism to evaluate the value of each training interaction trajectory. Only cases that correct causal relationships within the knowledge graph or possess typical tactical reference value are selected for the feedback queue. The desensitization processing logic generalizes or replaces sensitive entities to ensure privacy protection without disrupting the logical chain. Processed logical nodes are integrated into the graph through weighted aggregation, enabling the graph to perceive the latest criminal behavior patterns, such as new fraudulent tactics or methods of evading tracking.

[0063] Example 2: Based on Example 1, this example is designed with special features for the hardware deployment architecture and data flow protocol of the AI-based smart policing simulation training system, which is designed to meet the needs of large-scale city-level distributed training.

[0064] In this embodiment, a smart policing simulation training system based on a large artificial intelligence model adopts a distributed architecture of cloud-edge collaboration, including a cloud core processing cluster, multiple edge training base stations, and portable interactive terminals for trainees deployed at the edge.

[0065] The cloud-based core processing cluster houses the aforementioned spatiotemporal causal knowledge graph construction device and the central generation hub of a large-model-driven game-theoretic generator. Because the construction of the knowledge graph and the operation of large-scale language models require enormous computing resources (such as high-performance video memory and large-scale matrix operation units), the cloud cluster is configured to provide high-performance backend support. The cloud-based core processing cluster maintains communication with each edge training base station via an encrypted wide-area network channel, and is responsible for global script planning, evolutionary reasoning of complex causal logic, and feedback learning from the full training data.

[0066] The edge training base station is deployed within police training bases at various levels. It integrates components of the cross-modal data parsing device described in Example 1, a multi-agent reinforcement learning unit, a trainee behavior capture device, and a dual-track evaluation and feedback module. The edge training base station is configured to utilize low-latency local GPU computing power to perform latency-sensitive tasks, such as real-time skeleton extraction in video semantic deconstruction, speech emotion analysis, and real-time action decision-making for multi-agent NPCs. This distributed deployment effectively reduces the latency of the perception-action feedback loop in the virtual training environment, ensuring a highly immersive training experience.

[0067] The edge training base station includes an edge state synchronization unit. This unit is configured to maintain the consistency of the local virtual environment and, at a preset frequency, lightweight compresses key local training state data (such as the trainee's real-time location, key decision points, and NPC damage status) before synchronizing it to the cloud core processing cluster. Simultaneously, the cloud cluster sends policy update packages to the edge base station based on the latest evolution of the global graph, dynamically correcting the reward function parameters of the local agent.

[0068] At the data flow level, the cross-modal data parsing device of the edge training base station includes a real-time feature stream push unit. After performing preliminary feature extraction on the multimodal behavior of the local trainee, this real-time feature stream push unit no longer sends the original audio and video streams to the cloud, but instead sends a sequence of feature vectors that have undergone dimensionality reduction. This feature vector sequence not only contains the trainee's semantic instructions but also high-value information such as the energy characteristics of their actions and the probability distribution of their tactical intentions. The cloud-based core processing cluster uses these feature vectors to update the trainee's capability profile on a global scale.

[0069] Furthermore, the dual-track evaluation and feedback module within the edge training base station is divided into a real-time evaluation submodule and an offline deep evaluation submodule. The real-time evaluation submodule uses a lightweight decision tree or random forest model to provide second-level feedback on the trainee's tactical compliance. The offline deep evaluation submodule, after each training session, uploads all locally collected interaction logs to the cloud and uses the Pareto optimal policy evolution unit in the cloud to perform deep deviation analysis based on large-scale simulation, generating a more accurate personalized capability weakness profile.

[0070] The portable interactive terminal for trainees includes an augmented reality display, a bone conduction voice interaction component, and a handheld terminal integrating a police workflow simulator. This terminal connects to an edge training base station via Wi-Fi 6 or a 5G private network, and is responsible for presenting the generated combat scenarios to the trainees through visual, auditory, and other sensory channels. The terminal is equipped with a sensor fusion unit, capable of calibrating the trainee's positional mapping between physical and virtual spaces in real time, and capturing all query, scheduling, and data entry operations performed on the handheld device.

[0071] Through this cloud-edge collaborative architecture, this invention can support large-scale concurrent training while ensuring the complexity and emergence of scenarios. The cloud is responsible for the brain function, namely long-term knowledge accumulation and complex scenario deduction; the edge is responsible for the cerebellum function, namely agile interactive response and real-time behavior capture. This architecture greatly improves the system's scalability, enabling smart policing simulation training to cover a wider geographical area and more trainees.

[0072] Example 3: Based on the above examples, this example further optimizes the multi-agent game logic and environment modeling strategy of the system to meet the specific needs of smart policing in responding to mass emergencies or complex urban security environments.

[0073] In a smart policing simulation training system based on a large-scale artificial intelligence model, the spatiotemporal causal knowledge graph construction device specifically incorporates a sociopsychodynamic modeling unit. This unit is configured to extract common psychological characteristics of populations in specific regions from socio-demographic data, such as stress thresholds in high-density commercial areas and social cooperation tendencies in residential areas. These psychological parameters are quantified as latent variable nodes in the knowledge graph and associated with specific timestamps and spatial locations.

[0074] Correspondingly, the collaborative game engine in the large-scale model-driven game adversarial generator is configured to drive hundreds or thousands of collective intelligent agents with consistent behavioral logic. The collaborative game engine employs a path planning model based on the potential field method and a decision diffusion logic based on cellular automata. When trainees take inappropriate coercive measures in a virtual scenario, the parameters provided by the socio-psychological dynamics modeling unit trigger the emotional state transition of the collective intelligent agents, causing individual anger or panic to spread rapidly within their neighborhood, simulating the sudden evolution of collective emotion from silence to riot.

[0075] In this embodiment, the trainee behavior capture device adds a group control effectiveness evaluation unit. This unit uses complex spatial statistical algorithms to analyze the distribution of the deterrent force field formed by the trainee in the virtual space, as well as the offsetting effect of their rhetoric on the emotional index of the crowd. If the trainee's positioning, tone, or decision-making fails to effectively cut off the chain of emotional transmission, the system will determine that their evaluation score in the dimension of handling group events is low, and will record in detail the key time points and triggering actions of the emotional outburst.

[0076] The cognitive computing unit within the dual-track evaluation and feedback module incorporates multi-objective decision-making stress testing logic for such complex scenarios. During training, the system intentionally pushes multiple pieces of intelligence information—some true, some false, and some with conflicting timelines—through a virtual policing terminal. The cognitive computing unit quantifies the trainee's resilience and decision-making ability in an information fog environment by recording the trainee's response time to each piece of information, their processing priority, and the final logical judgment chain formed.

[0077] In this embodiment, the data feedback device functions as a group behavior pattern capture device. By clustering and analyzing the trajectories of a large number of trainees handling group incidents, the system can identify the optimal action combinations that can effectively de-escalate the situation and high-risk behaviors that are likely to escalate conflicts. These statistically validated behavioral patterns are transformed into logical edges with high confidence and reinforcing the spatiotemporal causal knowledge graph. This makes the system not only a training tool but also a platform for the research and evolution of policing tactics, enabling policing decision-making departments to discover and validate new handling strategies in complex law enforcement environments.

[0078] Furthermore, the system in this embodiment also includes a legal boundary constraint monitoring device. This device is connected to a large-model-driven game adversarial generator, and internally stores a complete code of law enforcement procedures and digital representations of the latest judicial interpretations. When generating adversarial strategies, this device acts as a legal reviewer, monitoring in real time whether the NPC agent's counter-surveillance strategies exceed the logical boundaries of the real world, and also monitoring in real time whether the trainee's actions cross legal boundaries. Once the trainee's behavior constitutes illegal law enforcement, the system will generate a red-line warning in the evaluation report and forcibly add relevant legal knowledge assessment scenarios in subsequent personalized closed-loop feedback.

[0079] Example 4: This example details the textual logic implementation of the Pareto optimal policy evolution unit in the dual-track evaluation and feedback module. This unit does not rely on simplified score calculations, but rather quantifies the trainee's comprehensive competence by constructing a high-dimensional policy space.

[0080] The Pareto optimal policy evolution unit is configured to execute the following logical flow: First, during the training scenario generation phase, the unit acquires all key environmental variables of the scenario, including the number of suspects, weapon possession status, surrounding civilian density, and the legal basis for law enforcement. Then, the unit initiates multiple parallel simulation instances, extensively sampling the policy space using a pre-defined heuristic search strategy. Each point in the policy space represents a complete decision scheme, which has a corresponding projection value on the evaluation dimension vector.

[0081] The evaluation dimension vector includes a safety dimension score, an efficiency dimension score, and a compliance dimension score. The safety dimension score is determined by a negative correlation function between civilian casualty rate, the probability of trainees being injured, and the probability of suspect escape. The efficiency dimension score is determined by the total time from receiving the report to completing the response and the redundancy of the allocated police resources. The compliance dimension score is determined by the degree of matching between the action sequence and law enforcement standards.

[0082] The Pareto optimal policy evolution unit selects the Pareto front set by comparing the dominance relationships between sampled points. Policy A is determined to be a Pareto optimal policy if and only if no policy is superior to policy A in all dimensions, and is strictly superior to policy A in at least one dimension. These policies constitute the ceiling for handling current extremely challenging environments.

[0083] After the trainee completes the exercise, the process mining and analysis unit transforms the trainee's actual decision path into a projection point on the same evaluation dimension vector. The dual-track evaluation and feedback module calculates the minimum Euclidean distance from this projection point to the Pareto front set. The Euclidean distance is defined as the trainee's capability gap index. If the distance value exceeds a first preset threshold, it indicates that the trainee's decision-making has room for improvement in multiple dimensions; if the distance value is less than a second preset threshold (the first preset threshold is greater than the second preset threshold), it indicates that their decision-making is close to the optimal solution in the current environment.

[0084] This module also performs component bias analysis. If a trainee's projection point is close to the Pareto front in the compliance dimension but significantly lags behind in the efficiency dimension, the system determines that the trainee has a cognitive tendency of excessive risk aversion leading to slow response. This fine-grained evaluation result will directly guide the large model-driven game adversarial generator to generate scenarios with greater time pressure in the next round, in order to deliberately train the trainee's decision-making decisiveness.

[0085] This Pareto optimality-based evaluation logic avoids the limitations of traditional evaluations that rely on a single point to determine victory or defeat. It acknowledges the objective reality of multiple optimal possibilities in police combat, providing trainees with more fair, scientific, and instructive feedback, and greatly improving the relevance and professionalism of the training.

[0086] Example 5: This example describes in detail the emotion annotation logic of the audio processing unit in the cross-modal data parsing device, and how it works in conjunction with the trainee behavior capture device to quantify the intensity of psychological confrontation.

[0087] The audio processing unit is configured to employ an emotion recognition method based on deep acoustic feature fusion. After acquiring the raw audio stream of the dialogue between the trainee and the suspect, the unit first performs silence cancellation and echo suppression processing. Subsequently, the unit performs frame segmentation, extracting the Mel-frequency cepstral coefficients, zero-crossing rate, and short-time energy features from each frame. These features are input into a Long Short-Term Memory (LSTM) network, which is configured to capture the dynamic evolution of the speech signal over time and identify various emotional states such as anger, fear, calmness, and hesitation.

[0088] The voice emotion analysis unit in the trainee behavior capture device simultaneously captures the trainee's voice features. The dual-track evaluation and feedback module receives these two sets of emotion data in real time and calculates the emotional valence hedging index. The emotional valence hedging index is defined as the ratio between the trainee's emotional stability and the suspect's emotional volatility. If, when faced with a suspect's high-decibel verbal attack, the trainee's voice features maintain low arousal and high stability, and the suspect's emotional index shows a downward trend under the trainee's verbal guidance, the system determines that the trainee's verbal control is excellent.

[0089] Conversely, if the trainee's voice signal exhibits obvious emotional lapses (such as a sudden increase in pitch or severe frequency fluctuations), and this triggers more aggressive confrontational behavior from the suspect (manifested as a shift in the suspect's node in the knowledge graph towards an extreme violent state), the system will record this critical mishandling error. This detailed data is not only used for the final evaluation score but is also incorporated into the data feedback device to train the NPC agent within the system to learn how to exploit the emotional weaknesses of law enforcement officers for psychological countermeasures.

[0090] This in-depth analysis of audio emotions elevates smart policing training from simple action drills to psychological game-theoretic exercises. It can effectively improve police officers' emotional regulation and communication skills under high-pressure environments, which is often of decisive significance in real law enforcement.

[0091] Example 6: This example aims to illustrate how the large model-driven game adversarial generator can utilize the emergent scenarios generated by the spatiotemporal causal knowledge graph to break through the limitations of traditional scripts.

[0092] In a simulated home search scenario, the central generation hub extracted historical features of the area from the knowledge graph: multiple drug-related cases had occurred there, and due to the complex building structure, wireless communication signals experienced periodic attenuation. Based on these factors, the large model generated a suspect NPC with a highly sophisticated counter-surveillance awareness.

[0093] When training begins, the trainee queries the property owner's information via a police terminal. This query, as a triggering event, is captured by the perception state update unit and mapped into the knowledge graph. The knowledge graph predicts through causal chains that, since the suspect NPC is set to have an inside contact in the property management office, the trainee's query behavior will be communicated to the suspect by the simulated inside contact. Therefore, the collaborative game engine immediately drives the suspect agent to destroy evidence and guides bystander agents (set as neighbors) to move large objects in the hallway, artificially creating physical obstacles to passage.

[0094] This scenario is not a pre-set script, but rather arises from the trainee's specific behavior triggering causal chains in the knowledge graph, which in turn drive multiple agents to autonomously collaborate based on their respective reward functions. This emergent scenario generation method ensures that the evolutionary path of each training session is unique.

[0095] After each training session, the dual-track evaluation and feedback module compares the trainee's response process with multiple optimal evolution paths pre-deduced by the system. If the trainee observes anomalies in the arrangement of clutter in the corridor, deduces in advance that the suspect is prepared, and decisively changes the assault plan to a perimeter blockade, this behavior, demonstrating outstanding practical wisdom, will be captured by the system and given a very high evaluation.

[0096] The data feedback device will de-identify and extract this successful handling case, and feed back the causal logic chain of abnormal debris in the corridor - vigilance - lockdown strategy to the knowledge graph with enhanced weights. As the number of training sessions increases, the knowledge graph will accumulate a large amount of such tactical knowledge derived from actual combat (simulation) but exceeding the script, thereby continuously improving the system's intelligence level.

[0097] Through the synergistic cooperation of the above embodiments, the present invention constructs a complete system capable of sensing multi-source data, automatically inferring causality, dynamically generating adversarial responses, accurately evaluating closed loops, and continuously self-evolving, thus completely solving the core technical problems of rigid training scenarios, single evaluation dimensions, and low data utilization in the prior art.

[0098] It should be understood that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. It should be noted that all hardware components and software modules mentioned in this specification can be physically implemented and logically constructed using commercially purchased servers, sensors, display terminals, and mature deep learning frameworks and database systems. The data interaction protocol between modules can follow industry standards, such as RESTful API, gRPC, or specific real-time data bus protocols. All algorithm parameters and threshold settings not detailed in this embodiment can be optimized and configured by those skilled in the art through a limited number of experiments according to the practical needs of specific police training, without departing from the technical essence of the present invention.

Claims

1. A smart policing simulation training system based on a large-scale artificial intelligence model, characterized in that, include: A cross-modal data parsing device is used to perform deep feature extraction and semantic association on multi-source heterogeneous police data; A spatiotemporal causal knowledge graph construction device is connected to the cross-modal data parsing device and is used to construct a dynamically evolving police knowledge hub. A large-model-driven game adversarial generator, connected to the spatiotemporal causal knowledge graph construction device, is used to generate emergent real-world adversarial scenarios. Trainee behavior capture device, used to acquire and quantify trainees’ multidimensional interaction data in real time throughout the training process; The system also includes a dual-track evaluation and feedback module, which is connected to the large model-driven game adversarial generator and the trainee behavior capture device, respectively. This module is used for a comprehensive evaluation of the process and results, and feeds back the correction parameters generated by the evaluation results to the large model-driven game adversarial generator to dynamically adjust the complexity and adversarial intensity of subsequent training scenarios.

2. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 1, characterized in that, The cross-modal data parsing device includes: a text parsing unit, used to receive historical police case files and convert unstructured text content into a structured event sequence with clear temporal characteristics through natural language understanding logic; The audio processing unit is used to analyze the audio stream captured by the law enforcement recorder and convert speech into text instructions with role identification and emotion labeling by identifying the fundamental frequency, energy distribution and speech rate fluctuations in acoustic features. The video semantic deconstruction unit is used to track human skeleton nodes in law enforcement recorder videos in real time. By analyzing the changes in skeleton displacement speed and angle, it identifies potential violent resistance to law enforcement or dangerous item extraction actions and converts them into semantic tags with risk level labels. The spatial feature transformation unit is used to transform geographic coordinates, building layouts, and public facility distribution data from a geographic information system into spatial features of a digital twin environment with topological relationships.

3. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 2, characterized in that, The spatiotemporal causal knowledge graph construction device includes: a semantic mapping unit, used to map the parsed heterogeneous data to a unified semantic space, encapsulate the relationship between basic police elements such as people, events, places, things, and organizations, and introduce fine-grained timestamp tags to form dynamic knowledge expression; The causal chain generation unit is used to construct the cause-and-effect relationship between events by mining the logical loops in historical cases, forming a graph base that supports logical deduction. The conflict detection unit is used to activate the conflict analysis logic when new input alarm data is received. By comparing the credibility weights of multi-source intelligence, it calculates whether the conflict point belongs to a strategic deviation under special circumstances and updates the causal probability distribution in the graph. The behavior probability modeling unit is used to calculate the probability of behavior patterns within a specific geographical area within a preset time period based on historical data, and to transform demographic data into a group behavior density function within a specific time period in order to adjust the base number and interaction frequency of non-player characters.

4. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 3, characterized in that, The large model-driven game adversarial generator includes: a central generation hub, which is used to extract key elements based on the causal logic in the spatiotemporal causal knowledge graph and the trainee's ability profile, and call the pre-trained large model to generate a training background with logical self-consistency and the personality settings of non-player characters. Multi-agent reinforcement learning units are used to take over the behavioral decisions of non-player characters, enabling each agent to have independent observation, decision-making and action capabilities in the virtual space; A collaborative game engine is used to guide cooperation between agents by introducing a global reward function and drive the crowd agents to produce agitated or obstructive behaviors based on a herd mentality model, simulating physical interference in real police environments. The perception state update unit is used to update the perception state of the suspect agent in real time after receiving the trainee's query for suspect information, so that the suspect agent can use the data in the spatiotemporal causal knowledge graph to autonomously find an escape path or create false interference information.

5. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 4, characterized in that, The trainee behavior capture device includes: an instruction acquisition unit, used to capture information retrieval keywords, resource scheduling requests and police feedback text input by the trainee through a connection to a police terminal interface; The trajectory tracking unit is used to record the patrol movement routes and search blind spots of trainees in simulated scenarios using virtual space positioning technology; The speech semantics analysis unit is used to analyze the tone, speech rate and keyword selection of trainees when they talk to non-player characters. By calculating emotional valence and arousal, it assesses the trainees' psychological stability and speech control. The tactical action recognition unit is used to capture the trainee's tactical action sequence through motion capture or visual analysis and convert it into a standardized semantic feature vector.

6. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 5, characterized in that, The dual-track evaluation and feedback module includes a Pareto optimal strategy evolution unit, which uses a game theory model to perform multi-objective sampling on the current scenario in terms of security, efficiency and compliance dimensions before training starts, and selects a set of Pareto optimal strategies that do not dominate each other. The process mining and analysis unit is used to transform the trainee's real-time behavior vector into a projection point on the evaluation dimension vector, and to calculate the minimum Euclidean distance from the projection point to the Pareto optimal policy set, wherein the Euclidean distance is defined as the trainee's ability gap index. The cognitive computing unit is used to analyze the logical connection process of trainees when processing multi-source contradictory intelligence and to identify redundancy or deficiency in the trainees' cognitive dimensions. The closed-loop feedback unit is used to generate a personalized capability gap profile based on the evaluation results and feed the parameters back to the game adversarial generator driven by the large model to adjust the scene parameters.

7. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 6, characterized in that, Also includes: The data feedback device, connected between the dual-track evaluation and feedback module and the spatiotemporal causal knowledge graph construction device, is used to automatically desensitize the successful decision sequences and failure cases of trainees during training, and use logical extraction operators to transform them into weighted logical nodes, which are then injected back into the spatiotemporal causal knowledge graph construction device to realize the evolution of causal knowledge in the system.

8. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 7, characterized in that, The video semantic deconstruction unit is used to perform the following logic: after receiving the video frame stream from the law enforcement recorder, it performs noise reduction, contrast enhancement and size normalization preprocessing; The human body heatmap in each frame of the image is calculated by a deep convolutional neural network, and the key points of the human skeleton are determined based on the peak coordinates of the heatmap. The key point location information is input into the spatiotemporal graph convolutional network, and the behavioral intention of the human body is evaluated by calculating the rate of change of Euclidean distance and angular offset between adjacent key points. When the degree of matching between the behavioral intention and the preset risk action template exceeds a preset threshold, a risk label with a timestamp is generated and pushed to the spatiotemporal causal knowledge graph construction device.

9. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 8, characterized in that, The speech semantic analysis unit is used to perform the following logic: after capturing the trainee's speech signal in real time, extract the Mel frequency cepstral coefficient feature vector; The feature vectors are mapped to a two-dimensional emotional coordinate space that includes the emotional valence dimension and the arousal dimension; The dual-track assessment and feedback module calculates the emotional valence hedging index in real time. The emotional valence hedging index is defined as the ratio between the trainee's emotional stability value and the suspect's emotional volatility value. If the trainee's voice characteristics maintain low arousal and high stability, and the suspect's emotional index shows a downward trend under the trainee's verbal guidance, then the trainee's verbal control is judged to be at the preset excellent level.

10. The intelligent policing simulation training system based on a large artificial intelligence model according to claim 9, characterized in that, The system adopts a distributed architecture that combines cloud and edge, including: a cloud core processing cluster, which is equipped with the spatiotemporal causal knowledge graph construction device and the central generation hub in the large model-driven game adversarial generator, responsible for global script planning, complex causal logic evolution and full data feedback learning; Multiple edge training base stations are deployed at the training site and integrate the cross-modal data parsing device, the multi-agent reinforcement learning unit, the trainee behavior capture device, and the dual-track evaluation and feedback module. They are responsible for performing low-latency video semantic deconstruction, voice emotion analysis, and real-time action decision-making for multi-agents. The trainee's portable interactive terminal, connected to the edge training base station, includes an augmented reality display, a bone conduction voice interaction component, and a handheld terminal integrating a police workflow simulator, which is responsible for presenting the generated combat scenarios to the trainee and capturing their operational behavior.

Images