Intelligent monitoring and evaluation system and method for prison night duty performance

Abstract

The invention discloses an intelligent monitoring and evaluation system and method for prison night duty performance. The system comprises a multi-source behavior perception unit, an edge real-time analysis unit and a cloud risk cognition and evolution unit. Non-blind area sensing is realized through UWB and a multi-camera network, and the edge unit carries out real-time abnormity identification and alarm; the cloud end cooperates with a multi-LLM agent through a performance knowledge graph, carries out deep correlation analysis on the event, actively discovers a cross-space-time potential composite risk mode, and generates a comprehensive assessment report; and the system can dynamically update the knowledge base and the detection rule by absorbing new knowledge and analyzing a natural language instruction, so that self-evolution is realized. According to the invention, the problems of passive monitoring, subjective evaluation and system stiffness in the prior art are solved, and the initiative, intelligent level and continuous adaptive capacity of prison night supervision are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and prison information management technology, and in particular to an intelligent monitoring and evaluation system and method for nighttime duty performance in prisons. Background Technology

[0002] Currently, prisons primarily rely on manual, timed patrols, fixed-point video surveillance, and basic electronic patrol systems to monitor the performance of nighttime police officers, patrol personnel, and on-duty personnel. These methods face significant limitations in practical application: fixed cameras have blind spots, failing to achieve uninterrupted coverage of the entire prison area; the detection of abnormal behavior heavily depends on monitoring personnel constantly watching screens, which can easily lead to missed detections due to fatigue, and there is a significant delay between detection and manual response, making it difficult to handle sudden emergencies. The evaluation of performance quality is largely based on manual records and subjective experience, lacking objective and quantitative data support, resulting in assessments becoming merely a formality. Furthermore, existing systems are functionally limited and have rigid rules, typically only able to issue alerts for simple rules such as "excessive absence from duty," unable to identify complex, multi-temporal and spatially correlated potential risk patterns (such as the correlation between repeated short absences and abnormal gatherings of people in specific areas), and lack the ability to self-optimize and update according to emerging risk situations or management requirements, remaining generally passive, lagging, and static. Summary of the Invention

[0003] To address the aforementioned problems, this invention proposes an intelligent monitoring and evaluation system and method for nighttime duty performance in prisons.

[0004] The main contents of this invention include: A smart monitoring and evaluation system for nighttime duty performance in prisons, comprising: The multi-source behavior perception unit is used to acquire the location coordinates, video images and patrol path data of the personnel performing their duties in real time. The edge real-time analysis unit is used to fuse and process the data from the multi-source behavior perception unit and identify preliminary abnormal events based on preset rules; The cloud-based risk cognition and evolution unit, connected to the edge real-time analysis unit, includes a duty performance knowledge graph module, a multi-agent collaborative judgment module, and a system adaptive optimization module. The duty performance knowledge graph module stores and associates duty performance standards, risk cases, and behavioral pattern knowledge. The multi-agent collaborative judgment module, composed of multiple large language model (LLM) agents, performs collaborative reasoning on preliminary abnormal events and historical correlation data to discover potential risk patterns exceeding preset rules. These potential risk patterns include complex risk scenarios formed by the correlation of multiple isolated events under specific temporal, spatial, or logical conditions. The system adaptive optimization module dynamically updates the content of the duty performance knowledge graph module and the identification rules of the edge real-time analysis unit based on the potential risk patterns output by the multi-agent collaborative judgment module and the corresponding handling feedback.

[0005] Furthermore, the multi-agent collaborative judgment module includes at least a first agent and a second agent; the first agent is used to judge the procedural compliance of behavior based on rules; the second agent is used to analyze the contextual relationships of events in conjunction with the duty performance knowledge graph module to determine whether they are potential signs of systemic risk. It also includes a third agent, which performs weighted fusion decision-making on the outputs of the first and second agents to generate the final risk judgment conclusion.

[0006] Furthermore, the system's adaptive optimization module includes a knowledge graph maintenance unit and a rule dynamic update unit. The knowledge graph maintenance unit is configured to extract structured knowledge from newly added regulations, accident reports, and handling records to update the knowledge graph. The rule dynamic update unit is configured to receive rule modification instructions through a natural language interaction interface, and after semantic parsing and testing, deploy the verified detection logic to the edge real-time analysis unit.

[0007] Furthermore, the multi-source behavior perception unit includes a UWB positioning module, a multi-camera monitoring network, and an electronic patrol module. The edge real-time analysis unit is deployed within the local edge computing gateway in the prison area to run computer vision target recognition algorithms, identify specific abnormal behaviors, and spatiotemporally correlate the recognition results with UWB positioning coordinates.

[0008] Correspondingly, the present invention also provides a method for intelligent monitoring and evaluation of nighttime duty performance in prisons, including: a data synchronization and acquisition step; a real-time edge processing step; a cloud-based collaborative analysis step; an early warning and report generation step; and a system adaptive optimization step.

[0009] Compared with existing technologies, the beneficial effects of the intelligent monitoring and evaluation system and method for nighttime duty performance in prisons proposed in this invention are as follows: (1) Through the collaboration of UWB high-precision positioning and multi-camera network, continuous and blind-spot-free perception of the entire prison area and personnel behavior is realized, providing a reliable data foundation for accurate judgment; (2) By deploying a multi-agent collaborative judgment module based on the duty knowledge graph in the cloud, the system can perform in-depth correlation analysis and reasoning on the preliminary anomalies reported from the edge, thereby discovering potential complex risk scenarios composed of multiple simple events, and realizing the leap from "passive alarm" to "active discovery".

[0010] (3) Through the system's adaptive optimization module, the performance knowledge graph and edge recognition rules are dynamically and continuously updated. This enables the system to absorb new management norms and handling experience, quickly adapt to changing risk patterns and management requirements, and possess the ability to self-evolve and continuously improve.

[0011] (4) By using the edge real-time analysis unit to perform local fusion and processing of multi-source data, the dependence on cloud bandwidth is reduced, ensuring real-time response. The hierarchical early warning mechanism provides on-site alerts for clear emergency events, improving emergency response efficiency. Meanwhile, potential risks requiring in-depth analysis are uploaded to the cloud, achieving optimized allocation of computing power. Attached Figure Description

[0012] Figure 1 This is a functional block diagram of the monitoring and evaluation system of the present invention; Figure 2 This is a flowchart of the monitoring and evaluation method of the present invention. Detailed Implementation

[0013] The technical solution protected by this invention will be described in detail below with reference to the accompanying drawings.

[0014] This invention provides an intelligent monitoring and evaluation system for nighttime duty performance in prisons, the hardware and software architecture of which are as follows: Figure 1 As shown, the system consists of a multi-source behavior perception unit deployed on-site in the prison area, an edge real-time analysis unit, and a cloud-based risk cognition and evolution unit deployed in a remote data center. The units interact with each other through a dedicated prison communication network.

[0015] The multi-source behavior perception unit, serving as the data input terminal of this system, comprises three types of heterogeneous sensing devices used to synchronously collect multimodal behavior data of personnel performing their duties. Specifically, the multi-source behavior perception unit includes a UWB positioning module, a multi-camera monitoring network, and an electronic patrol module.

[0016] More specifically, a positioning network is formed by deploying at least four UWB positioning base stations on the walls or ceilings of key areas such as prison corridors, passageways, and halls. Each personnel required to be monitored (such as on-duty police officers) wears a UWB tag with a unique ID. The base station measures the time of flight of the radio signal between itself and the tag, uses a time difference of arrival positioning algorithm to calculate and output the tag's three-dimensional spatial coordinates in real time, with positioning accuracy at the centimeter level. The coordinate data is transmitted via a wireless local area network at a specific frequency (such as 10Hz).

[0017] The multi-camera monitoring network can be composed of two types of cameras. For example, dome cameras supporting 360° horizontal and 90° vertical rotation can be deployed in large open areas (such as the central lobby of a prison area), while high-definition infrared cameras with fixed viewing angles can be deployed at key locations such as fixed entrances / exits and corners. Through topology optimization of the deployment locations of the two types of cameras, their fields of view are ensured to be interconnected, eliminating blind spots caused by physical obstructions. All cameras are connected via Ethernet and continuously output H.264 or H.265 encoded video streams.

[0018] The electronic patrol module is an enhanced version of the traditional electronic patrol wand, integrating a UWB positioning chip and a communication module. When personnel carrying the device arrive at a preset physical patrol point (such as a sensor installed on a wall) and perform a check-in operation, the device not only records the event timestamp but also simultaneously obtains the precise absolute location coordinates through its built-in UWB chip. This "time-location" data is uploaded via LoRa or Wi-Fi network.

[0019] The edge real-time analysis unit is used to fuse the data from the multi-source behavior perception unit and identify preliminary abnormal events based on preset rules. In one embodiment, the edge real-time analysis unit uses an industrial-grade edge computing gateway deployed in the prison area's computer room or duty room as a carrier. The gateway has a built-in GPU computing card and runs a lightweight containerized software stack. The data bus service built into this gateway simultaneously subscribes to personnel coordinate streams from the UWB positioning server, real-time video streams from the network video recorder, and event messages from the electronic patrol module. By stamping all data with a unified timestamp based on the network time protocol and unifying the spatial coordinate system, the spatiotemporal alignment of multi-source data is achieved. The edge real-time analysis unit also includes a computer vision analysis engine, which runs a specially trained YOLOv5 deep learning model. The model is trained using a large number of labeled prison scene images and video clips. Its output layer is designed to simultaneously detect and classify five preset abnormal behaviors: "personnel leaving their posts", "running fast", "accidental fall", "illegal climbing", and "gathering of people in specific areas". After decoding and sampling the input video stream, the engine feeds it into the model for real-time inference and outputs behavior category labels with confidence and corresponding image bounding box coordinates.

[0020] Furthermore, this gateway also incorporates a lightweight rules engine, which loads basic detection rules distributed from the cloud (e.g., "If no personnel are detected in area A01 within 30 seconds, trigger the 'area vacancy' event"). The rules engine receives behavior tags from the visual analysis engine and real-time coordinates from UWB. When abnormal behavior is detected, the engine immediately queries the UWB coordinates of the person at the same time, binding "behavior type - location - time" to generate a structured basic abnormal event. For example, the generated event might be: {"event_id": "E001", "type": "fall", "location": [x1, y1, z1], "timestamp": "2023-10-27 02:15:30.123", "confidence": 0.92}. Simultaneously, the rules engine compares the electronic patrol data with a pre-defined standardized patrol route map, calculating path deviation distance and time delay.

[0021] For primary events deemed high-urgency by the rule engine (such as "falling" with a confidence level > 0.9), or serious patrol deviations, the edge real-time analysis unit directly drives the audible and visual alarms deployed in the monitoring area to issue local alarms. All generated primary events, patrol compliance status, and associated video clip metadata are cached and uploaded in batches to the cloud unit.

[0022] The cloud-based risk perception and evolution unit is connected to the edge real-time analysis unit. This unit is deployed on a cloud server cluster and is the intelligent processing core of the system, consisting of three functional modules.

[0023] (1) Performance Knowledge Graph Module: This module is constructed and stored using the Neo4j graph database. The schema layer of the knowledge graph is defined by domain experts, and the core entity types include "regulatory system clauses", "historical risk cases", "abnormal behavior patterns", "physical areas", and "job roles". The relationships between entities include "belonging to", "occurring in", "easily leading to", and "mutually dependent". Instance data of the graph is injected in two ways: first, in batches during initialization, structured rules and regulations documents and historical event reports are imported; second, through continuous incremental updates via the system's adaptive optimization module. This module provides an efficient graph query interface for multi-agent modules to call.

[0024] (2) Multi-agent collaborative judgment module: This module deploys three agent services based on a large language model (e.g., using open-source LLaMA 3 or similar structure models for domain fine-tuning), each with its own function: Specifically, the three agents include a compliance verification agent, a context association analysis agent, and a decision fusion agent.

[0025] The compliance verification agent is fine-tuned to be proficient in explicit prison management regulations. Its input is a description of a primary event reported from the edge, and its task is to determine whether the event constitutes a clear violation based on the regulations in the knowledge graph (e.g., "failure to report location every 30 minutes as required during a nighttime solo patrol"). Its output is the compliance judgment and the cited regulations.

[0026] The contextual analysis agent is fine-tuned to possess powerful causal and relational reasoning capabilities. Its input includes not only the current event but also sequences of other events involving the same area and related individuals over a past period (e.g., the past hour), as well as environmental context queried from the knowledge graph module (e.g., "a conflict occurred in this area last month"). Its task is to analyze the spatiotemporal proximity and sequence patterns between events, inferring whether hidden causal chains or risk trends exist behind isolated events (e.g., a possible correlation between "two brief absences of police officers in Area A" and "abnormal activity of detainees in Area B during the same period"). Its output consists of potential risk association hypotheses and their confidence levels.

[0027] The decision fusion agent receives the outputs of the first two agents as input and performs decision fusion using a learnable weighted average algorithm. The weighted average decision algorithm fuses the detection results of each agent to generate the final risk score, as shown in the formula: Where n is the number of decision models. A similarity score for a single model.

[0028] The initial weights can be set as the reciprocal of each agent's accuracy on the historical validation set. The agent calculates the final risk score and determines whether to generate a "potential risk pattern" alert. For example, if the compliance check agent judges it as a "minor violation," but the context association analysis agent infers with high confidence that it belongs to a risk pattern, the decision fusion agent may comprehensively judge it as "medium risk" and trigger an alert.

[0029] (3) System Adaptive Optimization Module: This module ensures the sustainable evolution of system capabilities. Specifically, the system adaptive optimization module includes a knowledge graph maintenance unit and a rule dynamic update unit. The knowledge graph maintenance unit is configured to extract structured knowledge from newly added regulations, accident reports, and handling records, and integrate it with existing knowledge in the duty performance knowledge graph module to update the knowledge graph. For example, the knowledge graph maintenance unit periodically scans newly accessed PDFs of regulations and historical inspection record texts recognized by OCR. It uses a pre-trained natural language processing model (such as BERT) to perform named entity recognition and relation extraction, automatically extracting new "entity-relationship-entity" triples. Subsequently, through a conflict detection algorithm based on graph embedding similarity, the new triples are compared and integrated with existing knowledge in the knowledge graph. After administrator review or automatic confirmation, they are written into the graph database.

[0030] The rule dynamic update unit provides a natural language interactive interface. Administrators can input commands such as, "If anyone is found loitering near the northeast corner storage room for more than 2 minutes between 1 AM and 3 AM, please mark it as 'suspicious behavior' and notify the shift leader." This instruction is first converted into a structured logical expression by a syntax parser, and then validated in a simulated test environment containing various historical scenarios to evaluate its false positive and false negative rates. After successful validation, the rule is compiled into an executable script (such as a JSON-formatted rule description file) for the edge rule engine and distributed to the edge computing gateway via a secure channel to dynamically update its local rule base.

[0031] This invention also proposes a method for intelligent monitoring and evaluation of nighttime duty performance in prisons, such as... Figure 2 As shown, it includes: S1: Data Synchronization Acquisition Steps. Through the UWB positioning system, surveillance camera network, and electronic patrol equipment, the location coordinates, behavioral video images, and patrol route point data of personnel on duty are synchronously and in real time collected. Specifically, through a network of UWB positioning base stations deployed in key locations within the prison area, wireless signals emitted by UWB tags worn by personnel on duty are continuously received. The base stations use a time difference of arrival algorithm to calculate the precise three-dimensional coordinates of each tag in real time, forming a personnel location data stream, thus achieving the acquisition of positioning data. Through a monitoring network composed of dome cameras and fixed cameras, real-time video streams of various areas within the prison area are collected at a preset frame rate (e.g., 25 frames / second). The time of all cameras is synchronized with the edge server to achieve video data acquisition. When personnel on duty use electronic patrol equipment integrated with UWB modules to clock in at physical patrol points, the equipment simultaneously records a precise timestamp and the absolute location coordinates calculated by the built-in UWB chip, generating a "time-location" data pair, thus achieving patrol data acquisition.

[0032] S2: Real-time edge processing steps. On the local edge computing node in the prison area, the collected multi-source data is fused and processed, the target recognition algorithm is used to identify preliminary abnormal behavior and bind it with the location information, the compliance of the patrol path is checked, and local alarms are issued for emergency events.

[0033] Specifically, firstly, within the edge computing gateway, the data bus service receives and buffers data streams from the three sources mentioned above. Each data packet is then timestamped at the microsecond level based on NTP, and all location coordinates are uniformly transformed to the same spatial coordinate system, achieving precise temporal and spatial alignment of the multimodal data.

[0034] The aligned video stream is then fed into a computer vision analytics engine deployed on an edge gateway. This engine loads a YOLOv5 deep learning model trained on prison scene data, decodes and preprocesses the video frames, and performs real-time inference. The model output includes bounding boxes, category labels, and confidence scores for five preset abnormal behaviors (leaving the post, running, falling, climbing, and gathering).

[0035] Next, the rule engine within the edge gateway loads the initial detection rules downloaded from the cloud. The rule engine monitors the output of the visual analysis engine and the UWB location stream in real time. When abnormal behavior is detected, the engine immediately queries the UWB coordinates of the target person at the same timestamp, encapsulating the "behavior type-location-time-confidence" into a structured initial abnormal event object. Simultaneously, the rule engine compares the received electronic patrol "time-location" data with the preset standardized patrol route, calculating the Euclidean distance and time difference between the actual and planned path points to determine the compliance of the patrol.

[0036] Finally, for events that the rules engine determines meet the "emergency alarm" criteria (e.g., a fall with a confidence level higher than 0.9), the edge unit immediately triggers the locally deployed audible and visual alarms via its I / O interface. All generated primary abnormal events, patrol compliance judgment results, and corresponding video clip indexes (or keyframe features) are combined into a single data packet and uploaded to the cloud processing unit via an encrypted network channel.

[0037] S3: Cloud-based Collaborative Analysis Step. Preliminary abnormal events and related data are uploaded to the cloud. In the cloud, multiple LLM agents perform collaborative reasoning based on a pre-built duty performance knowledge graph to analyze deep relationships between events and identify potential complex risk patterns that go beyond simple rules and span across time and space.

[0038] Specifically, the cloud server receives data packets from multiple edge gateways, parses and stores them. The duty performance knowledge graph module provides a query interface. For each primary anomaly, the system queries the knowledge graph module based on dimensions such as the time, location, and personnel involved, retrieving relevant historical cases, policy provisions, regional risk levels, and other contextual information. Three LLM agents deployed in the cloud are triggered in parallel or sequentially to perform in-depth analysis of the events. The first intelligent agent takes the event description and relevant institutional provisions retrieved from the knowledge graph as input, runs its finely tuned language model, and outputs a conclusion on whether the event violates the explicit provisions and which provisions are violated, thus performing a compliance check. The second agent takes the current event, a sequence of recent related events (such as events in the same area within the previous hour), and environmental and historical risk information retrieved from a knowledge graph as input, and runs its reasoning model. This model analyzes the spatiotemporal correlations and sequence patterns between events, attempts to construct causal hypotheses, assesses whether they point to a systemic risk that has not yet materialized, and outputs correlation conclusions and confidence levels, thus achieving contextual correlation analysis. The third agent receives the outputs of the first two agents and performs decision fusion. It uses a dynamic weighted fusion algorithm (e.g., the weights are dynamically adjusted based on the accuracy of each agent's judgments on similar events over the past week) to calculate a comprehensive risk score. Based on a preset risk level threshold, it ultimately determines whether the event constitutes part of a "potential risk pattern" and determines its risk level (e.g., low, medium, high).

[0039] For events identified as potential risk patterns, the system will tag them with the corresponding pattern in their metadata and record their relationships with other events to form a risk map.

[0040] S4: Early Warning and Report Generation Steps. Based on the type and level of potential risk patterns, tiered early warnings are issued. The warning level is dynamically determined by combining rules such as event attributes, whether it constitutes part of a risk pattern, and regional risk levels. Simultaneously, a quantitative assessment report is generated to analyze risk trends.

[0041] Specifically, based on the risk level and risk pattern label output from step S3, a tiered response is executed. The early warning rule base defines the actions corresponding to each risk level, for example: Low-level risks: Only recorded in the system log and displayed as information on the management interface.

[0042] Intermediate risk: Automatically generate early warning messages and push them to the mobile terminals of relevant personnel or the computer screens in the duty room.

[0043] High-risk: Immediately triggers a remote audible and visual alarm, and notifies the on-duty leader and emergency response team via telephone, SMS, and other means.

[0044] Generate comprehensive risk reports. The report generation module is activated periodically (e.g., per shift, daily) or as needed. It aggregates all event data, risk assessment results, and early warning response records from the database within a specified time period. Utilizing templates and natural language generation technology, it automatically generates structured performance evaluation reports. Report content includes, but is not limited to: patrol plan completion rate, distribution of the number and type of abnormal events, average response time, a list of identified potential risk patterns, risk trend analysis charts, and targeted management recommendations.

[0045] S5: System Adaptive Optimization Steps. Based on newly discovered potential risk patterns and their handling feedback, the content of the cloud-based knowledge graph is dynamically updated, and the optimized detection rules are deployed to the edge by parsing natural language instructions.

[0046] Specifically, the system's adaptive optimization steps include a knowledge update sub-step and a rule update sub-step. The knowledge update sub-step involves extracting structured knowledge from newly added regulations and handling records to update the performance knowledge graph. By monitoring a designated data source directory, the processing flow is automatically initiated when new regulations documents (PDF / Word) or reviewed and completed event handling reports are stored. First, OCR and text parsing tools are used to extract text content. Then, a pre-trained NLP model (such as a BERT-based sequence labeling model) is called to perform named entity recognition and relation extraction, extracting new entities (such as new clauses, new equipment, and new personnel types) and relations (such as "a certain clause applies to a certain region"). Next, a graph similarity algorithm is used to determine whether the new knowledge conflicts with or is redundant with existing knowledge. After administrator confirmation or automatic rule verification, the new knowledge is incrementally written into the graph database.

[0047] The rule update sub-step includes generating or modifying detection rules by parsing natural language instructions, testing and verifying them, and then deploying them to the edge real-time processing step. Through a web interface or API, administrators can describe new detection requirements in natural language (e.g., "If any person is found lingering in restricted area C for an extended period, and abnormal sounds are simultaneously detected, raise the alarm level"). After sending this instruction to the cloud, a semantic parser first converts it into a formal logical expression. Then, this expression is verified on a test platform containing a large amount of historical data and simulation scenarios to evaluate its triggering conditions and effects. Once verified, the new rule is converted into a format recognizable by the edge rule engine (such as JSON or a specific DSL script) and distributed to all relevant edge computing gateways through a secure update channel. Upon receiving the new rule, the gateway dynamically updates its local rule base, taking effect without requiring a service restart.

[0048] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A smart monitoring and evaluation system for nighttime duty performance in prisons, characterized in that, include: The multi-source behavior perception unit is used to acquire the location coordinates, video images and patrol path data of the personnel performing their duties in real time. The edge real-time analysis unit is used to fuse and process the data from the multi-source behavior perception unit and identify preliminary abnormal events based on preset rules. A cloud-based risk perception and evolution unit, connected to the edge real-time analysis unit, includes: The job performance knowledge graph module is used to store and associate job performance norms, risk cases, and behavioral patterns. The multi-agent collaborative judgment module consists of multiple large-scale language model agents, which are used to perform collaborative reasoning on the preliminary abnormal events and historical related data to discover potential risk patterns that exceed preset rules. The potential risk patterns include composite risk scenarios formed by the association of multiple isolated events in a specific time, space or logic. The system adaptive optimization module is used to dynamically update the content of the duty performance knowledge graph module and the identification rules of the edge real-time analysis unit based on the potential risk patterns and corresponding handling feedback output by the multi-agent collaborative judgment module.

2. The intelligent monitoring and evaluation system for nighttime duty performance in prisons according to claim 1, characterized in that, The multi-agent collaborative judgment module includes at least a first agent and a second agent; the first agent is used to judge the procedural compliance of behavior based on rules; the second agent is used to analyze the contextual relationship of events in conjunction with the duty performance knowledge graph module to determine whether they are potential signs of systemic risk.

3. The intelligent monitoring and evaluation system for nighttime duty performance in prisons according to claim 2, characterized in that, The multi-agent collaborative judgment module also includes a third agent, which is used to perform weighted fusion decision on the outputs of the first agent and the second agent to generate the final risk judgment conclusion; wherein, the weight of each agent is dynamically adjusted according to its historical judgment accuracy.

4. The intelligent monitoring and evaluation system for nighttime duty performance in prisons according to claim 1, characterized in that, The system adaptive optimization module includes a knowledge graph maintenance unit; the knowledge graph maintenance unit is configured to extract structured knowledge from newly added rules and regulations, accident reports and handling records, and integrate it with existing knowledge in the duty performance knowledge graph module to update the knowledge graph.

5. The intelligent monitoring and evaluation system for nighttime duty performance in prisons according to claim 1, characterized in that, The system adaptive optimization module includes a rule dynamic update unit; the rule dynamic update unit is configured to receive rule modification instructions through a natural language interaction interface, and after performing semantic parsing and testing on the instructions, deploy the verified detection logic to the edge real-time analysis unit.

6. The intelligent monitoring and evaluation system for nighttime duty performance in prisons according to claim 1, characterized in that, The multi-source behavior sensing unit includes: UWB positioning module, used to obtain the location coordinates of personnel; A multi-camera surveillance network is deployed using a combination of rotatable PTZ cameras and fixed-view cameras to cover blind spots in the prison area. The electronic patrol module is used to record the location and timestamp of the patrol points.

7. The intelligent monitoring and evaluation system for nighttime duty performance in prisons according to claim 1, characterized in that, The edge real-time analysis unit is deployed in the edge computing gateway in the prison area. It is used to run computer vision target recognition algorithms to identify behaviors such as leaving the post, running, falling, climbing, and gathering, and to spatiotemporally correlate the recognition results with UWB positioning coordinates.

8. A method for intelligent monitoring and evaluation of nighttime duty performance in prisons, characterized in that, The method includes the following steps: Data synchronization and collection steps: Synchronously and in real time collect the location data of personnel performing their duties, video surveillance data, and electronic patrol data; Edge real-time processing steps: On the computing nodes deployed locally in the prison area, the collected data is fused and preliminarily analyzed to identify preliminary abnormal events that meet the preset rules; Cloud-based collaborative analysis steps: The preliminary abnormal events and related historical and environmental data are transmitted to the cloud; In the cloud, multiple large-scale language model agents perform collaborative reasoning based on a pre-built duty performance knowledge graph to identify potential risk patterns that exceed preset rules and are formed by the association of multiple events in a specific time, space or logic. Early warning and report generation steps: Based on the type and level of the potential risk patterns, perform graded early warnings and generate a comprehensive assessment report that includes risk evolution analysis; System adaptive optimization steps: Based on the identified potential risk patterns and feedback on the effects of subsequent handling, dynamically update the content of the duty performance knowledge graph and the identification rules applied in the edge real-time processing steps.

9. A method for intelligent monitoring and evaluation of nighttime duty performance in prisons according to claim 8, characterized in that, The cloud-based collaborative analysis steps specifically include: The first intelligent agent performs procedural compliance checks on the event based on plaintext rules; The second intelligent agent, in conjunction with the aforementioned duty performance knowledge graph, performs contextual analysis on the occurrence of the event to determine whether it is a sign of systemic risk. A third intelligent agent integrates the assessment results of the first two to make a decision, generating the final risk assessment conclusion.

10. A method for intelligent monitoring and evaluation of nighttime duty performance in prisons according to claim 8, characterized in that, In the early warning and report generation step, the level of the potential risk pattern is determined by the following steps: based on the type and duration of a single abnormal behavior, it is mapped to a predefined basic alarm level; If the single abnormal behavior is identified as part of the potential risk pattern, or if its location is in a high-risk area, the warning level will be increased from the basic alarm level. If multiple related abnormal behaviors are identified in the same time period and in the same area, the highest warning level will be triggered directly. The system adaptive optimization steps include a knowledge update sub-step and a rule update sub-step, wherein... The knowledge update sub-step includes extracting structured knowledge from newly added rules and regulations and handling records to update the duty performance knowledge graph; The rule update sub-step includes generating or modifying detection rules by parsing natural language instructions, and after testing and verification, deploying them to the edge real-time processing step for application.

Images