Closed-loop reconstruction video analysis system and method based on self-evolution MoE Agent architecture

Abstract

The invention discloses a closed-loop reconstruction video analysis system and method based on a self-evolution MoE Agent architecture, and the system is realized based on a five-layer architecture, and comprises a camera access layer which accesses a camera to receive video stream data input; the video data analysis and basic algorithm layer is used for carrying out decoding, preprocessing and basic feature extraction on the video stream; the self-evolution Agent core layer is used for configuring an Agent agent and a related model and transmitting and memorizing the Agent agent and the related model; the self-evolution engine and continuous learning layer is used for converting the error attribution into a model optimization instruction based on the self-evolution Agent core layer; and the closed-loop control and feedback layer is used for converting the final decision of the Agent into an external action. According to the scheme, the core problems that an existing video analysis system is rigid in algorithm, cannot continuously adapt to new tasks, is low in large-scale model edge deployment efficiency, is low in closed-loop reasoning reliability in complex tasks and the like in a dynamic environment are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent video analytics technology, and in particular to a closed-loop reconstructed video analytics system and method based on a self-evolving MoE Agent architecture. Background Technology

[0002] Existing video surveillance systems are widely used in industrial security, smart parks and other scenarios, but they still face three challenges: rigid algorithms in traditional video analysis systems and low edge deployment efficiency and weak continuous learning ability of the new generation of MLLM Agent systems.

[0003] Limitations of fixed algorithms: Traditional video analytics relies on pre-set, fixed algorithms, making it difficult to quickly adapt to ever-changing business needs or emerging recognition tasks. Whenever new analytical capabilities are required, it often necessitates time-consuming and labor-intensive retraining of models and re-integration of the system.

[0004] Reliability and adaptability issues of static agents: While existing MLLM agents can provide deep semantic understanding, their decision-making logic and algorithm configurations are usually pre-set and static. When the system faces unknown anomalies or environmental changes beyond the training data range, the reliability of its reasoning and decision-making drops sharply, and it lacks autonomous learning and self-evolution capabilities. They cannot iteratively optimize and autonomously correct errors based on their own failures, which is a fatal flaw in critical closed-loop control tasks.

[0005] The computational bottleneck of large-scale model edge deployment: Powerful multimodal large models require massive computing resources due to their huge number of parameters. This is fundamentally contradictory to the constraints of real-time performance (low latency) and edge device resources (low power consumption, low cost NVR / AIBox) in actual monitoring scenarios. Existing dynamic scheduling systems only focus on load balancing and fail to solve the edge computing efficiency problem at the model architecture level.

[0006] Therefore, there is an urgent need in this field for an innovative system and method that can combine powerful MLLM agent inference capabilities, highly reliable autonomous learning mechanisms, and efficient edge deployment architecture. Summary of the Invention

[0007] To address the aforementioned technical issues, this invention provides a closed-loop reconstructed video analysis system and method based on a self-evolving MoE Agent architecture. This involves technologies such as artificial intelligence, computer vision, large-scale language model agents, continuous learning, sparse activation model architecture, and edge computing. In particular, it presents a self-evolving intelligent agent system based on a sparse activation Mixture-of-Experts architecture, used to achieve deep semantic-level analysis, task planning, closed-loop control, and continuous integration of autonomous knowledge from camera videos. This system is suitable for high real-time performance and resource-constrained edge monitoring scenarios.

[0008] This invention is achieved using the following technical solution: Firstly, a closed-loop reconstructed video analytics system based on a self-evolving MoE Agent architecture is implemented using a five-layer architecture, specifically including: Camera access layer: receives video stream data input from cameras; Video data parsing and basic algorithm layer: Decoding, preprocessing and basic feature extraction of video streams; Self-evolving Agent Core Layer: Configures the Agent and related models, and performs emission and memorization; Self-evolving engine and continuous learning layer: Based on the core layer of self-evolving agent, error attribution is transformed into model optimization instructions; Closed-loop control and feedback layer: Transforms the final decision of the agent into external actions.

[0009] Specifically, the self-evolving Agent core layer includes: Algorithm Scheduling and AaaS Plugin Module: Dynamically loads and instantiates containerized algorithm AaaS plugins based on the pipeline planned by the Agent. Edge-Native MoE MLLM Agent Model: Configures a sparse activation MoE architecture as the core inference model, dynamically activating a small subset of experts based on the input, responsible for semantic understanding, contextual reasoning, and task planning; Reflection and Memory Module: Receives execution results and external feedback, performs decision evaluation and structured error attribution, and stores the results.

[0010] Specifically, the results of the reflection and memory module are stored as three types of memory: procedural memory, strategic memory, and tool memory.

[0011] Specifically, the self-evolutionary engine and continuous learning layer include: MoE Continuous Tuning Module: Based on PEFT parameter fine-tuning, a new MoE expert network is trained in a targeted manner and injected into the Edge-Native MoE MLLM Agent model to achieve incremental knowledge integration and continuous learning; Edge MoE Deployment Manager: Deploys the Edge-Native MoE MLLM Agent model in a distributed edge environment for efficient operation; MoE Expert Scheduling Module: Monitors edge device resources, dynamically allocates and coordinates expert computing tasks based on the expert subset activated by the Edge-Native MoE MLLM Agent model, and assigns them to different edge nodes for collaborative inference; Model synchronization module: manages new or optimized expert networks trained by the self-evolutionary engine and continuous learning layers, and registers and synchronizes them to edge devices.

[0012] Specifically, the Edge-Native MoE MLLM Agent model includes a gating network and several expert networks, wherein: The gated network sparsely activates a subset of experts to perform inference based on the input video analytics metadata. When the Edge-Native MoE MLLM Agent model activates a subset of experts, the Edge MoE deployment manager dynamically allocates the activated computing tasks to different edge devices for collaborative inference based on the experts' computing needs, the real-time load of the edge devices, and network latency.

[0013] Specifically, the decision evaluation and structured error attribution of the reflection and memory module include: Collect structured event logs, control command execution status, and possible user feedback; Determine whether the inference confidence of the agent is lower than the threshold, and calculate the deviation between the actual execution result and the expected goal; If a deviation is detected, the agent uses its memory system to attribute the error.

[0014] Specifically, the error attribution includes: If the reasoning logic is correct, but the algorithm pipeline causes the result to be inaccurate, it is attributed to program failure; If the program logic and strategic reasoning are correct, but the underlying visual algorithm fails to recognize newly emerging objects or environmental changes, it is attributed to tool failure. If the underlying algorithm yields accurate results, but the agent misunderstands the scene context or causal relationships, it is attributed to strategic failure. The reflection and memory module transforms attribution results into structured new training instructions or optimization guidelines, updating the agent's procedural memory, strategic memory, and tool memory.

[0015] Specifically, the model optimization of the self-evolutionary engine and continuous learning layer also includes: Based on the structured error attribution results report of the reflection and memory module, a task-aware discriminator is used to evaluate the nature of the failure report and determine whether the required knowledge is a fine-tuning of existing experts or the creation of entirely new experts. An adversarial hybrid expert continuous tuning architecture based on LoRA is used to train and fine-tune the experts; The newly trained expert network is injected into the new Agent expert pool via the Edge MoE Deployment Manager.

[0016] On the other hand, the closed-loop reconstructed video analysis method based on the self-evolving MoE Agent architecture includes the following steps: Step S1: Task initialization, user-configured task, Edge-Native MoE MLLM Agent model generates initial algorithm execution pipeline based on its strategic memory and procedural memory; Step S2: Access video stream data, and under the coordination of the edge MoE deployment manager, the algorithm scheduling and AaaS plug-in modules execute the pipeline to output structured analysis results; Step S3: The Edge-Native MoE MLLM Agent model performs semantic understanding and preliminary decision-making on the analysis results, and generates control instructions; Step S4: The reflection and memory module receives semantic understanding, preliminary decision-making, and control instructions, and evaluates the reliability of the decision and whether the execution result is successful; Step S5: If the evaluation fails or the confidence level is low, the reflection and memory modules perform error attribution and trigger the self-evolution engine and continuous learning layer; the self-evolution engine and continuous learning layer train a new expert network, update the knowledge of the Edge-Native MoE MLLM Agent model, and update the three types of memory in the reflection and memory modules; the Edge-Native MoE MLLM Agent model uses the updated memory to autonomously reconstruct the algorithm pipeline or optimize the decision logic; the process returns to step S2 and is executed again until success or the iteration limit is reached; Step S6: If the evaluation is successful, the task ends and a log is recorded.

[0017] The beneficial effects of this invention are as follows: This invention proposes a closed-loop reconstruction video analysis system and method based on a self-evolving MoEAgent architecture. It introduces a closed-loop reflection mechanism (H) driven by a continuous learning engine (J), and utilizes a sparse activation Mixture-of-Experts (MoE) architecture and an edge MoE deployment manager (K) to achieve efficient edge deployment. This solves the core problems of existing video analysis systems in dynamic environments, such as algorithm rigidity, inability to continuously adapt to new tasks, low efficiency of large-scale model edge deployment, and low reliability of closed-loop inference in complex tasks. It also has the following technical effects: It enables agents to continuously and autonomously evolve and accumulate knowledge: Through the MoE-CL architecture, the system can dynamically learn new tasks and autonomously integrate new knowledge, completely solving the problems of algorithm rigidity and knowledge updating in traditional agent systems.

[0018] Extremely high inference efficiency and edge adaptability: The Edge-Native MoE MLLM Agent, combined with EDM(K), leverages the sparse activation characteristics of MoE and collaborative computing between edge devices to significantly reduce the resource consumption and real-time latency of deploying complex MLLMs at the edge.

[0019] Improved reliability of decision-making in complex tasks: The introduction of a reflection and memory module (H) enables closed-loop correction of decisions and optimization of instructions through iterative evaluation and error attribution, ensuring the accuracy of judgments in high-risk scenarios.

[0020] It enables the agent to autonomously manage its own tools: the agent can not only call algorithms, but also autonomously evaluate and improve the algorithm plugins (tool memory) and execution logic (program memory) it uses based on the running results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0022] Figure 1 This is an architecture diagram of a closed-loop reconstructed video analysis system based on a self-evolving MoE Agent architecture, as described in this embodiment of the invention. Figure 2 This is a flowchart of the closed-loop reconstructed video analysis method based on the self-evolving MoE Agent architecture in this embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] The following is in conjunction with the appendix Figure 1-2 The following describes some embodiments of the present invention in detail. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0026] This invention proposes a closed-loop reconstructed video analysis system and method based on a self-evolving MoE Agent architecture. In a preferred embodiment, the system is implemented based on a five-layer architecture, specifically including: Camera access layer: receives video stream data input from cameras; Video data parsing and basic algorithm layer: Decoding, preprocessing and basic feature extraction of video streams; Self-evolving Agent Core Layer: Configures the Agent and related models, and performs emission and memorization; Self-evolving engine and continuous learning layer: Based on the core layer of self-evolving agent, error attribution is transformed into model optimization instructions; Closed-loop control and feedback layer: Transforms the final decision of the agent into external actions.

[0027] In this embodiment, the self-evolving Agent core layer specifically includes: Algorithm Scheduling and AaaS Plugin Module: Dynamically loads and instantiates containerized algorithm AaaS plugins based on the pipeline planned by the Agent. Edge-Native MoE MLLM Agent Model: Configures a sparse activation MoE architecture as the core inference model, dynamically activating a small subset of experts based on the input, responsible for semantic understanding, contextual reasoning, and task planning; Reflection and Memory Module: Receives execution results and external feedback, performs decision evaluation and structured error attribution, and stores the results.

[0028] In this embodiment, the self-evolutionary engine and the continuous learning layer specifically include: MoE Continuous Tuning Module: Based on PEFT parameter fine-tuning, a new MoE expert network is trained in a targeted manner and injected into the Edge-Native MoE MLLM Agent model to achieve incremental knowledge integration and continuous learning; Edge MoE Deployment Manager: Deploys the Edge-Native MoE MLLM Agent model in a distributed edge environment for efficient operation; MoE Expert Scheduling Module: Monitors edge device resources, dynamically allocates and coordinates expert computing tasks based on the expert subset activated by the Edge-Native MoE MLLM Agent model, and assigns them to different edge nodes for collaborative inference; Model synchronization module: manages new or optimized expert networks trained by the self-evolutionary engine and continuous learning layers, and registers and synchronizes them to edge devices.

[0029] In one specific embodiment, the system architecture is as follows: Figure 1 As shown, the core layer includes the connections and data flows between the self-evolution engine (J), the reflection and memory module (H), the Edge-Native MoE MLLM Agent (D), and the Edge MoE Deployment Manager (K). These components collectively drive the algorithm scheduling layer (C) and implement closed-loop control. The system composition specifically includes: Camera access layer (A): Responsible for receiving video stream input; Video Data Parsing and Basic Algorithm Layer (B): Decodes, preprocesses, and extracts basic features from the video stream.

[0030] Self-Evolving Agent Core Layer: Algorithm Scheduling and AaaS Plugin Module (C): Dynamically loads and instantiates containerized algorithm plugins (AaaS) based on the pipeline planned by the Agent. Edge-Native MoE MLLM Agent(D): The core reasoning model, which adopts a sparse activation MoE architecture, dynamically activates a small subset of experts based on the input, and is responsible for semantic understanding, contextual reasoning and task planning; The Reflection and Memory Module (H) receives execution results and external feedback, performs decision evaluation and structured error attribution (tool, procedure, strategic failure), and stores the results as three types of memory: procedure memory, strategic memory, and tool memory.

[0031] Self-evolutionary engine and continuous learning layer (J): Responsible for translating the error attribution of the H module into model optimization instructions, including: MoE Continuous Tuning Module (MoE-CLTuning): Based on PEFT technology, it trains new MoE expert networks in a targeted manner and injects them into the Edge-Native MoE MLLM Agent to achieve incremental knowledge integration and continuous learning; Edge MoE Deployment Manager (EDM)(K): Responsible for the efficient operation of MoEMLLMAgent in a distributed edge environment; MoE Expert Scheduling Module: Monitors edge device resources and dynamically allocates and coordinates expert computing tasks to different edge nodes for collaborative inference based on the subset of experts activated by MoEAgent. Model synchronization module: manages new or tuned expert networks trained in the J layer and efficiently registers and synchronizes them to edge devices.

[0032] Closed-loop control and feedback layer (F): Transforms the Agent's final decision into external actions (alarms, control commands I).

[0033] The design of some specific technologies in the system of this invention will be described in detail below: (1) Edge deployment of Edge-Native MoE MLLM Agent(D) This invention employs a MoE architecture for MLLM agents; this model comprises a gating network and several expert networks. The gating network sparsely activates a subset of experts for inference based on the input video analytics metadata (S105). The Edge MoE Deployment Manager (EDM) (K) monitors the computing resource pools of edge computing devices (e.g., AI boxes equipped with NVIDIA Jetson or domestically produced chips (such as Suaneng, Huawei, Lingxi, etc.), smart NVRs, etc.) in real time. When the MoE Agent activates a subset of experts, the EDM dynamically allocates the activated computing tasks to different edge devices for collaborative inference based on the experts' computing needs, the real-time load of the edge devices, and network latency. For example, if the Agent needs experts $E_1$ (for personnel posture analysis) and $E_2$ (for specific hazardous material identification), the EDM can schedule $E_1$ to AI Box A and $E_2$ to AI Box B. The two exchange intermediate activation results through the edge network and finally aggregate the results to the Agent (D). This method effectively utilizes the sparsity of MoE, achieving low-latency, high-throughput MLLM inference in resource-constrained environments.

[0034] (2) Structured error attribution of the reflection and memory module (H) The reflection and memory module (H) intervenes immediately after the agent's decision (S106) is translated into actual action (S108). Its core steps include: Results feedback collection: Collect structured event logs (E), control command execution status (F), and possible user feedback (such as confirmation / denial of alarms). Confidence and Bias Check: Determine if the Agent's inference confidence is below a threshold, and calculate the deviation between the actual execution result and the expected goal (e.g., if the Agent plans to close the gate, but the sensor indicates that the gate is not closed, record the deviation). Level 3 error attribution: If a bias is detected, the agent uses its memory system to make an attribution.

[0035] In this embodiment, the three-level error attribution specifically includes: Program failure: If the reasoning logic is correct, but the algorithm pipeline (the order or parameters of the AaaS plugins) causes the result to be inaccurate, it is attributed to program failure. Tool failure: If the program logic and strategic reasoning are correct, but the underlying vision algorithm (such as the YOLO model) fails to recognize new objects or environmental changes, it is attributed to tool failure. Strategic failure: If the underlying algorithm is accurate, but the agent misunderstands the scene context or causal relationship, it is attributed to strategic failure. Memory Update: The H module transforms the attribution results into structured new training instructions or optimization guidelines, updating the agent's procedural memory, strategic memory, and tool memory. These memories are stored using vector embedding for the agent to access in steps S101 and S110.

[0036] (3) Continuous optimization of the MoE of the self-evolution engine (J) After receiving a failure report from module H, the self-evolutionary engine (J) initiates the autonomous learning process: Task identification: The task-aware discriminator evaluates the nature of failure reports to determine whether the required knowledge is a fine-tuning of existing experts or the creation of entirely new experts; MoE Expert Training (MoE-CL): This approach employs an adversarial hybrid expert continuous tuning architecture based on LoRA (Parameter-Efficient Fine-Tuning, PEFT). This method allows the system to efficiently train or fine-tune specific experts with a very small number of parameters: for entirely new knowledge requirements, a new sparse expert is trained specifically to handle new recognition tasks or failure scenarios. During training, the system ensures that the integration of new knowledge does not lead to a performance degradation of existing experts on previous tasks (overcoming catastrophic forgetting). New Knowledge Injection and Reconstruction (S110): J injects the newly trained expert network into the expert pool of Agent (D) through EDM (K). At the same time, Agent uses the updated memory of H to reconstruct its algorithm pipeline of S101. For example, it incorporates the AaaS plugin corresponding to the new expert into the execution process, realizing a complete closed loop from failure perception, knowledge learning to autonomous reconstruction.

[0037] In this embodiment, the continuous learning method of the self-evolutionary engine (J) adopts the MoE-CL architecture to achieve incremental knowledge updates and task adaptability for the agent. The closed-loop reflexive reconstruction mechanism (H) achieves autonomous reconstruction and reliability enhancement of agent decision-making through iterative evaluation and updates of three types of memory (programs, strategies, and tools). The system architecture (K) of the edge MoE deployment manager (EDM) enables efficient collaborative scheduling and resource management of the MoE sparse activation model in a distributed edge computing environment, ensuring that MLLM can run in a low-latency, resource-constrained edge environment.

[0038] The video intelligent analysis process based on the above system is as follows: Figure 2 As shown, after S107 (reflection evaluation), if it fails, the process proceeds to S109 (continuous learning) and S110 (pipeline refactoring), demonstrating how the system performs autonomous knowledge updates and iterative optimization, specifically including: Task initialization (S100, S101): User-configured task (G). Agent (D) generates an initial algorithm execution pipeline (C) based on its strategic memory and procedural memory; Data processing and execution (S102~S105): Access video streams (A, B). The algorithm scheduling layer (C) executes the pipeline under the coordination of EDM(K) and outputs structured analysis results (S105); Agent reasoning and preliminary decision-making (D, S106): MoEAgent(D) performs semantic understanding and preliminary decision-making (S106) and generates control instructions (S108); Reflection evaluation (S107): The reflection and memory module (H) receives the results of S106 / S108 and evaluates the reliability of the decision and whether the execution result is successful; Failure Path (Autonomous Reconstruction) (S109, S110): If the evaluation fails or the confidence level is low, H performs error attribution and triggers the self-evolutionary engine (J) (S109). J starts the MoE-CL process, trains a new expert network, updates the knowledge of Agent (D), and updates H's three types of memories. The Agent uses the updated memories to autonomously reconstruct the algorithm pipeline or optimize the decision logic (S110). The process returns to S104 and executes again until success or the iteration limit is reached; Success path: If the evaluation is successful, the task ends and a log is recorded (E_END).

[0039] In a specific application scenario: warehouse monitoring, which requires identifying "unauthorized intrusion outside of working hours," the system's application specifically includes: Initial Execution and Failure: Agent Initial Configuration Pipeline: Personnel Detection (AaaS1) -> Time Comparison (AaaS2) -> Agent (D) Judgment. One weekend, personnel intrusion was detected, and AaaS1 reported "Personnel Confidence 99%". Agent (D) determined "Unauthorized Intrusion" and triggered an alarm (S108). However, security personnel reported that the intrusion was actually caused by a large animal (such as a deer), and the alarm was a false alarm.

[0040] Reflection and Attribution (H): The H module receives "false alarm" feedback and determines that the task has failed. The H attribution result is that the tool has failed: AaaS1 (person detection model) cannot effectively distinguish between people and large animals. The H module generates a new training instruction: "Train the model to distinguish between people and deer / large animals", and updates the tool memory.

[0041] Self-evolution (J) and MoE expert generation: J receives a tool failure report and initiates the MoE-CL process. The system uses newly acquired image data of "deer" and "humans" to train a new MoE expert, focusing on finely distinguishing the features of humans and large animals. The new expert is injected into Agent (D) via EDM (K).

[0042] Pipeline Reconstruction and Iteration Success (S110): Agent (D) updates its program memory and reconstructs the pipeline: Personnel Detection (AaaS1) -> New Expert-Assisted Judgment (AaaS3) -> Time Comparison (AaaS2) -> Agent (D) Final Judgment. When a similar event occurs again, Edge-NativeMoEAgent (D) activates a new expert (cooperatively scheduled to the edge device), accurately identifies the "large animal," and based on the safety rules in the strategic memory ("large animals do not require immediate alarm"), ultimately determines it as a "risk-free event," thereby avoiding false alarms and completing closed-loop iterative optimization.

[0043] In one embodiment, the invention also includes alternative solutions, such as: MoE expert scheduling algorithm: More complex load balancing algorithms, QoS-aware scheduling strategies, or prediction-based expert preloading strategies can be used in EDM(K). Continuous learning techniques: The MoE-CL module can be replaced with other PEFT methods, such as AdapterTuning, PrefixTuning, etc., or an incremental learning framework can be adopted; Memory storage format: Program, strategy, and tool memory can adopt different storage structures, such as knowledge graphs, vector databases, or hybrid storage; Error attribution mechanism: More complex causal inference models or rule-based expert systems can be introduced to assist in error attribution.

[0044] This invention also features scalable capabilities, such as: Multimodal perception fusion: Agent(D) can further fuse other modal information such as audio and text to enhance understanding and decision-making capabilities; Active sample collection: The self-evolutionary engine (J) can proactively request labeled samples from human operators based on low confidence or novel scenarios, thereby accelerating the learning process; Enhanced security and privacy: Implement differential privacy or federated learning mechanisms at the edge to protect video data privacy; Agent collaboration and knowledge sharing: Deploy multiple agents and achieve collaboration and capability enhancement among agents through shared memory or knowledge distillation mechanisms.

[0045] For the foregoing embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to this application.

[0046] The above embodiments describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Modifications and variations made by those skilled in the art without departing from the spirit and scope of the invention should be within the protection scope of the appended claims.

Claims

1. A closed loop reconfigurable video analytics system based on a self-evolving MoE Agent architecture, characterized in that, Based on the implementation of a five-layer architecture, specifically including: Camera access layer: access camera to receive video stream data input; Video data analysis and basic algorithm layer: decode, preprocess and extract basic features from video stream; Self-evolution Agent core layer: configure Agent and related models, and perform transmission and memory; Self-evolution engine and continuous learning layer: based on the self-evolution Agent core layer, error attribution is converted into model optimization instructions; Closed-loop control and feedback layer: convert the final decision of Agent into external action.

2. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 1 wherein, The self-evolution Agent core layer specifically includes: Algorithm scheduling and AaaS plug-in module: dynamically load and instantiate containerized algorithm AaaS plug-ins according to the pipeline planned by Agent; Edge-Native MoE MLLM Agent model: configure sparse activation MoE architecture as the core inference model, dynamically activate a few expert subsets according to input, responsible for semantic understanding, context reasoning and task planning; Reflection and memory module: receive execution results and external feedback, perform decision evaluation and structured error attribution, and store the results.

3. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 2, wherein, The reflection and memory module results are stored as three types of memory: procedural memory, strategic memory and tool memory.

4. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 3, wherein, The self-evolution engine and continuous learning layer specifically includes: MoE continuous optimization module: based on PEFT parameter fine-tuning, train new MoE expert networks and inject them into the Edge-Native MoE MLLM Agent model to realize incremental knowledge integration and continuous learning; Edge MoE deployment manager: deploy the Edge-Native MoE MLLM Agent model in a distributed edge environment for efficient operation; MoE expert scheduling module: monitor edge device resources, dynamically allocate and coordinate expert computing tasks according to the activated expert subsets of the Edge-Native MoE MLLM Agent model, and allocate them to different edge nodes for collaborative reasoning; Model synchronization module: manage new experts or optimized expert networks trained by the self-evolution engine and continuous learning layer, and register and synchronize them to edge devices.

5. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 4, wherein, The Edge-Native MoE MLLM Agent model includes a gating network and several expert networks, wherein: The gating network activates part of the experts for reasoning according to the input video analysis metadata; when the Edge-Native MoE MLLM Agent model activates the expert subsets, the edge MoE deployment manager dynamically allocates the activated computing tasks to different edge devices for collaborative reasoning according to the computing requirements of the experts, the real-time load of the edge devices and the network delay.

6. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 5 wherein, The decision evaluation and structured error attribution of the reflection and memory module specifically includes: Collect structured event records, control instruction execution status and possible user feedback; Determine whether the reasoning confidence of Agent is below the threshold, and calculate the deviation between the actual execution result and the expected target; If deviation is detected, the Agent uses its memory system to attribute the error.

7. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 6 wherein, The error attribution specifically includes: If the reasoning logic is correct, but the algorithm pipeline leads to inaccurate results, it is attributed to program failure; If the program logic and strategic reasoning are correct, but the underlying visual algorithm fails to recognize new objects or environmental changes, it is attributed to tool failure; If the underlying algorithm results are accurate, but the Agent's understanding of the scene context or causal relationships is incorrect, it is attributed to strategic failure; The reflection and memory module converts the attribution results into structured new training instructions or optimization guidelines, updating the Agent's program memory, strategic memory, and tool memory.

8. The closed loop reconfigurable video analytics system based on a Self- Evolving MoE Agent architecture as claimed in claim 7, wherein, The self-evolution engine and continuous learning layer model optimization also includes: Based on the structured error attribution results reported by the reflection and memory module, the task-aware discriminator assesses the nature of the failure report to determine whether the required knowledge is a fine-tuning of existing experts or the creation of new experts; Using the LoRA-based adversarial hybrid expert continuous optimization architecture to train and fine-tune the experts; The trained new expert network is injected into the new Agent expert pool through the edge MoE deployment manager.

9. The closed-loop reconfigurable video analysis method based on the self-evolution MoE Agent architecture, implemented by the closed-loop reconfigurable video analysis system based on the self-evolution MoE Agent architecture according to any one of claims 1-8, characterized in that, The process includes the following steps: Step S1: Task initialization, user configures task, Edge-Native MoE MLLM Agent model generates initial algorithm execution pipeline based on its strategic memory and program memory; Step S2: Access video stream data, algorithm scheduling and AaaS plugin modules execute the pipeline under the coordination of the edge MoE deployment manager, output structured analysis results; Step S3: Edge-Native MoE MLLM Agent model performs semantic understanding and preliminary decision-making on the analysis results, and generates control instructions; Step S4: The reflection and memory module receives semantic understanding, preliminary decision-making, and control instructions, evaluates the reliability of the decision and whether the execution result is successful; Step S5: If the evaluation fails or the confidence is low, the reflection and memory module attributes the error and triggers the self-evolution engine and continuous learning layer; the self-evolution engine and continuous learning layer train new expert networks, update the knowledge of the Edge-Native MoE MLLM Agent model, and update the three types of memory of the reflection and memory module; the Edge-Native MoE MLLM Agent model uses the updated memory to automatically reconstruct the algorithm pipeline or optimize the decision logic; the process returns to step S2 for execution again until success or the iteration limit is reached; Step S6: If the evaluation is successful, the task is completed, and the log is recorded.

Images