A central nervous network host and system for training a super agent cluster

Abstract

This invention belongs to the field of AI swarm intelligence training technology, and discloses a central neural network host and system for professionally training super agent clusters. The core is a rack-mounted central neural network host, which includes a heterogeneous computing array, a central neural network chip, an algorithm-encrypted globally shared storage unit, low-latency high-speed interconnect interfaces, a hardware-level risk control and computing power allocation unit, and a stable power supply and heat dissipation module. Each module is independently powered and supports hot-swapping. As the cluster core, this host achieves millisecond-level scheduling, microsecond-level communication, and hardware-level security management of the super agent cluster, significantly improving the training efficiency and computing power utilization of the agent cluster, ensuring the security and stability of the training process, and providing a domestically produced hardware foundation for the unified training and collaborative evolution of super agent clusters, adapting to the needs of large-scale AI applications in multiple scenarios.

Description

Technical Field

[0001] This invention belongs to the field of AI swarm intelligence training technology, specifically involving a central neural network host and system for professionally training super agent clusters. Background Technology

[0002] Traditional AI often employs a single-agent architecture, exhibiting weak cross-scenario collaboration capabilities and poor scenario adaptability, failing to meet the large-scale intelligent operation needs of diverse scenarios such as physical commerce and one-person companies (OPCs). Related AI training and deployment lack a unified technological foundation, resulting in inefficient computing power scheduling and significant waste of computing resources. Platform subsidies and reward mechanisms are easily exploited by fraudulent activities such as order manipulation, cash-out schemes, and self-trading, causing substantial economic losses. Furthermore, AI training data faces security risks and non-compliant circulation issues, hindering the effective monetization of data value. Existing AI training systems lack domestically developed and controllable end-to-end solutions, making core hardware and software susceptible to external technological limitations. In addition, the low efficiency of single-agent training and the lack of effective collaborative evolution and global scheduling mechanisms for clustered agents further restrict the practical application of AI swarm intelligence. Summary of the Invention

[0003] To overcome the above-mentioned technical problems, the present invention provides a central neural network host and system for professionally training super agent clusters.

[0004] The present invention adopts the following technical solution: A central neural network host for training super agent clusters includes: Heterogeneous computing array: integrates multi-core CPU, training-grade GPU, edge inference NPU, and intelligent scheduling DPU to support distributed training and inference parallelism of large models; Central neural network chip: Built-in swarm intelligence routing algorithm, task decomposition engine, agent scheduling model and evolutionary algorithm module to achieve millisecond-level scheduling of agent clusters; Global shared storage unit: includes DDR5 high-speed cache and solid-state persistent storage area, TEE trusted execution environment encryption based on national cryptographic SM4 algorithm, supports second-level backup and off-site disaster recovery; High-speed interconnect interfaces: Equipped with PCIe 5.0, CXL 2.0, and 100G optical ports, enabling microsecond-level communication between the master node and edge nodes; Risk control and computing power rights confirmation unit: a hardware-level independent module that integrates encryption and behavior recognition chips, and has a built-in abnormal behavior feature library to achieve hardware-level data verification and real-time blocking of abnormal behavior; Power and thermal management module: It adopts a redundant dual power supply and a liquid cooling + air cooling composite heat dissipation solution to support 24 / 7 high-density rack-mounted steady-state operation.

[0005] This invention also discloses a central neural network system for professionally training super agent clusters, comprising: Central neural network host; Super Agent Cluster Layer: Includes digital store managers, digital employees, and OPC-specific agents. Initial deployment supports tens of thousands of concurrent users and can be dynamically scaled up or down. Each agent is finely trained based on a large model base. Swarm intelligence training platform: It adopts a training mode that combines reinforcement learning, federated learning and model distillation, and provides capabilities for task distribution, hierarchical rewards and lightweight model deployment; Computing power voucher ownership confirmation and redemption system: Built on a domestic consortium blockchain, it realizes the reliable generation, ownership confirmation, circulation and redemption of computing power vouchers, and the whole process is on-chain and auditable; Risk control and anti-fraud engine: It adopts a dual risk control mode of hardware verification + software analysis to intercept fraudulent transactions, self-trading, cash-out accounts, and low-entry-high-exit behaviors in real time; Data circulation and value realization module: Connects to compliant data trading venues to achieve data anonymization, listing, trading, settlement and full-chain traceability.

[0006] Preferably, in the super agent cluster layer, the digital store manager agent is adapted to offline physical stores such as supermarkets, convenience stores, and fresh food stores. It can connect to hardware such as store POS machines, inventory management systems, and customer flow cameras, and has full-process operation capabilities such as product display planning, inventory management, customer flow analysis, promotion planning, and cashier settlement.

[0007] Preferably, in the super agent cluster layer, the digital employee agent is divided into four subcategories: content creation, intelligent customer service, data statistics, and financial accounting. It supports multi-channel 24 / 7 operation and is adapted to the content production, customer service, data management, and financial accounting needs of individual businesses and micro-enterprises.

[0008] Preferably, in the super agent cluster layer, the OPC exclusive agent is customized for one-person company entrepreneurs, integrating business registration, qualification processing, supply chain connection, customer development, and compliance risk control capabilities. It can seamlessly collaborate with digital store manager and digital employee agents across scenarios to provide one-stop intelligent entrepreneurship services.

[0009] Preferably, the model distillation module of the swarm intelligence training platform can distill the large model knowledge on the host side into a lightweight model, compressing the model volume to 1 / 10 of the original model, increasing the inference speed by 5 times, realizing the lightweight deployment of edge agent nodes, and reducing computing power consumption.

[0010] Preferably, in the computing power voucher confirmation and redemption system, computing power vouchers are generated only from real consumption, real fulfillment, and real and valid data. One unit of computing power voucher corresponds to 1 yuan of real consumption data value. The minimum redemption unit is 10 units. Redemption categories include physical goods, cloud services, entrepreneurship training, and computing power services. Free transfer between compliant entities is supported, while paid buying and selling and illegal monetization are prohibited.

[0011] Preferably, the risk control and anti-fraud engine includes a device fingerprint and behavior profiling module, a transaction link verification module, a logistics / service fulfillment verification module, a cash-out account identification and blocking module, and a loss accounting module. The loss accounting module can calculate platform losses such as payment fees, taxes, rewards, and operating costs caused by fraudulent behavior in real time and generate loss accounting reports.

[0012] Preferably, the data circulation and value realization module removes sensitive information from the raw data collected by the Agent. After the data is desensitized, it is listed in a compliant venue after being verified by the host. The bidding transaction mode is adopted. After the transaction is completed, the data delivery and fund settlement are completed, and the settlement funds are distributed to the data producer and the platform in proportion.

[0013] Preferably, the host executes a distributed collaborative training process for the super agent cluster, including: Receive scenario-based training and operation tasks, and break them down into sub-tasks with execution requirements, priorities, and completion deadlines; Subtasks are dynamically allocated based on Agent scenario adaptability, computing load, and historical execution efficiency, with high-priority core tasks being prioritized for allocation to high-quality Agent nodes. Establish a dynamic communication topology network for agents participating in the task, designate a master / cooperating agent, and automatically reassign tasks and adjust the topology when a node fails. A reward function is constructed based on real and valid behavioral data, with data validity, task completion rate, and operational efficiency as the core, and computing power vouchers are issued in conjunction with the computing power voucher system. The system uses both hardware and software to intercept abnormal orders, high-frequency transactions from the same device / account, and fake orders without logistics in real time. The model is iteratively trained using task results, agent behavior data, and reward values ​​as samples. The weights are optimized using the gradient descent algorithm and then fed back to all agent nodes, simultaneously completing lightweight model distillation. The Agent's capabilities are evaluated in simulated scenarios after iteration. If the Agent fails to meet the standards, it is retrained with parameter tuning. If it meets the standards, the task loop is completed.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The central neural network host and system of this invention achieve full-chain domestic production and independent control. The core hardware and software all adopt domestic solutions, breaking through external technical limitations and ensuring the security of AI training and application. The heterogeneous computing array on the host and the distributed collaborative training mode improve the training efficiency of the super agent cluster, and the high utilization rate of host computing power greatly improves the utilization efficiency of computing power. The hardware and software dual risk control system improves the interception rate of fraudulent activities, eliminates fraudulent activities such as order brushing and cash-out from the source, significantly reduces the economic losses of the platform, and ensures the compliant closed-loop issuance and circulation of computing power vouchers; The Super Agent cluster supports seamless cross-scenario collaboration and can be scaled up for various scenarios such as physical businesses, digital employees, and OPC startups, providing one-stop intelligent services for different entities. A data compliance circulation system has been built to achieve data anonymization, listing and trading, and full-chain traceability, enabling the commercial value of training data to be effectively realized and forming a value closed loop of the AI ​​ecosystem; The system adopts a microservice architecture and supports elastic scaling. The host can achieve stable operation 24 / 7, with a low equipment failure rate. It is compatible with tens of thousands of agents concurrently scheduling and has strong practicality and scalability. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the host architecture of the present invention; Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0016] This embodiment addresses the composite scenario of a new retail ecosystem for physical businesses and one-person company entrepreneurship services. It details the entire implementation plan for the central neural network host and system, including hardware deployment, software configuration, training process, risk control execution, circulation of computing power vouchers, and monetization of data value. In this embodiment, the system is adapted to the concurrent scheduling of tens of thousands of super agents, covering three major scenarios: offline retail stores, online e-commerce stores, and individual entrepreneurial OPCs, achieving full-link autonomous control from agent training and deployment to operation.

[0017] I. Implementation Environment and Basic Configuration (a) Hardware deployment environment This embodiment adopts a high-density rack-mounted data center deployment. The central neural network host is deployed on the main node of the core data center, and the super agent cluster is distributed and deployed on edge nodes (local servers in offline stores and lightweight computing nodes in the cloud). The main node and edge nodes achieve microsecond-level communication through a high-speed interconnection network. The overall hardware environment meets the requirements of domestic production, and the core chips, computing units and storage devices all use domestically produced and controllable products.

[0018] (II) System Basic Parameter Configuration 1. Central Neural Network Host: Supports high-speed interconnection via PCIe 5.0, CXL 2.0, and 100G optical ports; total peak computing power of heterogeneous computing array ≥10PFlops; total capacity of global shared storage unit ≥100TB; TEE encryption module adopts the national cryptographic SM4 algorithm. 2. Super Agent Cluster: Initial deployment of 10,000 agents, including 3,000 digital store manager agents, 6,000 digital employee agents, and 1,000 OPC-specific agents, supporting dynamic elastic scaling. 3. Computing Power Coupon System: 1 unit of computing power coupon corresponds to 1 yuan of real consumption data value. The minimum redemption unit of computing power coupon is 10 units. The redeemed goods / services cover four major categories: physical goods, cloud services, entrepreneurship training, and computing power services. 4. Risk control thresholds: A single device with a daily transaction frequency of ≥20 times will trigger an alert; the same account with ≥15 cross-scenario transactions without fulfillment vouchers will trigger an interception; and the response delay for identifying fraudulent orders will be ≤100ms.

[0019] II. Specific Implementation of the Central Neural Network Host Hardware Module In this embodiment, the host is a 4U standard rack unit. Each module is physically integrated, independently powered, and supports hot-swapping. The specific module implementation details are as follows: 1. Heterogeneous computing array: integrates multi-core CPU, training-level GPU, edge inference NPU, and intelligent scheduling DPU. The CPU is responsible for system scheduling and logical operations, the GPU undertakes distributed training and inference of large models, the NPU is used for lightweight computing of edge agents, and the DPU realizes computing power offloading and network acceleration. Each unit achieves parallel computing through an internal high-speed bus, supporting the simultaneous training of tens of thousands of agents. 2. Central Neural Network Chip: Built-in swarm intelligence routing algorithm, task decomposition engine, agent scheduling model and evolutionary algorithm module, chip main frequency ≥2.5GHz, supports millisecond-level scheduling of agent clusters, and can update agent communication topology in real time; 3. Global shared storage unit: It is divided into a cache area and a persistent storage area, both of which are encrypted based on the TEE trusted execution environment. The cache area stores the Agent weights and task execution status of real-time training, while the persistent storage area stores historical task graphs, Agent behavior data, and computing power voucher circulation records. The stored data supports second-level backup and off-site disaster recovery. 4. High-speed interconnection interface: Configured with 16 PCIe 5.0 interfaces, 8 CXL 2.0 interfaces, and 12 100G optical ports. The master node and edge nodes are directly connected via 10 Gigabit fiber optic cables. Edge agents achieve computing power pooling and data sharing through CXL interfaces. The communication latency is controlled within 50 microseconds to meet the low latency requirements of cluster collaborative decision-making. 5. Risk control and computing power confirmation unit: This is a hardware-level independent module that integrates an encryption chip and a behavior recognition chip. It has a built-in abnormal behavior feature library and can directly perform hardware-level verification on transaction data, device data, and account data without going through the software layer. After identifying anomalies, it can directly block transactions and trigger account risk control. 6. Power and thermal management module: It adopts redundant dual power supply and supports 380V industrial power input; the heat dissipation adopts liquid cooling + air cooling composite solution. Liquid cooling is responsible for the heat dissipation of high heat units such as GPU and NPU, while air cooling is responsible for the remaining modules. The host operating temperature is controlled at 25-40℃, and it supports 7×24 hours of high-density rack-mounted stable operation.

[0020] III. Specific Implementation of the Super Agent Cluster System Software Module The system software layer adopts a microservice architecture, with each module decoupled and supporting independent upgrades. The central neural network host serves as the core brain of the system, connecting the super agent cluster layer, the swarm intelligence training platform, the computing power voucher rights confirmation and redemption system, the risk control and anti-fraud engine, and the data circulation and value realization module. The implementation details of each module are as follows: (a) Super Agent Cluster Layer Each agent is fine-tuned and trained based on a large model foundation, possessing scenario-based autonomous decision-making capabilities and supporting cross-agent collaborative interaction. Specific implementation types include: 1. Digital Store Manager Agent: Adapted to offline supermarkets, convenience stores, fresh food stores and other physical stores, it has full-process operation capabilities such as merchandise display planning, inventory management, customer flow analysis, promotional activity planning, and cashier settlement. It can connect to store POS machines, inventory management systems, customer flow cameras and other hardware to collect store operation data in real time and feed it back to the host. 2. Digital Employee Agent: Divided into four subcategories: content creation, intelligent customer service, data statistics, and financial accounting. The content creation agent can generate product copy and short video scripts. The intelligent customer service agent supports 24 / 7 consultation and response through multiple channels. The data statistics agent can automatically generate operational reports. The financial accounting agent is adapted to the simple accounting and tax filing needs of individual businesses and micro-enterprises. 3. OPC Dedicated Agent: Customized for one-person company entrepreneurs, integrating capabilities such as business registration, qualification processing, supply chain connection, customer development, and compliance risk control. It can formulate personalized entrepreneurial plans based on the entrepreneur's industry, capital, and resources, and link with the digital employee agent to provide OPC with one-stop intelligent entrepreneurial services.

[0021] (II) Swarm Intelligence Training Platform The platform provides full lifecycle training capabilities for the super agent cluster, employing a combined training mode of reinforcement learning, federated learning, and model distillation to avoid data silos and improve training efficiency. Specific functional implementation includes: 1. Task Distribution Module: Receives task instructions from the central neural network host and dynamically distributes subtasks to the corresponding agents based on the agent's scenario adaptability, computing load, and historical execution efficiency. It supports setting task priorities (core operation tasks > training and optimization tasks > data collection tasks). 2. Reinforcement Learning Module: Establish a reinforcement learning mechanism of "task execution - result feedback - reward and punishment", using the agent's task completion rate, operational efficiency, and data validity as evaluation indicators. If the indicators are met, the agent's weight is increased; if the indicators are not met, the weight is decreased and the agent is retrained. 3. Reward Mechanism Module: Linked with the computing power voucher rights confirmation system, computing power vouchers are issued to corresponding merchants / OPCs based on the real consumption data and fulfillment data collected by the Agent. No vouchers are issued if there is no real data or the data is invalid. 4. Model Distillation Module: Distills the knowledge of the large model trained on the host into a lightweight model, which is then deployed to the edge agent nodes to reduce the computing power consumption of the edge agents while ensuring the decision-making ability of the agents. After distillation, the model size is compressed to 1 / 10 of the original model, thereby improving the inference speed.

[0022] (III) Computing Power Voucher Confirmation and Redemption System The system is built on blockchain technology to realize the trusted generation, confirmation, circulation, and exchange of computing power vouchers. The entire process is on-chain and auditable. Specific implementation rules are as follows: 1. Computing Power Voucher Generation: The system automatically generates computing power vouchers only when there is a genuine consumption behavior, a genuine fulfillment behavior, and a genuine data upload. The generation record is uploaded to the blockchain in real time and includes information such as the generating entity, the generation time, and the corresponding real data number. It cannot be tampered with. 2. Computing Power Voucher Ownership Confirmation: Ownership of computing power vouchers belongs to the merchants / OPCs that generate real data. The vouchers are bound to device fingerprints, account information and blockchain addresses, supporting ownership inquiry and confirmation verification, and prohibiting illegal transfer by third parties. 3. Computing Power Voucher Redemption: Establish an integrated online and offline redemption platform. Redemption categories include physical goods, cloud services, startup services, and computing power services. After the redemption application is submitted, the system verifies the validity of the computing power voucher. Once the verification is approved, the redemption will be fulfilled within 24 hours, and the redemption record will be uploaded to the blockchain. 4. Transfer of computing power vouchers: Supports the free transfer of computing power vouchers between compliant entities, prohibits paid buying and selling and illegal monetization. The transfer process must be verified by the risk control and anti-fraud engine, and abnormal transfers will be blocked.

[0023] (iv) Risk control and anti-fraud engine The engine employs a dual risk control mode of hardware verification and software analysis. The hardware layer is implemented by the host's risk control and computing power rights confirmation unit, while the software layer consists of five modules that intercept fraudulent activities in real time. Specific implementation details are as follows: 1. Device fingerprint and behavior profile module: Generates a unique device fingerprint (including hardware information, network information, and operating habits) for each terminal device, and establishes a behavior profile for each account. When the device / account behavior deviates from the profile threshold, an alert is triggered. 2. Transaction Link Verification Module: Performs full-link data verification on each transaction, including payment vouchers, order information, merchant information, and fund flow. If any verification item is missing or the data is inconsistent, the transaction is judged as abnormal. 3. Logistics / Service Fulfillment Verification Module: Connects to mainstream logistics platforms and service supervision platforms to verify the logistics trajectory of goods and service completion certificates. Orders without fulfillment certificates are judged as fraudulent. 4. Cash-out account identification and blocking module: Built-in cash-out behavior feature database (such as low-in, high-out, cross-account mutual brushing, frequent redemption of high-value goods). For identified cash-out accounts, the power to generate and redeem computing power coupons will be blocked immediately. In serious cases, the account will be frozen and reported to the platform management terminal. 5. Loss Calculation Module: Real-time statistics of platform losses caused by fraudulent activities, including payment fees, taxes, computing power voucher rewards, and platform operating costs, generating loss calculation reports to provide data support for the platform to optimize subsidy strategies and adjust risk control thresholds.

[0024] (v) Data Circulation and Value Realization Module The module is built on a data compliance listing and trading model, connecting to domestic compliant data trading venues to achieve data anonymization, listing, trading, settlement, and traceability for agent-collected data, ensuring the legality and security of data circulation. Specific implementation details are as follows: 1. Data anonymization: The raw data collected by the Agent is anonymized to remove sensitive content such as personal identity information and core business information of merchants, and only the anonymized behavioral data, operational data and consumption data are retained. The anonymized data does not affect its commercial value. 2. Data listing: After the data has been de-identified and verified by the host, it is listed on a compliant data trading platform, with information such as data type, data volume, collection scenario, and data value marked. The listing information is synchronized to the blockchain in real time. 3. Data Trading: Supports compliant entities such as enterprises and institutions to participate in data trading. It adopts an auction trading model. After the transaction is completed, the module completes the data delivery and fund settlement. The settlement funds are distributed to the data producer and the platform in proportion. 4. Data traceability: The entire process of data collection, de-identification, listing, transaction, and delivery information is recorded. All records are on the blockchain and can be checked, realizing full-chain traceability of data from generation to monetization and ensuring the traceability of data transactions.

[0025] IV. Specific Implementation of the Central Neural Network Host Training Process In this embodiment, the host adopts a distributed collaborative training mode to perform unified training and iteration on the super agent cluster. The training process is closed-loop and supports real-time optimization. The specific steps are as follows: Step 1: Task Reception and Decomposition The host receives scenario-based training and operation tasks from the platform management terminal. Through the task decomposition engine of the central neural network chip, the total task is decomposed into multiple sub-tasks, and execution requirements, priorities and completion deadlines are set for each sub-task.

[0026] Step 2: Dynamically allocate subtasks The host dynamically allocates subtasks to the corresponding Agent nodes based on the scenario adaptability of the super Agent cluster, real-time computing load, and historical execution efficiency through a swarm intelligence routing algorithm. For high-priority core tasks, they are preferentially allocated to Agents with sufficient computing power and high execution efficiency. For edge data collection tasks, they are allocated to the nearest edge Agent node to reduce data transmission latency.

[0027] Step 3: Establishing the inter-agent communication topology The host establishes a dynamic communication topology network for the agents participating in task execution, designating a main agent and cooperating agents. The main agent is responsible for task coordination and result aggregation, while the cooperating agents are responsible for sub-task execution and data uploading. Agents communicate point-to-point through the CXL interface and 10 Gigabit optical port. The host monitors the communication topology status in real time. When an agent node fails, the host automatically reassigns tasks and adjusts the topology to ensure continuous task execution.

[0028] Step 4: Distributed Task Execution and Data Acquisition Each agent performs distributed execution according to the assigned sub-tasks, while collecting contextual data in real time. After local lightweight verification, the collected data is uploaded to the global shared storage unit of the host. The hardware-level risk control unit performs real-time verification on the uploaded data, filtering out invalid and false data.

[0029] Step 5: Collaborative Decision Making and Result Aggregation The main agent collects the subtask execution results from each collaborating agent and makes collaborative decisions based on the host's preset decision model. If the decision result does not meet the task requirements, the main agent sends feedback information to the collaborating agents and re-executes the subtask. If the decision result meets the requirements, the main agent uploads the summarized task results to the host. The host performs unified aggregation and verification of the results and generates a task execution report.

[0030] Step 6: Reward Function Construction and Computing Power Voucher Distribution The host constructs a reinforcement learning reward function based on real and valid behavioral data. The core reward indicators are data validity, task completion rate, and operational efficiency. The reward value for each Agent is calculated according to the reward function. The host also links with the computing power voucher rights confirmation and redemption system to issue the corresponding number of computing power vouchers to the data producers (merchants / OPCs). Computing power vouchers are only issued for valid data. No reward is given if there is no valid data.

[0031] Step 7: Real-time interception and risk control of abnormal behavior Throughout the entire process of task execution and data collection, the host's risk control and computing power confirmation unit, along with the risk control and anti-fraud engine, achieve dual real-time risk control: for abnormal orders, high-frequency transactions on the same device / account, fake orders without logistics, and cash-out through order brushing, the hardware layer directly blocks them, and the software layer triggers account warnings / bans. At the same time, abnormal behavior records are uploaded to the blockchain and incorporated into the device / account behavior profile.

[0032] Step 8: Model Iteration and Weight Backpropagation The host uses task execution results, Agent behavior data, and reward values ​​as training samples to iteratively train the basic model of the super Agent. The model weights are optimized using the gradient descent algorithm. After training, the updated model weights are sent back to all Agent nodes. At the same time, the large model weights are distilled into lightweight weights that can be adapted to edge Agents through model distillation technology, thereby improving the overall capabilities of the Agent cluster. The iterated model weights are stored in a globally shared storage unit, and TEE encryption is used to prevent tampering.

[0033] Step 9: Evaluation of Training Results and Task Closure The host performs capability assessment on the Agent after model iteration. It tests the Agent's decision-making ability, operational efficiency, and collaborative ability through simulated scenario tasks. If the assessment meets the standard, the training is completed and a task loop is formed; if the assessment does not meet the standard, it returns to step 2, reallocates sub-tasks, and adjusts training parameters until the assessment meets the standard.

[0034] V. System Overall Operation and Maintenance Implementation (I) Daily Operations 1. Host side: Dedicated operation and maintenance personnel monitor the host's operating status 24 / 7, and view the Agent cluster training progress, task execution status, and risk control interception records in real time through the management terminal; 2. Agent side: Edge agent nodes operate automatically without manual intervention. The host only issues an alarm when a node fails, and maintenance personnel then conduct on-site troubleshooting. 3. Computing Power Voucher Terminal: Real-time monitoring of the generation, circulation, and redemption of computing power vouchers, with key verification of large-amount redemptions and high-frequency circulation to ensure a closed loop in the computing power voucher system; 4. Data side: Regularly organize and list the anonymized data, track the progress of data transactions, and complete fund settlement and profit distribution.

[0035] (II) System Maintenance 1. Hardware maintenance: All modules of the host support hot-swapping, and faulty modules can be replaced in real time. A comprehensive hardware inspection is carried out every month, and a hardware upgrade is carried out every year. The hardware of the edge Agent nodes is regularly inspected, and aging equipment is replaced in a timely manner. 2. Software Maintenance: The system software layer adopts a microservice architecture, and each module can be upgraded independently. A minor version update is performed weekly, and a major version update is performed quarterly. 3. Data maintenance: Daily backup of data in the global shared storage unit, off-site disaster recovery storage, annual organization and archiving of historical data, deletion of invalid data, and release of storage space; 4. Risk control maintenance: The abnormal behavior feature library of the risk control and anti-fraud engine is updated monthly, and the risk control threshold is adjusted according to the platform's operation to improve the accuracy of fraudulent behavior identification.

[0036] VI. Verification of Implementation Results After six months of deployment and operation, the system in this embodiment achieved the following technical effects through actual testing: 1. Training efficiency and computing power utilization: The training efficiency of the super agent cluster is 8 times higher than that of the traditional single agent training method, and the host computing power utilization is stable at over 90%, far exceeding the preset target of ≥85%. 2. Cross-scenario collaboration capability: Digital store manager agent, digital employee agent, and OPC exclusive agent can achieve seamless cross-scenario collaboration. For example, the OPC exclusive agent can work with the digital employee agent to complete entrepreneurial copywriting and intelligent customer service, and work with the digital store manager agent to complete offline store operations, supporting the large-scale implementation of physical businesses and one-person companies, and has served more than 5,000 merchants / OPCs. 3. Risk control interception effect: The hardware + software dual risk control mode achieves a fraudulent behavior interception rate of ≥99.5%, far exceeding the preset target of ≥99.2%. The platform's losses caused by order brushing, cash-out, and self-trading are reduced by more than 98%, and the compliance issuance rate of computing power coupons is 100%. 4. Data circulation and value realization: More than 1,000 batches of anonymized data have been listed and traded, generating more than 5 million yuan in data transaction revenue, realizing the compliant circulation and commercial value realization of data; 5. System stability: The central neural network host operates steadily 24 / 7 with a device failure rate of ≤0.1% and the agent cluster concurrent scheduling response latency of ≤100ms, meeting the needs of large-scale operation in multiple scenarios.

[0037] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A central neural network host for professionally training super agent clusters, characterized in that, include: Heterogeneous computing array: integrates multi-core CPU, training-grade GPU, edge inference NPU, and intelligent scheduling DPU to support distributed training and inference parallelism of large models; Central neural network chip: Built-in swarm intelligence routing algorithm, task decomposition engine, agent scheduling model and evolutionary algorithm module to achieve millisecond-level scheduling of agent clusters; Global shared storage unit: includes DDR5 high-speed cache and solid-state persistent storage area, TEE trusted execution environment encryption based on national cryptographic SM4 algorithm, supports second-level backup and off-site disaster recovery; High-speed interconnect interfaces: configured with PCIe 5.0, CXL, and 10 Gigabit optical ports to enable microsecond-level communication between the master node and edge nodes; Risk control and computing power rights confirmation unit: a hardware-level independent module that integrates encryption and behavior recognition chips, and has a built-in abnormal behavior feature library to achieve hardware-level data verification and real-time blocking of abnormal behavior; Power and thermal management module: It adopts a redundant dual power supply and a liquid cooling + air cooling composite heat dissipation solution to support 24 / 7 high-density rack-mounted steady-state operation.

2. A central neural network system for professionally training super agent clusters, characterized in that, include: Central neural network host; Super Agent Cluster Layer: Includes digital store managers, digital employees, and OPC-specific agents. Initial deployment supports tens of thousands of concurrent users and can be dynamically scaled up or down. Each agent is finely trained based on a large model base. Swarm intelligence training platform: It adopts a training mode that combines reinforcement learning, federated learning and model distillation, and provides capabilities for task distribution, hierarchical rewards and lightweight model deployment; Computing power voucher ownership confirmation and redemption system: Built on a domestic consortium blockchain, it realizes the reliable generation, ownership confirmation, circulation and redemption of computing power vouchers, and the whole process is on-chain and auditable; Risk control and anti-fraud engine: It adopts a dual risk control mode of hardware verification + software analysis to intercept fraudulent transactions, self-trading, cash-out accounts, and low-entry-high-exit behaviors in real time; Data circulation and value realization module: Connects to compliant data trading venues to achieve data anonymization, listing, trading, settlement and full-chain traceability.

3. The system according to claim 2, characterized in that, In the super agent cluster layer, the digital store manager agent is adapted to offline physical stores such as supermarkets, convenience stores, and fresh food stores. It can connect to hardware such as store POS machines, inventory management systems, and customer flow cameras, and has full-process operation capabilities such as product display planning, inventory management, customer flow analysis, promotion planning, and cashier settlement.

4. The system according to claim 2, characterized in that, In the super agent cluster layer, the digital employee agent is divided into four subcategories: content creation, intelligent customer service, data statistics, and financial accounting. It supports multi-channel 24 / 7 operation and is adapted to the content production, customer service, data management, and financial accounting needs of individual businesses and micro-enterprises.

5. The system according to claim 2, characterized in that, In the super agent cluster layer, the OPC exclusive agent is customized for one-person company entrepreneurs, integrating business registration, qualification processing, supply chain connection, customer development, and compliance risk control capabilities. It can seamlessly collaborate with digital store manager and digital employee agents across scenarios to provide one-stop intelligent entrepreneurship services.

6. The system according to claim 2, characterized in that, The model distillation module of the swarm intelligence training platform can distill large model knowledge on the host side into a lightweight model, compressing the model size to 1 / 10 of the original model, increasing the inference speed by 5 times, realizing lightweight deployment of edge agent nodes, and reducing computing power consumption.

7. The system according to claim 2, characterized in that, In the computing power voucher confirmation and redemption system, computing power vouchers are generated only from real consumption, real fulfillment, and real and valid data. One unit of computing power voucher corresponds to 1 yuan of real consumption data value. The minimum redemption unit is 10 units. Redemption categories include physical goods, cloud services, entrepreneurship training, and computing power services. Free transfer between compliant entities is supported, but paid buying and selling and illegal monetization are prohibited.

8. The system according to claim 2, characterized in that, The risk control and anti-fraud engine includes a device fingerprint and behavior profiling module, a transaction link verification module, a logistics / service fulfillment verification module, a cash-out account identification and blocking module, and a loss accounting module. The loss accounting module can calculate platform losses such as payment fees, taxes, rewards, and operating costs caused by fraudulent behavior in real time and generate loss accounting reports.

9. The system according to claim 2, characterized in that, The data circulation and value realization module removes sensitive information from the raw data collected by the Agent. After the data is desensitized, it is verified by the host and then listed in a compliant venue. It adopts an auction trading model. After the transaction is completed, the data is delivered and the funds are settled. The settlement funds are distributed to the data producer and the platform in proportion.

10. The host computer according to claim 1 or the system according to claim 2, characterized in that, The host executes the distributed collaborative training process of the super agent cluster, including: Receive scenario-based training and operation tasks, and break them down into sub-tasks with execution requirements, priorities, and completion deadlines; Subtasks are dynamically allocated based on Agent scenario adaptability, computing load, and historical execution efficiency, with high-priority core tasks being prioritized for allocation to high-quality Agent nodes. Establish a dynamic communication topology network for agents participating in the task, designate a master / cooperating agent, and automatically reassign tasks and adjust the topology when a node fails. A reward function is constructed based on real and valid behavioral data, with data validity, task completion rate, and operational efficiency as the core, and computing power vouchers are issued in conjunction with the computing power voucher system. The system uses both hardware and software to intercept abnormal orders, high-frequency transactions from the same device / account, and fake orders without logistics in real time. The model is iteratively trained using task results, agent behavior data, and reward values ​​as samples. The weights are optimized using the gradient descent algorithm and then fed back to all agent nodes, simultaneously completing lightweight model distillation. The Agent's capabilities are evaluated in simulated scenarios after iteration. If the Agent fails to meet the standards, it is retrained with parameter tuning. If it meets the standards, the task loop is completed.

Images