Big data driven intelligent information resource management platform

Abstract

The invention relates to the technical field of industrial data scheduling and intelligent manufacturing, and provides a big-data-driven intelligent information resource management platform which adopts a closed-loop architecture composed of five modules, namely a data perception and characterization module, a knowledge cognition and constraint modeling module, an online dynamic optimization decision module, a resource execution and coordination module and a meta-cognition and evolution module. Real-time quantitative evaluation of data quality is realized through a triple quality vector; automatically converting the text business rule into a computable mathematical constraint and knowledge graph; carrying out dynamic scheduling decision making by adopting a knowledge enhancement reinforcement learning agent fused with a double reward mechanism; node-level real-time fine tuning is realized through an edge actuator; and closed-loop self-evolution of the scheduling model and the knowledge base is realized by means of a meta-cognitive mechanism. According to the method, the scheduling efficiency, the business compliance and the system adaptive capacity are cooperatively improved, and the method is suitable for industrial quality inspection data scheduling scenes with high real-time performance and high compliance requirements.

Description

Technical Field

[0001] This invention relates to the field of industrial data scheduling and intelligent manufacturing technology, specifically to a big data-driven intelligent information resource management platform. Background Technology

[0002] In the current process of intelligent transformation in industrial manufacturing, the efficient collaborative scheduling of massive amounts of heterogeneous data from multiple sources (such as real-time video, equipment status, and production work orders) has become a key challenge. Traditional scheduling systems typically employ rule engines or static optimization algorithms, which suffer from three prominent technical problems: First, the data from multiple sources has inconsistent formats and varying quality, lacking a unified quantitative evaluation system, leading to unreliable scheduling inputs. Second, business rules (such as safety compliance and energy efficiency constraints) are mostly in text or configuration form, making it difficult to directly integrate them into real-time decision-making models, resulting in a trade-off between scheduling efficiency and business compliance. Third, the systems are mostly designed in an open-loop manner, unable to dynamically optimize and update knowledge based on execution feedback, lacking adaptability to complex disturbances, and prone to performance degradation over long-term operation.

[0003] Therefore, a big data-driven intelligent information resource management platform is proposed. First, it achieves unified access to multi-source data through multi-protocol adaptation and spatiotemporal tagging, and innovatively proposes a triplet quality vector to quantify data quality in real time. Second, it automatically transforms text rules into computable mathematical constraints and structured knowledge graphs, and achieves distributed decomposition of global constraints through a hierarchical constraint propagation mechanism. Third, it adopts a knowledge-enhanced dual-reward reinforcement learning agent to simultaneously optimize efficiency and compliance goals in scheduling decisions. At the same time, it relies on edge executors to achieve real-time node-level fine-tuning to ensure accurate execution of instructions. Finally, it achieves model self-optimization and dynamic updating of the knowledge base through a metacognitive module, enabling the platform to have continuous evolution capabilities. Summary of the Invention

[0004] To address the technical problems in existing industrial scheduling systems, such as uncontrollable multi-source data quality, difficulty in integrating business rules into real-time decision-making, and lack of adaptive optimization capabilities, this invention provides a big data-driven intelligent information resource management platform.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution: a big data-driven intelligent information resource management platform, comprising:

[0006] The data perception and representation module is used to access multi-source heterogeneous quality inspection data through multiple preset protocol adapters, synchronize the timestamps of the accessed data to a unified time server, and attach spatiotemporal location tags to each data entry based on a preset resource directory, thereby calculating the triplet quality vector of each data stream in real time. , For completeness, For timeliness, , where is the confidence level.

[0007] The knowledge cognition and constraint modeling module, connected to the data perception and representation module, is used to generate directly calculable mathematical constraints by parsing the text of quality inspection business rules, and to construct an industrial knowledge graph based on historical scheduling data; and to distribute and decompose global resource constraints to each scheduling sub-unit using a hierarchical constraint propagation network.

[0008] The online dynamic optimization decision module connects the data perception and representation module and the knowledge cognition and constraint modeling module. It is used to construct a comprehensive state vector based on the quality vector, mathematical constraints and knowledge graph, and input it into a deep reinforcement learning network model enhanced with knowledge graph information to output scheduling actions.

[0009] The resource execution and coordination module, connected to the online dynamic optimization decision module, is used to convert the scheduling actions into commands that can be executed at the underlying level and to deploy edge executors on each computing node to execute predefined micro-actions based on local context information.

[0010] The metacognition and evolution module connects the resource execution and coordination module, the online dynamic optimization decision-making module, and the knowledge cognition and constraint modeling module. It is used to collect scheduling execution results and trigger the fine-tuning training of the deep reinforcement learning network model and dynamically optimize the industrial knowledge graph.

[0011] Furthermore, the data perception and representation module calculates the timeliness in the triplet quality vector. When this happens, perform the following steps:

[0012] Acquire the delay from data generation to data access in this module. and preset timeliness qualification threshold .

[0013] According to the formula Calculate the timeliness score, where The preset attenuation coefficient, This represents the time difference between the data generation and arrival at the module.

[0014] Furthermore, when the knowledge recognition and constraint modeling module generates directly computable mathematical constraints by parsing business rule text, it performs the following steps:

[0015] The rule text is parsed and keywords are identified using a deterministic finite automaton.

[0016] The rule semantics corresponding to the identified mandatory keywords are mapped to the adaptation type in the hard constraints of location and resource type.

[0017] The semantic rules corresponding to the identified priority keywords are mapped to the task delay weight coefficients in the objective function of the decision model.

[0018] Furthermore, when the knowledge cognition and constraint modeling module decomposes global resource constraints using a hierarchical constraint propagation network, it performs the following steps:

[0019] Initialize the Lagrange multipliers used to coordinate global constraints and local objectives. And broadcast it to each scheduling subunit.

[0020] Each scheduling subunit has a local cost function Independent optimization is performed below, among which For local costs, For local variables related to global constraints, and the optimized Report it to higher authorities.

[0021] The central node reports data from each sub-unit. Calculate global variables It is then compared with the global constraint threshold and dynamically adjusted. The value of .

[0022] Iteratively execute the above steps until... The global constraints are satisfied.

[0023] Furthermore, the deep reinforcement learning network model in the online dynamic optimization decision module is trained and used for decision-making using a dual reward function, which is:

[0024] ,

[0025] in, For the total reward, Efficiency rewards are calculated based on average task latency and total resource idle rate; The calculation formula for compliance rewards is as follows:

[0026] ,

[0027] in For the first The penalty weight of the constraint, For the first The degree of violation of the constraint; and when hour, According to the formula To achieve adaptive increase, where This is the preset magnification factor. For indicator functions, when hour ,when hour .

[0028] Furthermore, the edge executor in the resource execution and coordination module periodically performs the following operations:

[0029] Read the context information of the local computing node, which includes one or more of the following: CPU utilization, memory usage, core temperature.

[0030] Based on the Thompson sampling algorithm, the micro-actions to be executed are selected from a set of predefined micro-actions for different node running scenarios.

[0031] After executing the selected micro-action, update the beta distribution parameter corresponding to the micro-action based on the positive feedback obtained. .

[0032] Furthermore, the conditions under which the metacognition and evolution module triggers the fine-tuning training of the deep reinforcement learning network model include at least one of the following: reaching a preset periodic trigger time; or detecting that the error between the model's predicted reward and the actual reward exceeds a preset threshold.

[0033] Furthermore, when performing model fine-tuning training, the metacognition and evolution module employs importance sampling technology to assign weights to historical interaction data sampled from the experience replay buffer.

[0034] Compared with existing technologies, this invention provides a big data-driven intelligent information resource management platform, which has the following beneficial effects:

[0035] 1. This solution achieves a fundamental shift in industrial scheduling systems from passive execution to proactive optimization by constructing a closed-loop self-evolving architecture encompassing perception, cognition, decision-making, execution, and evolution. Its core innovation lies in digitizing business knowledge into computable constraints and graphs, and integrating a dual-reward mechanism for reinforcement learning agents. This enables scheduling decisions to not only dynamically adapt to real-time operating conditions but also embed compliance logic, resolving the technical contradiction in traditional scheduling methods where efficiency optimization and rule adherence are difficult to coordinate.

[0036] 2. This solution employs a collaborative mechanism of hierarchical decomposition of global constraints and real-time fine-tuning at the edge, ensuring plant-wide compliance while empowering each computing node to autonomously respond to local disturbances. By combining macro-level precise guidance with micro-level autonomous adaptation, it effectively overcomes the dual shortcomings of sluggish response in centralized scheduling and global loss of control in distributed scheduling, thereby achieving simultaneous improvements in system stability, scheduling efficiency, and business compliance in complex industrial scenarios. Attached Figure Description

[0037] Figure 1 This is a diagram showing the overall system architecture and closed-loop data flow of the present invention.

[0038] Figure 2 This is a flowchart of the data perception and representation module of the present invention;

[0039] Figure 3 This is a flowchart of the knowledge cognition and constraint modeling module of the present invention;

[0040] Figure 4 This is a flowchart of the online dynamic optimization decision-making module of the present invention;

[0041] Figure 5 This is a flowchart of the resource execution and coordination module of the present invention;

[0042] Figure 6 This is a flowchart of the metacognition and evolution module of the present invention;

[0043] Figure 7 This is a flowchart illustrating the operation of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1:

[0046] Please see Figures 1 to 7 This invention relates to a big data-driven intelligent information resource management platform. This embodiment uses real-time video quality inspection data scheduling in an automotive parts manufacturing plant as an application scenario. The system includes: a data perception and representation module, a knowledge cognition and constraint modeling module, an online dynamic optimization decision-making module, a resource execution and coordination module, and a metacognition and evolution module. The data flow consists of standardized quality inspection data and quality vectors processed by the data perception and representation module, which are transmitted to the knowledge cognition and constraint modeling module and the online dynamic optimization decision-making module. The knowledge flow consists of computable mathematical constraints and knowledge graph information generated by the knowledge cognition and constraint modeling module, which are transmitted to the online dynamic optimization decision-making module. The feedback flow consists of the scheduling execution results from the resource execution and coordination module and the model optimization and knowledge base update results from the metacognition and evolution module. The former is transmitted to the metacognition and evolution module, and the latter is fed back to the knowledge cognition and constraint modeling module and the online dynamic optimization decision-making module. These five modules form a closed-loop collaborative technical architecture.

[0047] Module 1: Data Perception and Representation Module

[0048] The core function of this module is to enable standardized access to multi-source quality inspection data and quantitative assessment of data quality, providing subsequent modules with quality inspection data in a unified format and quantitative quality assessment results.

[0049] This module is responsible for standardizing all input data, providing a unified interface for decision-making. It is configured with multiple protocol adapters: an RTSP adapter for camera video streams, a Kafka adapter for MES system work order information, and an OPC UA adapter for production equipment status data. These adapters enable unified access to multi-source quality inspection data.

[0050] To ensure data consistency across time and space, the module establishes a unified time server for the entire plant, synchronizing the timestamps of all incoming data to this baseline. It also maintains a resource directory containing hierarchical relationships of "factory-workshop-production line-equipment," automatically attaching corresponding location tag information to each data packet based on its source address. For example, a video frame from "Robot A on Welding Workshop Line 3" has the tag information "Location Information: Factory 01 - Welding Workshop - Line 3 - Robot A, Timestamp: 1723456789.123".

[0051] Based on this, the module will also calculate the triplet quality vector for each data stream in real time, the expression of which is:

[0052]

[0053] In the formula, For completeness, For timeliness, For confidence levels, all characters are simplified single-letter characters. Integrity is evaluated based on a preset sampling frequency. If the video stream is preset to 30 frames per second, and 29 frames are received within the current 1-second window, then... .

[0054] The delay between the generation of timeliness calculation data and its arrival in this module. Set threshold (Timeliness compliance threshold), the scoring formula is:

[0055]

[0056] In the formula, The preset attenuation coefficient, , This refers to the time difference between the data generation and arrival at the module. If... ,but .

[0057] The confidence level is verified using rules, such as checking whether the average pixel value of the image frame is within a reasonable historical range [10, 250]. If it exceeds this range, then... .

[0058] The final output of this module includes the original data, spatiotemporal label information, and a quality vector. The standardized quality inspection data is synchronously transmitted to the knowledge cognition and constraint modeling module and the online dynamic optimization decision-making module, providing a standardized and quantitative evaluation data source for subsequent constraint modeling and scheduling decisions, and achieving the beneficial effects of unified multi-source data format and quantifiable data quality.

[0059] Module Two: Knowledge Cognition and Constraint Modeling

[0060] The core function of this module is to realize the digitization and computability of business knowledge, the hierarchical and precise decomposition of global constraints, and to provide directly callable mathematical constraints and knowledge graph information for scheduling decisions.

[0061] This module receives standardized quality inspection data and quality vectors output by the data perception and representation module, transforming textual or configured quality inspection business knowledge into mathematical constraints and knowledge graphs to provide compliance basis and optimization direction for the decision-making module. Its constraint formalization process begins with rule parsing. After inputting rule text, it is parsed and segmented, for example, "The inspection video of safety components (brake discs) must be processed on the local secure GPU cluster, and the analysis task priority is the highest." Then, a deterministic finite automaton identifies keywords: "must" is determined as a hard constraint (cannot be violated), and "highest priority" is determined as an optimization target weight item (affecting decision priority).

[0062] After keyword recognition is completed, the module further performs semantic mapping: "Safety component (brake disc)" is mapped to the data attribute "Data type = safety critical class"; "Must be processed in the local security zone GPU cluster" is mapped to the location hard constraint "Tasks derived from safety critical class data must satisfy 'location = local security zone and type = GPU' for allocable resources"; "Highest priority" is mapped to "The completion delay weight of this type of task in the objective function of the decision model". Set to a constant much larger than that of a normal task. ",in This is the task delay weighting coefficient. This is a high-priority dedicated weight constant.

[0063] In addition to constraint transformation, the module also constructs an industrial knowledge graph based on quality inspection business rules and historical scheduling data, and stores entities and relationships in a graph database to form a structured knowledge network. For example, it creates entity nodes "task T1, video stream V1, GPU server S1" and creates the relationship "T1 consumes V1, V1 is generated by L1, E1 is located in the local security zone". At the same time, it records historical statistical weights, such as "the average execution time of T1 type tasks on server S1 = 50ms", to provide historical experience support for decision-making.

[0064] For plant-wide global constraints (such as "total real-time power consumption ≤ 1000kW"), the module decomposes them into local constraints executable by each workshop sub-unit through a hierarchical constraint propagation network. The specific process is as follows: Initialize the Lagrange multipliers. (Constraining and coordinating multipliers to balance global constraints and local optimization objectives), broadcasting to each workshop scheduling subunit. Each sub-unit is optimized independently under its local cost function, the expression of which is:

[0065]

[0066] In the formula, For the first Sub-unit power consumption cost For the first The optimal power consumption of each sub-unit. Each workshop sub-unit will calculate the... The data is reported to the central node in real time, and the central node calculates the total power consumption.

[0067]

[0068] In the formula, This represents the total power consumption of the entire plant. This represents the total number of workshop scheduling subunits.

[0069] like but ,like but Repeat the above steps until the following conditions are met: It is stable for three iteration cycles, ultimately achieving a balance between distributed optimization and global compliance.

[0070] The final output of this module is computable mathematical constraints and structured industrial knowledge graph information, which are synchronously transmitted to the online dynamic optimization decision-making module. This achieves the beneficial effects of digitizing business knowledge and accurately decomposing global constraints, giving scheduling decisions a clear basis for business compliance.

[0071] Module 3: Online Dynamic Optimization Decision Module

[0072] The core function of this module is to combine the computable constraints of the knowledge cognition and constraint modeling module to build a dual reward mechanism that takes into account both efficiency and compliance, and to output efficient and compliant scheduling decisions through reinforcement learning agents.

[0073] This module is the core of scheduling decision-making. It integrates standardized quality inspection data and quality vectors from the data perception and representation module, and computable mathematical constraints and knowledge graph information from the knowledge cognition and constraint modeling module. It employs a knowledge-enhanced reinforcement learning model to achieve efficient and compliant scheduling decisions. At each decision-making moment... (Decision time node identifier, interval set as needed), the module will first construct a comprehensive state vector. (Comprehensive representation of the decision-making environment) fully reflects the current scheduling scenario.

[0074] Specifically, it includes three parts: characteristics of the task queue to be scheduled (task type, data volume, quality vector). The system includes a list of constraint identifiers, resource pool status (real-time utilization of computing nodes, memory, temperature, etc.), and constraint state vectors (total power consumption to limit ratio, backlog of high-priority tasks, and other constraint satisfaction status), providing comprehensive environmental awareness for agent decision-making.

[0075] The knowledge-enhanced double-reward deep reinforcement learning agent in the module adopts the Dueling Double DQN architecture. The unique feature of this architecture is that it splits into two parallel sub-networks after the last fully connected layer: a value stream network... ( (For the set of value stream parameters, evaluate the state-based value) and the dominant stream network. ( (Assessing the relative advantage of actions using the set of dominant flow parameters). The final action value calculation formula is:

[0076]

[0077] In the formula, For the total set of network parameters, To achieve the mean value, avoid bias and ensure the accuracy of action value assessment.

[0078] The core of the intelligent agent is the design of a dual reward function, expressed as:

[0079]

[0080] In the formula, For the total reward, Rewards for efficiency To achieve a balance between efficiency and compliance, a compliance reward is provided. The efficiency reward calculation formula is as follows:

[0081]

[0082] In the formula, To average task latency, The total resource idle rate represents the overall resource utilization rate; the lower the latency, the higher the utilization rate. The larger the value, the more directly it relates to the efficiency of scheduling.

[0083] The formula for calculating compliance rewards is as follows:

[0084]

[0085] In the formula, For the total number of binding clauses, For the first Constraint penalty weights For the first The degree of constraint violation, such as the amount of power consumption exceeding the limit. Its core innovation lies in the adaptive penalty weight, updated by the following formula:

[0086]

[0087] In the formula, This is the magnification factor, with a value of 0.5; For indicator functions, when hour ,when hour Once the constraint is violated, the penalty weight immediately increases, forcing the agent to correct the violation first.

[0088] After being pre-trained in a simulation based on a factory quality inspection data dataset, the intelligent agent is connected to the real scheduling system. When making decisions, it uses... As input, output the Q-values ​​of all scheduling actions, and select the action with the highest Q-value as the final scheduling action. Simultaneously, samples generated through interaction with the scheduling environment are also included. It will continuously store the experience replay buffer, where This is the current state. In order to perform the action, As a feedback reward, For the next state, the experience replay buffer uses a rolling storage rule to save the latest interaction samples for subsequent network fine-tuning, continuously improving the model's adaptability to actual scheduling scenarios.

[0089] The final output of this module is a specific scheduling action, which is passed to the resource execution and coordination module to achieve the beneficial effects of intelligent scheduling with knowledge constraints and optimization of both efficiency and compliance objectives.

[0090] Module 4: Resource Execution and Coordination

[0091] The core function of this module is to transform macro-level scheduling instructions into low-level executable commands, while using edge executors to achieve real-time fine-tuning of node disturbances, ensuring the accurate implementation of scheduling instructions and the stable operation of nodes.

[0092] This module receives scheduling actions from the online dynamic optimization decision-making module, transforms macro-level scheduling instructions into low-level executable commands, and handles micro-level disturbances at the computing node level to ensure stability and efficiency at the execution level. Upon receiving a scheduling action (such as "assign task T1 to server S1"), the module first converts it into a low-level executable command. Specifically, it issues an instruction to the Kubernetes cluster to schedule the container (Pod) carrying task T1 to node S1 and sets a network priority label to ensure the priority of task transmission and execution, guaranteeing the accurate implementation of macro-level scheduling instructions.

[0093] In addition to global instruction execution, to address real-time disturbances at the node level, edge executors based on context-based multi-armed automata are deployed locally on each computing node for millisecond-level local fine-tuning to adapt to the node's real-time operating status. The executors predefine a set of micro-actions for common node scenarios. The trigger conditions for each micro-action are exemplary conditions that can be set according to the actual scenario: Micro-action 1 = "Follow the central task order" (applicable to load balancing scenarios, exemplary judgment condition is node resource utilization in the 40%-60% range); Micro-action 2 = "Prioritize short tasks with a duration of <10ms" (applicable to queue backlog scenarios, exemplary judgment condition is task queue length exceeding a preset threshold of 50 or quality vector integrity). Micro-action 3 = "When the temperature is >80℃, computationally intensive tasks run at reduced frequency" (applicable to high-temperature node scenarios, the exemplary judgment condition is that the node core temperature exceeds 80℃).

[0094] Every The executor reads local context information (CPU utilization, memory usage, core temperature) and uses the Thompson sampling algorithm to balance exploration and utilization, where... For local decision intervals, millisecond-level response requirements are matched. Each micro-action corresponds to a beta prior distribution:

[0095]

[0096] In the formula, This is the sequence number of the micro-action. For the first The first parameter of the beta distribution of each micro-action For the first The second parameter of each micro-action beta distribution is initially set to 1, indicating an unbiased state. The actuator samples the reward value from each micro-action distribution based on the current context. (No. (Sampling benefit of individual micro-motions), select The execution of the largest micro-movements.

[0097] After execution, the executor will receive binary benefit feedback based on indicators such as whether queue latency is reduced and node temperature is controlled. (Profit feedback, 0 indicates no positive profit, 1 indicates positive profit), and update the parameters accordingly:

[0098]

[0099]

[0100] By continuously iterating and optimizing local decision-making, the stability of node operation is ensured. This module transmits the scheduling execution results to the metacognition and evolution module in real time, achieving the beneficial effects of precise implementation of scheduling instructions and real-time response to node disturbances.

[0101] Module 5: Metacognition and Evolution

[0102] The core function of this module is to realize the self-optimization iteration of the scheduling model and the dynamic updating of the business knowledge graph, thereby forming a closed-loop self-evolution capability of the scheduling system.

[0103] This module acts as the system's "self-optimization hub," receiving the scheduling and execution results from the resource execution and coordination module. It enables the system to self-monitor, learn from feedback, and iteratively optimize, ensuring long-term stable operation and continuous performance improvement. The module first records the complete trajectory of each round of decisions, including the current state. Execution of actions Predicted Rewards (Model-predicted reward), Actual reward (Actual reward generated), details of constraint violation (constraint identifier, degree of violation, scope of impact), and actual task completion time.

[0104] Based on these records, the module... and The difference is calculated to determine the prediction error and assess the model's accuracy. Simultaneously, compliance risk points are analyzed by constraining violation details, providing a precise basis for subsequent optimization. The module initiates fine-tuning training of the agent model according to exemplary dual trigger conditions: first, periodic triggering, starting once every 24 hours; second, threshold triggering, triggered immediately when the model's prediction error exceeds a preset threshold of 15%. During training, recently generated real-world interaction samples are used. An experience replay buffer is added to the agent, and an importance sampling technique is used to assign appropriate weights to old data to avoid excessive dilution of old data and prevent the "forgetting" of historical effective experience due to learning new data (i.e., catastrophic forgetting), thus ensuring the continuous stability of model performance.

[0105] In addition, the module will analyze long-term constraint violation records and scheduling effect data to achieve dynamic optimization of the industrial knowledge graph: if a constraint (such as "a certain type of task must use a specific algorithm library") is found to be frequently violated but does not cause business loss (such as the result accuracy does not decrease), a rule revision prompt will be sent to the administrator; if a specific scheduling experience (such as "use micro-action 2 when the context is under high load") is found to be effective in a certain scenario, it will be converted into an experience rule and updated to the knowledge graph.

[0106] This module updates the model fine-tuning results to the reinforcement learning agent of the online dynamic optimization decision module in real time, and feeds back the optimization results of the knowledge graph to the knowledge cognition and constraint modeling module, realizing the beneficial effects of self-optimization iteration of the scheduling model and dynamic updating of the knowledge base, enabling the entire scheduling system to have closed-loop self-evolution capability.

[0107] The operation process is as follows:

[0108] 1. Video stream V1 (belonging to the safety-critical data category) and work order data are connected to the data perception and representation module through the corresponding protocol adapter. After spatiotemporal alignment, tagging, and quality vector calculation, they are assigned spatiotemporal tag information and quality vectors. In the formula This is the quality vector corresponding to video stream V1, enabling the standardization and quantitative quality assessment of multi-source quality inspection data.

[0109] 2. The knowledge cognition and constraint modeling module receives the above standardized data, identifies the hard constraint associated with video stream V1 as "must be processed by the GPU in the local safe zone", and runs the HCPN algorithm through the hierarchical constraint propagation network. It calculates the power consumption budget of each resource in combination with the current state of the entire plant's resources, decomposes the global power consumption constraint into local constraints, and realizes the computability of business knowledge and the accurate decomposition of global constraints.

[0110] 3. The online dynamic optimization decision-making module integrates the task characteristics corresponding to video stream V1, the real-time operating status of each server, and the power consumption constraint status to construct the current comprehensive state vector. Reinforce learning agents to Using this as input, and combining hard constraints, the Q-values ​​of all possible scheduling actions are calculated, thus realizing scheduling decision modeling with knowledge constraints.

[0111] 4. Due to the hard constraints corresponding to video stream V1, a high penalty weight is set in the reward function. All actions that assign the task to resources outside the local security zone have extremely low Q values. The agent ultimately chooses the compliant action with the highest Q value—assigning the task to GPU server 02 in the local security zone—and outputs the optimal scheduling action that balances efficiency and compliance.

[0112] 5. The resource execution and coordination module then converts the above scheduling actions into executable commands for the Kubernetes cluster and sends them to GPU server 02. At the same time, the edge executor of this server reads its own high-load local context information (such as CPU utilization of 85% and memory usage of 70%), selects micro-action 2 (prioritizing short tasks) through Thompson sampling, and performs local rearrangement of the task queue issued by the center to achieve accurate implementation of scheduling instructions and real-time response to node disturbances.

[0113] 6. After the task is completed, the metacognition and evolution module records the actual time taken, server power consumption, and other execution data of the task, and then records the interaction data of this round. An experience replay buffer is added and fine-tuning training is initiated to optimize the parameters of the reinforcement learning agent. Since no constraint violations are detected, the industrial knowledge graph remains stable, enabling feedback optimization of the scheduling model and ensuring the system's self-evolution capability.

[0114] Through the above implementation methods, the present invention realizes intelligent, compliant and adaptive optimization of industrial data resource scheduling, effectively improving the scheduling efficiency and business compliance of industrial quality inspection data. At the same time, it has the closed-loop self-evolution capability of model self-optimization and dynamic knowledge updating, and is suitable for industrial quality inspection data scheduling scenarios with high real-time and high compliance requirements.

[0115] Finally, it should be noted that the above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Any obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A big data-driven intelligent information resource management platform, characterized in that: include: The data perception and representation module is used to access multi-source heterogeneous quality inspection data through multiple preset protocol adapters, synchronize the timestamps of the accessed data to a unified time server, and attach spatiotemporal location tags to each data entry based on a preset resource directory, thereby calculating the triplet quality vector of each data stream in real time. , For completeness, For timeliness, Confidence level; The knowledge cognition and constraint modeling module, connected to the data perception and representation module, is used to generate directly calculable mathematical constraints by parsing the quality inspection business rule text, and to construct an industrial knowledge graph based on historical scheduling data; and to distribute and decompose global resource constraints to each scheduling sub-unit using a hierarchical constraint propagation network. The online dynamic optimization decision module connects the data perception and representation module and the knowledge cognition and constraint modeling module. It is used to construct a comprehensive state vector based on the quality vector, mathematical constraints and knowledge graph, and input it into a deep reinforcement learning network model enhanced with knowledge graph information to output scheduling actions. The resource execution and coordination module, connected to the online dynamic optimization decision module, is used to convert the scheduling action into commands that can be executed at the underlying level and to deploy edge executors on each computing node to execute predefined micro-actions based on local context information. The metacognition and evolution module connects the resource execution and coordination module, the online dynamic optimization decision-making module, and the knowledge cognition and constraint modeling module. It is used to collect scheduling execution results and trigger the fine-tuning training of the deep reinforcement learning network model and dynamically optimize the industrial knowledge graph.

2. The big data-driven intelligent information resource management platform according to claim 1, characterized in that: The data perception and representation module calculates the timeliness in the triplet quality vector. When this happens, perform the following steps: Acquire the delay from data generation to data access in this module. and preset timeliness qualification threshold ; According to the formula Calculate the timeliness score, where The preset attenuation coefficient, This represents the time difference between the data generation and arrival at the module.

3. The big data-driven intelligent information resource management platform according to claim 1, characterized in that: When the knowledge recognition and constraint modeling module generates directly calculable mathematical constraints by parsing business rule text, it performs the following steps: Rule text parsing and keyword identification are performed using deterministic finite automata; The semantic mapping of the rules corresponding to the identified mandatory keywords is used as the adaptation type in the hard constraints of location and hard constraints of resource type. The semantic rules corresponding to the identified priority keywords are mapped to the task delay weight coefficients in the objective function of the decision model.

4. The big data-driven intelligent information resource management platform according to claim 1, characterized in that: When the knowledge recognition and constraint modeling module decomposes global resource constraints using a hierarchical constraint propagation network, it performs the following steps: Initialize the Lagrange multipliers used to coordinate global constraints and local objectives. And broadcast it to each scheduling subunit; Each scheduling subunit has a local cost function Independent optimization is performed below, among which For local costs, For local variables related to global constraints, and the optimized Report to higher authorities; The central node reports data from each sub-unit. Calculate global variables It is then compared with the global constraint threshold and dynamically adjusted. The value; Iteratively execute the above steps until... The global constraints are satisfied.

5. The big data-driven intelligent information resource management platform according to claim 1, characterized in that: The deep reinforcement learning network model in the online dynamic optimization decision module uses a dual reward function for training and decision-making. The dual reward function is as follows: ， in, For the total reward, Efficiency rewards are calculated based on average task latency and total resource idle rate; The calculation formula for compliance rewards is as follows: ， in For the first The penalty weight of the constraint, For the first The degree of violation of the constraint; and when hour, According to the formula To achieve adaptive increase, where This is the preset magnification factor. For indicator functions, when hour ,when hour .

6. The big data-driven intelligent information resource management platform according to claim 1, characterized in that: The edge executor in the resource execution and coordination module periodically performs the following operations: Read the context information of the local computing node, which includes one or more of the following: CPU utilization, memory usage, core temperature; Based on the Thompson sampling algorithm, the micro-actions to be executed are selected from a set of predefined micro-actions for different node running scenarios; After executing the selected micro-action, update the beta distribution parameter corresponding to the micro-action based on the positive feedback obtained. .

7. The big data-driven intelligent information resource management platform according to claim 1, characterized in that: The conditions under which the metacognition and evolution module triggers the fine-tuning training of the deep reinforcement learning network model include at least one of the following: reaching a preset periodic trigger time; The error between the model's predicted reward and the actual reward was detected to exceed a preset threshold.

8. The big data-driven intelligent information resource management platform according to claim 7, characterized in that: When performing model fine-tuning training, the metacognition and evolution module uses importance sampling technology to assign weights to historical interaction data sampled from the experience replay buffer.

Images