Intelligent algorithm center resource regulation system and method fusing multi-dimensional perception and physical information

Abstract

The application discloses a kind of fusion multi-dimensional perception and physical information's intelligent calculation center resource regulation system and method, belong to intelligent calculation center resource management technical field.System is unified by multi-dimensional perception module Resource task state view is constructed.Intelligent decision module is based on the view driving multi-agent reinforcement learning model to generate collaborative strategy;Its collaborative mechanism is based on generalized reactive temporal logic formalization, by the team strategy of synthesis is decomposed into synchronous reward automaton, to guide multi-agent distributed collaboration.Execution module is linked based on the fine energy efficiency control instruction of physical information neural network, realize resource global optimization.Digital twin module is responsible for strategy verification and model iteration.The application solves the problem of cross-domain collaborative temporal logic security guarantee and physical level fine energy efficiency optimization, significantly improves the autonomy, reliability and energy efficiency of system.

Description

Technical Field

[0001] This invention relates to the field of intelligent computing center resource management technology, specifically to an intelligent computing center resource control system and method that integrates multi-dimensional sensing and physical information. Background Technology

[0002] With the explosive growth of artificial intelligence and high-performance computing tasks, intelligent computing centers face the severe challenge of imbalanced utilization of multi-dimensional resources such as computing, storage, network, and energy. Existing dynamic scheduling systems mostly rely on threshold-based passive responses or optimization algorithms for single resource domains, lacking the ability to accurately characterize and coordinate the complex coupling relationships and global temporal constraints between multiple resource domains. This makes it difficult for the system to achieve global optimization of resource utilization and energy efficiency while ensuring task performance and service level agreements when facing sudden loads, complex task chains, and cross-domain interference, resulting in "resource control silos" and "energy efficiency bottlenecks."

[0003] While existing research has attempted to introduce resource knowledge graphs for resource relationship modeling or utilize multi-agent reinforcement learning for distributed decision-making, these methods inherently fall short in ensuring the security (e.g., avoiding deadlock) and liveness (e.g., task completion) of multi-agent collaborative strategies, failing to formally circumvent the risks of deadlock, livelock, or violation of critical runtime constraints. Furthermore, current solutions have not deeply integrated the underlying physical laws of cyber-physical systems (e.g., thermodynamics, power consumption models) into the scheduling decision-making closed loop, resulting in energy efficiency optimization remaining at a coarse-grained stage based on empirical formulas or statistical models, with limited accuracy and reliability. Therefore, how to construct an intelligent computing center resource autonomous management system that can simultaneously ensure cross-domain collaborative temporal security and achieve physical-level fine-grained energy efficiency control has become a core technical challenge urgently needing to be solved in this field. Summary of the Invention

[0004] The present invention aims to provide a resource regulation system and method for intelligent computing centers that integrates multi-dimensional perception and physical information, in order to solve the technical problems of existing technologies lacking the ability to guarantee the complex temporal logic constraints in the collaboration of multiple resource domains, and failing to achieve fine energy efficiency regulation based on physical laws.

[0005] This invention provides a resource control system for an intelligent computing center that integrates multi-dimensional sensing and physical information, characterized in that it includes:

[0006] The multi-dimensional perception and fusion module constructs a unified resource task status view based on the resource knowledge graph of the intelligent computing center, which is based on pre-built physical information.

[0007] The intelligent analysis and decision-making module is used to predict the load based on the resource task status view and drive the multi-agent reinforcement learning model to generate a collaborative regulation strategy; wherein, the collaborative mechanism of the multi-agent reinforcement learning model is based on the cross-domain collaborative task specification formally defined by generalized reactive temporal logic, and constructs a set of synchronous, decomposed reward automata by decomposing the synthesized high-level team strategy.

[0008] The coordinated control and execution module is used to execute the coordinated control strategy and output dynamic allocation instructions for computing, network, and storage resources, as well as linkage adjustment instructions for energy infrastructure.

[0009] The physical information fusion module is used to run an end-to-end prediction model based on a physical information neural network. This model takes a resource scheduling plan as input, fuses physical environment simulation data, and generates refined energy efficiency control instructions. These energy efficiency control instructions are used in conjunction with the adjustment instructions output by the collaborative regulation and execution module to achieve cross-layer dynamic resource optimization.

[0010] Preferably, the construction of the resource knowledge graph specifically includes: using cross-language patent text mining and semantic association rules to automatically discover and expand the entities and relationships between computing tasks, physical resources, and energy efficiency indicators, and establishing a dedicated resource knowledge graph for the intelligent computing center scheduling field through iterative learning.

[0011] Preferably, the time series prediction model adopts a model based on... The architecture of the time-series information multi-agent model models multi-dimensional time-series data into sequences of arbitrary length to extract long-range dependencies, and serves as a general pre-training basis for the multi-agent reinforcement learning model.

[0012] Preferably, in the multi-agent reinforcement learning model, the weight coefficients of the joint reward function used to evaluate the joint strategy of the agents are dynamically adjusted by a meta-reinforcement learning module. This module takes historical scheduling effect evaluation data as input, takes the long-term benefits generated by the joint reward function on the validation set as the optimization objective, and automatically learns and adjusts the weights to adapt to different business cycles.

[0013] Preferably, the real-time migration of computing instances is specifically executed by an intelligent migration decision engine. Based on the real-time and predicted data in the resource task status view, the engine calls the digital twin simulation and iteration module to perform millisecond-level forward-looking simulations of candidate migration schemes in the digital twin environment, predicting the comprehensive performance of the target node and associated links in a short period after the migration is completed, so as to screen the optimal migration target and timing.

[0014] Preferably, the system further includes a digital twin simulation and iteration module, which is used to construct a high-fidelity digital twin, perform simulation verification and risk simulation of the collaborative control strategy, and continuously iterate and optimize the multi-agent reinforcement learning model based on feedback data. The digital twin simulation and iteration module is connected to the intelligent analysis and decision-making module and the collaborative control and execution module, and is used to perform simulation verification and risk simulation before strategy execution. The high-fidelity digital twin constructed by the digital twin simulation and iteration module supports the injection of historical or constructed extreme failure modes to test the resilience of the collaborative control strategy and the effectiveness of the emergency control link.

[0015] Preferably, the multi-dimensional perception and fusion module further includes a cross-architecture state perception engine for heterogeneous computing power, which monitors the throughput and latency of communication links between different chip architecture instances in a fine-grained manner, and dynamically identifies and avoids systemic bottleneck links caused by specific hardware combinations.

[0016] This invention also provides a method for a resource regulation system of an intelligent computing center that integrates multi-dimensional sensing and physical information, characterized in that it is applied to the regulation system described above, and the method includes:

[0017] A unified view of resource task status is constructed based on the resource knowledge graph of the intelligent computing center, which is based on pre-built physical information.

[0018] Load prediction is performed based on the resource task state view, and a multi-agent reinforcement learning model is driven to generate a collaborative regulation strategy. The collaborative mechanism of the multi-agent reinforcement learning model is based on a cross-domain collaborative task specification formally defined by generalized reactive temporal logic, which constructs a set of synchronous, decomposed reward automata by decomposing the synthesized high-level team strategy.

[0019] The coordinated control strategy is executed to output dynamic allocation instructions for computing, network, and storage resources, as well as coordinated adjustment instructions for energy infrastructure.

[0020] An end-to-end predictive model based on a physical information neural network is run. This model takes a resource scheduling plan as input, integrates physical environment simulation data, and generates refined energy efficiency control commands.

[0021] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the intelligent computing center resource control system method as described above, which integrates multi-dimensional sensing and physical information.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] Solved the reliability and security challenges of cross-domain collaboration: by introducing a system based on... Formal logic's strategy synthesis and decomposition mechanism transforms complex temporal attributes such as security and activity into an intrinsic reward structure that can be embedded in the distributed learning process, fundamentally ensuring the logical correctness and system resilience of collaborative strategies under complex temporal constraints.

[0024] Achieved physical-level fine control for energy efficiency optimization: through deep fusion of physical information neural networks The scheduling instructions and the physical model of the data center environment enable end-to-end accurate prediction and linkage from logical resource layout to physical power consumption and thermal field distribution, which conforms to the laws of physics. This breaks through the bottleneck of energy efficiency optimization accuracy based on traditional empirical formulas or statistical models.

[0025] A fully closed-loop autonomous system was constructed, encompassing perception, decision-making, execution, and simulation iteration. Through fault injection and forward-looking simulation in a high-fidelity digital twin, the strategy underwent "sandbox testing" and self-evolution, significantly improving the system's adaptability to unknown loads and sudden failures, as well as its long-term operational stability. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of the intelligent computing center resource control system that integrates multi-dimensional perception and physical information according to the present invention.

[0027] Figure 2 This is a flowchart of a multi-agent collaborative decision-making mechanism based on formal specifications in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the energy efficiency linkage control principle of integrating physical information neural networks in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of the closed loop of digital twin simulation verification and model iteration in an embodiment of the present invention. Detailed Implementation

[0030] 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.

[0031] like Figure 1-4 As shown, this embodiment of the invention provides a resource control system for an intelligent computing center that integrates multi-dimensional sensing and physical information, comprising:

[0032] The multi-dimensional perception and fusion module constructs a unified resource task status view based on the resource knowledge graph of the intelligent computing center, which is based on pre-built physical information.

[0033] The intelligent analysis and decision-making module is used to predict the load based on the resource task status view and drive the multi-agent reinforcement learning model to generate a collaborative control strategy. The collaborative mechanism of the multi-agent reinforcement learning model is based on the cross-domain collaborative task specification formally defined by generalized reactive temporal logic. It guides multiple decision agents to perform distributed strategy collaboration by decomposing the synthesized high-level team strategy into a set of synchronous, decomposed reward automata.

[0034] The coordinated control and execution module is used to execute the coordinated control strategy and output dynamic allocation instructions for computing, network, and storage resources, as well as linkage adjustment instructions for energy infrastructure.

[0035] The physical information fusion module is used to run an end-to-end prediction model based on a physical information neural network. This model takes a resource scheduling plan as input, fuses physical environment simulation data, and generates refined energy efficiency control instructions. These energy efficiency control instructions are used in conjunction with the adjustment instructions output by the collaborative regulation and execution module to achieve cross-layer dynamic resource optimization.

[0036] In this embodiment, the core of the intelligent computing center resource control system, which integrates multi-dimensional perception and physical information, lies in constructing a closed-loop intelligent management system from perception to execution. The system first aggregates multi-dimensional raw data from computing chips, storage devices, network links, and power distribution and cooling units in real time through widely deployed sensors and software probes, and simultaneously acquires metadata such as task type and priority. This data is sent to a fusion center based on a resource knowledge graph. This center automatically establishes and maintains dynamic semantic relationships between tasks, virtual resources, physical devices, and energy consumption indicators, thereby generating a globally unified, reasonable view of resource and task status, providing the system with unprecedented overall situational understanding capabilities. Building upon this foundation, a forward-looking intelligent decision-making engine begins operation. It utilizes time-series models to predict future load changes and drives a multi-agent reinforcement learning model to generate scheduling strategies. The unique feature of this learning model lies in its collaborative mechanism: it formalizes high-level management objectives (such as "ensuring zero disruption to critical business operations") using generalized reactive temporal logic and automatically synthesizes these abstract specifications into specific high-level team strategies. These strategies are then decomposed and constructed into a set of synchronously running reward automata to precisely guide the learning and decision-making processes of multiple distributed agents, including computing, networking, storage, and energy efficiency, thereby fundamentally ensuring the logical correctness and temporal security of the collaborative strategy. After the decision is generated, the collaborative execution module translates the strategy into control instructions for each infrastructure layer, enabling dynamic resource allocation. Crucially, an independent physical information fusion module is simultaneously activated. Utilizing a neural network model embedded with physical laws, it accurately predicts the future power consumption and thermal load that the scheduling strategy will trigger, generating corresponding refined energy efficiency control instructions, achieving a deep integration of information space decision-making and physical world response. Finally, before being deployed in production, all strategies undergo simulation verification and extreme failure stress testing in a high-fidelity digital twin environment. The real-world performance data after strategy execution is then fed back to continuously iterate and optimize the model. This complete closed loop enables the system not only to perform global optimization but also to possess inherent security guarantees, physical-level precise control, and continuous self-evolution capabilities, fundamentally improving the autonomy, reliability, and energy efficiency of the intelligent computing center.

[0037] In some implementations, the construction of the resource knowledge graph specifically includes: using cross-language text mining and semantic association rules to automatically discover and expand three types of entities—computing tasks, physical resources, and energy efficiency indicators—and the relationships between them, and establishing a dedicated resource knowledge graph for the intelligent computing center scheduling domain through iterative learning.

[0038] In this embodiment, the process of constructing a resource knowledge graph in the system is a continuous self-evolving process. Specifically, the system does not rely on statically preset graph templates, but instead uses natural language processing technology to automatically scan and analyze cross-language technical documents, patent documents, operation and maintenance reports, and logs. This process can identify and extract novel hardware entities, software task paradigms, performance indicators, and potential relationships between them. For example, it can discover from text that "a certain type of graph computing task in..." The system incorporates implicit knowledge such as "sensitive to inter-chip interconnect bandwidth" when running on accelerators. These newly discovered entities and relationships are dynamically integrated into the existing resource knowledge graph network after confidence evaluation. Through this iterative learning, the domain knowledge graph constructed by the system can continuously adapt to the rapidly changing intelligent computing technology ecosystem. This allows the resource task state view to not only reflect real-time data but also contain rich domain prior knowledge and deep semantic relationships, thereby providing more accurate and insightful contextual information for intelligent decision-making and significantly improving the system's ability to understand and schedule complex tasks and new resources.

[0039] In some implementations, the time series model used for load forecasting in the intelligent analysis and decision-making module is based on... The architecture of the time-series information multi-agent model models multi-dimensional time-series data into sequences of arbitrary length to extract long-range dependencies, and serves as a general pre-training basis for the multi-agent reinforcement learning model.

[0040] In this embodiment, the time series prediction unit in the intelligent decision engine adopts a method based on... The architecture employs a time-series information-based multi-agent model. When processing multi-dimensional historical data sequences from resource task state views, the encoder uses a self-attention mechanism to compute the interdependence weights of different resource indicators at different time points in parallel. This effectively captures long-range causal relationships across domains and time periods, such as "sudden increases in network latency often lead to queuing of subsequent computational tasks." After training on massive amounts of historical data, the model can not only accurately predict load and performance indicators at multiple future time points, but more importantly, the deep feature representations of the temporal dynamics of complex systems learned internally are extracted as a shared foundation for all subsequent decision-making agents to initialize or fine-tune their respective policy networks. This is equivalent to providing all distributed agents with a unified, high-quality "temporal cognitive prior," significantly reducing the exploration difficulty and instability of multi-agent reinforcement learning in complex environments, accelerating the convergence of cooperative strategies, and enhancing the system's generalization and rapid adaptation capabilities when facing unprecedented load patterns.

[0041] In some implementations, the multi-agent reinforcement learning model includes four decision agents: computing resources, network resources, storage resources, and energy efficiency management resources; and a joint reward function for evaluating the joint policy, the weight coefficients of which are dynamically adjusted by a meta-reinforcement learning submodule.

[0042] In this embodiment, the multi-agent reinforcement learning model comprises multiple decision-making agents, each with its own specific role, all evaluated by a joint reward function. This joint reward function integrates multiple objectives, including task completion time, resource utilization, energy consumption, and service level agreement violation rate. The relative importance weights of these objectives are not fixed but dynamically managed by a meta-reinforcement learning submodule. This meta-learning module observes actual scheduling performance data over a historical period and analyzes which weight configurations guide the system to achieve better long-term overall benefits under different external conditions (such as peak business periods, off-peak nighttime periods, and different electricity price periods). Then, it automatically adjusts the weight coefficients in the joint reward function for the next stage through an online learning algorithm. This allows the system's optimization objectives to intelligently switch and dynamically balance according to the actual operating rhythm and external policy requirements, much like an experienced administrator. For example, while ensuring core business performance, it can more actively optimize energy efficiency during off-peak electricity price periods, thereby achieving adaptive and refined adjustment of the global optimization strategy.

[0043] In some implementations, the decision to migrate computing instances in real time is executed by an intelligent migration decision engine. This engine, based on the resource task status view and by calling the digital twin simulation and iteration module, performs forward simulations of candidate migration schemes in the digital twin environment to predict the results and select the optimal migration target and timing.

[0044] In this embodiment, the decision-making process of the system during real-time migration of computing instances is highly optimized by an intelligent migration engine. When a migration request is triggered, the engine does not simply select the node with the lightest current load, but generates several candidate migration schemes and immediately requests support from the digital twin simulation environment. In the twin environment, each candidate scheme undergoes an accelerated, end-to-end simulation: not only simulating the transfer process of memory and storage data, but also loading the predicted short-term business load onto the virtual resource layout after the migration for "pre-running." By analyzing the performance curves of the simulation output, the engine can proactively assess the service response latency jitter, the impact on related businesses, and the changes in overall system energy efficiency that each scheme may cause. Ultimately, the engine selects the scheme with the best overall performance and the most stable performance in the simulation for implementation. This decision-making mechanism based on digital twin forward-looking simulation transforms migration from a passive response based on instantaneous state to a proactive planning based on comprehensive simulation, maximizing business continuity and achieving smooth and accurate load rebalancing.

[0045] In some implementations, the prediction model based on the physical information neural network specifically incorporates the future resource distribution map derived from the resource scheduling strategy, computational fluid dynamics simulation data of the airflow organization in the data center, and the operating modal parameters of the cooling system as input.

[0046] In this embodiment, the core of the prediction model in the physical information fusion module is a physical information neural network. This model receives three types of inputs: first, a future load distribution map of computing and storage nodes parsed from the scheduling strategy; second, three-dimensional temperature and velocity field grid data of the data center obtained through computational fluid dynamics simulation; and third, specific operating parameters of the cooling equipment, such as chilled water temperature, flow rate, and fan speed. During network training, constraint terms consisting of discretized physical laws such as energy conservation and momentum conservation are explicitly added to the traditional mean square error loss function. This design forces the model to ensure that its output basically conforms to physical laws while learning data features, thereby significantly improving the reliability and physical rationality of the model in operating conditions not covered by training data or when making long-term extrapolation predictions. Therefore, this model can output extremely high-precision, rack-level predictions of future power consumption and hotspot distribution, providing a reliable physical basis for subsequent precise cooling control and achieving a high degree of synergy from bitstream to thermal management.

[0047] In some implementations, a digital twin simulation and iteration module is also included, used to construct a high-fidelity digital twin, to perform simulation verification and risk extrapolation on the collaborative control strategy, and to continuously iterate and optimize the multi-agent reinforcement learning model based on feedback data; wherein, the high-fidelity digital twin supports the injection of extreme failure modes, used to test the emergency response capability and system resilience of the collaborative control strategy.

[0048] In this embodiment, the digital twin simulation environment integrates a powerful automated fault injection and testing framework. This framework allows for the definition and execution of various single or combined extreme fault scenarios, such as simulating batch failures of specific hard drives, intermittent core network link outages, or progressive performance degradation of cooling water pumps. Any control strategies to be deployed must be re-verified and tested in the twin environment after such fault injection. By observing and recording the strategy's behavioral response, fault isolation effectiveness, and system recovery time under simulated faults, the robustness of the strategy and the effectiveness of emergency links can be quantitatively evaluated. This proactive "chaos engineering" verification method can expose and repair potential defects of the strategy under extreme conditions in virtual space in advance, systematically and preventively honing and improving the resilience and high availability of the entire intelligent computing center service.

[0049] In some implementations, the multi-dimensional perception and fusion module further includes a cross-architecture state perception engine for fine-grained monitoring of the performance metrics of communication links between computing instances of different chip architectures, and dynamically identifying systemic bottleneck links caused by hardware heterogeneity.

[0050] In this embodiment, the system's perception layer specifically includes a cross-architecture state-aware engine to address the increasingly prevalent heterogeneous computing environments. This engine uses lightweight probes deployed at the host layer to monitor and analyze computing units of different architectures (such as...) with fine granularity. , , The engine analyzes the performance characteristics of data exchange and synchronization between devices, including point-to-point communication bandwidth and cross-device memory access latency. Its built-in analytical model correlates these performance metrics with hardware models, driver versions, and interconnect topologies, dynamically diagnosing systemic communication bottlenecks caused by specific hardware combinations or software configurations. This diagnosed "performance affinity" and "potential degradation" information is transformed into a special resource tag and integrated into the global state view. This allows upper-layer scheduling decisions to fully consider the compatibility between task communication patterns and the underlying hardware topology when allocating resources to tasks, proactively avoiding deploying highly communication-intensive tasks on node combinations with hidden bottlenecks. This achieves deeper, hardware topology-aware performance optimization within the heterogeneous computing pool.

[0051] This invention also provides a resource regulation method for intelligent computing centers that integrates multi-dimensional sensing and physical information, applied to the regulation system described in any of the preceding claims, the method comprising:

[0052] A unified view of resource task status is constructed based on the resource knowledge graph of the intelligent computing center, which is based on pre-built physical information.

[0053] Load prediction is performed based on the resource task state view, and a multi-agent reinforcement learning model is driven to generate a collaborative regulation strategy. The collaborative mechanism of the multi-agent reinforcement learning model is based on a cross-domain collaborative task specification formally defined by generalized reactive temporal logic, which constructs a set of synchronous, decomposed reward automata by decomposing the synthesized high-level team strategy.

[0054] The coordinated control strategy is executed to output dynamic allocation instructions for computing, network, and storage resources, as well as coordinated adjustment instructions for energy infrastructure.

[0055] An end-to-end predictive model based on a physical information neural network is run. This model takes a resource scheduling plan as input, integrates physical environment simulation data, and generates refined energy efficiency control commands.

[0056] In this embodiment, the control method based on the aforementioned system embodies a complete intelligent closed loop from perception and cognition to physical execution. The method begins with real-time acquisition of operational data across the entire intelligent computing center and deep fusion based on resource knowledge graphs, thereby establishing a unified and understandable view of the environment's state. Next, the method expands the temporal dimension of this view using a temporal prediction model and drives a multi-agent reinforcement learning model guided by formal specifications to generate a safe and reliable collaborative scheduling strategy. Subsequently, the method executes two key paths in parallel: first, decomposing the strategy into specific control actions on computing, network, and storage resources; second, inputting the strategy into a physical information neural network to generate precisely matched, refined energy efficiency control instructions. The outputs of both paths take effect synchronously in the physical world, achieving cross-domain resource linkage. Finally, the method emphasizes that all strategies must undergo simulation verification and fault testing in a digital twin environment, and uses actual execution feedback data for continuous iterative learning of the model. This method, by organically integrating formal verification, physical modeling, and data-driven learning, systematically solves the challenges of security, accuracy, and adaptability in large-scale resource coordination.

[0057] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0058] In this embodiment, when the program is executed by a computing device, all steps defined by the aforementioned control method will be fully implemented. This means that the technical solution of the present invention can be solidified, replicated, and widely deployed in the form of a software product. The program drives the processor to complete the entire process from multi-dimensional data perception, resource knowledge graph construction, formally guided intelligent decision-making, energy efficiency prediction through physical information fusion, to digital twin verification and iterative optimization, thereby instantiating a highly autonomous, secure, reliable, and energy-efficient resource management system on various intelligent computing center infrastructures, enabling the standardized, productized delivery and application of the technical benefits of the present invention.

[0059] The following are specific embodiments of the present invention:

[0060] See Figure 1 The system in this embodiment includes five core modules: multi-dimensional perception and fusion module, intelligent analysis and decision-making module, collaborative control and execution module, physical information fusion module, and digital twin simulation and iteration module. Each module communicates loosely with the message middleware through a high-speed data bus.

[0061] The specific implementation of the multi-dimensional perception and fusion module includes: deploying lightweight acquisition agents on physical servers, network switches, storage arrays, and intelligent power distribution units, while extracting task metadata from the virtualization management platform and container orchestration engine. The scope of the collected data covers... The module displays utilization and performance counters, storage read / write throughput and latency, network link bandwidth utilization and packet loss rate, task types and dependencies, and real-time power consumption at the server, rack, and cooling system levels. Its built-in resource knowledge graph engine utilizes natural language processing to automatically extract entity relationships from operational documents and patent texts, dynamically constructing and updating a domain knowledge graph depicting the multi-dimensional relationships between "tasks, resources, and energy efficiency," forming a unified view of resource and task status.

[0062] The intelligent analysis and decision-making module is the core of this invention. See also... Figure 2 This module first utilizes a based The time-series prediction submodule performs multi-step predictions on key metrics in the resource task state view. The prediction results, along with the current state, are then fed into an innovative multi-agent reinforcement learning decision core.

[0063] The core decision-making system employs dedicated decision-making agents for four resource domains: computing, networking, storage, and energy efficiency. This is achieved by introducing a formal reduction processing unit. This unit receives data in the form of generalized reactive sequential logic. The task specification written, such as " (Rack power exceeded) (Migrate at least 50% of the load within 3 control cycles), where " " indicates "always", "->" indicates "implies", The term "finally" indicates the automatic synthesis of a high-level deterministic policy that satisfies the specification through model detection technology. Subsequently, a policy decomposition and reward automaton construction unit decomposes this high-level policy into behavioral constraints for each of the four agents and constructs a set of synchronous, decomposed reward automata. Each reward automaton provides an immediate reward signal to its corresponding agent, which not only encourages the achievement of the global goal but also embeds the temporal logic requirements that satisfy the formal specification. Under the decentralized partially observable Markov decision process framework, the agents perform distributed training and decision-making based on this reward structure, ultimately outputting a coordinated and consistent regulatory policy. The weights of the joint reward function are dynamically optimized by a meta-learning submodule according to the business stage.

[0064] The coordinated control and execution module is responsible for the precise execution of the strategy. On the computing resource side, the intelligent migration engine combines real-time perception and digital twin forward-looking assessment results to select the optimal migration target and timing. On the network side, flow tables are issued through the software-defined network controller. On the storage side, data replica adjustments are triggered.

[0065] The physical information fusion module is key to achieving precise energy efficiency control. (See also...) Figure 3In this embodiment, the module runs a physical information neural network model. This model takes as input a future resource distribution map translated from the scheduling strategy, computational fluid dynamics simulation grid data for the data center, and outdoor temperature and humidity predictions. By embedding physical conservation equations (such as energy conservation equations) as soft constraints into its loss function, it outputs refined rack-level power consumption and hotspot predictions for the next few minutes. These predictions are directly fed back to the energy efficiency management agent and transformed into precise control commands for the uninterruptible power supply's operating mode, server power consumption status, and the data center's air conditioning supply temperature and chilled water flow rate.

[0066] The digital twin simulation and iteration module constructs a digital twin that runs parallel to the production environment, maintaining state consistency through virtual-real synchronization technology. See also... Figure 4 Any strategy generated by the intelligent analysis and decision-making module must be simulated and extrapolated within this twin before being delivered to the production system for execution. The simulation not only evaluates performance gains but also proactively introduces extreme failure scenarios such as "optical link interruption" and "sudden increase in single-cabinet power consumption" to test the strategy's resilience. The real-world performance data after strategy execution, along with the simulation data, is stored as experience samples for periodic offline batch retraining of the multi-agent reinforcement learning model, forming a complete evolutionary closed loop of "decision-simulation-execution-evaluation-learning."

[0067] In summary, this invention constructs a highly intelligent, reliable, and efficient autonomous resource management system for intelligent computing centers through the deep integration of "collaborative decision-making guaranteed by formal logic," "precise energy efficiency regulation through physical information fusion," and "self-evolution driven by high-fidelity digital twins."

[0068] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions made to the technical solutions of the present invention cannot depart from the spirit and scope of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. A resource control system for an intelligent computing center that integrates multi-dimensional sensing and physical information, characterized in that, include: The multi-dimensional perception and fusion module constructs a unified resource task status view based on the resource knowledge graph of the intelligent computing center, which is based on pre-built physical information. The intelligent analysis and decision-making module is used to predict the load based on the resource task status view and drive the multi-agent reinforcement learning model to generate a collaborative regulation strategy; wherein, the collaborative mechanism of the multi-agent reinforcement learning model is based on the cross-domain collaborative task specification formally defined by generalized reactive temporal logic, and constructs a set of synchronous, decomposed reward automata by decomposing the synthesized high-level team strategy. The coordinated control and execution module is used to execute the coordinated control strategy and output dynamic allocation instructions for computing, network, and storage resources, as well as linkage adjustment instructions for energy infrastructure. The physical information fusion module is used to run an end-to-end prediction model based on a physical information neural network. This model takes a resource scheduling plan as input, fuses physical environment simulation data, and generates refined energy efficiency control instructions. These energy efficiency control instructions are used in conjunction with the adjustment instructions output by the collaborative regulation and execution module to achieve cross-layer dynamic resource optimization.

2. The system according to claim 1, characterized in that, The construction of the resource knowledge graph specifically includes: using cross-language text mining and semantic association rules to automatically discover and expand three types of entities—computing tasks, physical resources, and energy efficiency indicators—and the relationships between them; and establishing a dedicated resource knowledge graph for the intelligent computing center scheduling domain through iterative learning.

3. The system according to claim 1, characterized in that, In the intelligent analysis and decision-making module, the time series model used for load forecasting is based on The architecture of the time-series information multi-agent model models multi-dimensional time-series data into sequences of arbitrary length to extract long-range dependencies, and serves as a general pre-training basis for the multi-agent reinforcement learning model.

4. The system according to claim 1, characterized in that, The multi-agent reinforcement learning model comprises four decision agents: computing resources, network resources, storage resources, and energy efficiency management resources; and a joint reward function for evaluating the joint policy, the weight coefficients of which are dynamically adjusted by a meta-reinforcement learning submodule.

5. The system according to claim 1, characterized in that, In the collaborative control and execution module, the decision to migrate computing instances in real time is executed by an intelligent migration decision engine. This engine is based on the resource task status view and calls the digital twin simulation and iteration module to perform forward simulation of candidate migration schemes in the digital twin environment, so as to predict the results and select the optimal migration target and timing.

6. The system according to claim 1, characterized in that, The prediction model based on physical information neural networks incorporates the future resource distribution map derived from resource scheduling strategies, computational fluid dynamics simulation data of airflow organization in data center computer rooms, and operating modal parameters of the cooling system as inputs.

7. The system according to claim 1, characterized in that, It also includes a digital twin simulation and iteration module, which is used to construct a high-fidelity digital twin, perform simulation verification and risk simulation of the collaborative control strategy, and continuously iterate and optimize the multi-agent reinforcement learning model based on feedback data. The digital twin simulation and iteration module is connected to the intelligent analysis and decision-making module and the collaborative control and execution module, and is used to perform simulation verification and risk simulation before the strategy is executed. The high-fidelity digital twin supports the injection of extreme failure modes to test the emergency response capability and system resilience of the collaborative control strategy.

8. The system according to claim 1, characterized in that, The multi-dimensional perception and fusion module also includes a cross-architecture state perception engine, which is used to monitor the performance indicators of communication links between computing instances of different chip architectures in a fine-grained manner and dynamically identify systemic bottleneck links caused by hardware heterogeneity.

9. A resource regulation method for an intelligent computing center that integrates multi-dimensional sensing and physical information, applied to the regulation system as described in any one of claims 1 to 8, characterized in that, The method includes: A unified view of resource task status is constructed based on the resource knowledge graph of the intelligent computing center, which is based on pre-built physical information. Load prediction is performed based on the resource task state view, and a multi-agent reinforcement learning model is driven to generate a collaborative regulation strategy. The collaborative mechanism of the multi-agent reinforcement learning model is based on a cross-domain collaborative task specification formally defined by generalized reactive temporal logic, which constructs a set of synchronous, decomposed reward automata by decomposing the synthesized high-level team strategy. The coordinated control strategy is executed to output dynamic allocation instructions for computing, network, and storage resources, as well as coordinated adjustment instructions for energy infrastructure. An end-to-end predictive model based on a physical information neural network is run. This model takes a resource scheduling plan as input, integrates physical environment simulation data, and generates refined energy efficiency control commands.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 9.

Images