A method and device for safe control of multimodal thermal load in data centers

Abstract

This invention relates to the field of data center operation and maintenance and intelligent control technology, specifically providing a method and apparatus for safe control of multimodal thermal load in data centers, comprising the following steps: S1, data acquisition and preprocessing; S2, multimodal semantic representation generation; S3, construction of a trusted action cone; S4, security policy generation and execution; S5, control feedback and policy iteration. Compared with existing technologies, this invention can solve the problems of difficulty in fusing multimodal data, lack of security boundaries in black-box control, and difficulty in cross-target collaborative optimization in existing data center thermal management solutions.

Description

Technical Field

[0001] This invention relates to the field of data center operation and maintenance and intelligent control technology, specifically providing a method and device for safe control of multimodal thermal load in data centers. Background Technology

[0002] As large data centers continue to expand in scale and computing cluster density rises, the energy consumption structure and thermal management complexity of data centers are increasing dramatically. Existing data centers typically rely on the coordinated operation of devices such as temperature probes, PDU power meters, chiller controllers, and damper actuators, with basic operations and maintenance systems (such as BMS and EMS) periodically collecting data and performing cooling and load scheduling based on threshold-triggered rules or empirical models. However, this approach suffers from the following significant problems: First, there is a lack of fusion and modeling capabilities among multimodal data. Data formats such as temperature points, wind speed images, and equipment operation logs are heterogeneous, and information is scattered across different systems. This makes it impossible for control logic to uniformly process spatiotemporal context relationships, forming "perception islands" and resulting in a slow response to sudden hot spots or joint anomalies of multiple nodes.

[0003] Secondly, the control strategy lacks safety and transparency. Currently widely used control models, such as traditional PID control or black-box deep reinforcement learning methods, are difficult to explicitly constrain the operating boundaries. In extreme cases, they may output operating commands that exceed the equipment's carrying capacity limits, leading to safety hazards such as cold-cooling overload and air pressure imbalance, which are particularly prominent when the load fluctuates drastically or there is local thermal runaway.

[0004] Third, control objectives are often focused on a single metric. For example, minimizing PUE (Power Usage Effectiveness) while ignoring Service Level Agreement (SLA) requirements, or focusing solely on reducing hotspot temperature at the expense of system stability, makes it difficult to achieve a dynamic balance between energy efficiency, safety, and service quality.

[0005] To address the aforementioned issues, there is an urgent need for a novel control method that integrates multimodal data perception, possesses security constraint mechanisms, and supports multi-objective dynamic optimization. In particular, a large-scale model with semantic understanding capabilities should be introduced as a "scheduling hub" to break down data barriers and achieve intelligent and refined upgrades to data center heat-load collaborative control.

[0006] Therefore, how to solve the problems of difficulty in integrating multimodal data, lack of security boundaries in black-box control, and difficulty in cross-target collaborative optimization in existing data center thermal management solutions is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] This invention addresses the shortcomings of the prior art by providing a highly practical method for safe control of multimodal thermal load in data centers.

[0008] A further technical objective of this invention is to provide a reasonably designed, safe, and applicable data center multimodal thermal load safety control device.

[0009] The technical solution adopted by this invention to solve its technical problem is: A method for safe control of multimodal thermal load in data centers includes the following steps: S1. Data Acquisition and Preprocessing; S2, Multimodal semantic representation generation; S3. Construct a trustworthy action cone; S4. Security policy generation and execution; S5, Control Feedback and Strategy Iteration.

[0010] Furthermore, in step S1, the multimodal data access module first continuously collects data from temperature sensors, wind speed field cameras, PDU power metering devices, cooling equipment status, and operation and maintenance log systems. The collection frequency is dynamically allocated according to the type of equipment, and the log data is asynchronously written through a trigger mechanism. After collection, the data undergoes a preprocessing process to standardize and transform it. Time-series data is filled with missing values ​​using sliding window resampling and linear interpolation. Image data is enhanced with edge enhancement and thermal field saliency region extraction. Text logs are constructed with semantic prompts through regular expression filtering and key term highlighting. All data is tagged with a unified timestamp and source label.

[0011] Furthermore, in step S2, the preprocessed multimodal data enters the unified semantic representation module, and is converted into a unified high-dimensional embedding vector through the large language model and the multimodal fusion structure. The text log is input into the large language model to obtain a summary vector containing contextual risk semantics. The wind field image is input into the visual Transformer to extract local vortex and abnormal flow direction spatial structure features. Meanwhile, temperature and power consumption time-series data are used to construct sliding window vectors before input, and a time coding mechanism is used to enhance the perception of trends and abrupt changes. Multimodal embedding is finally fused into a unified state vector and sent to the next action constraint generation module.

[0012] Furthermore, in step S3, using the current unified semantic representation, combined with the device control limit, industry safety standards and service level agreements, semantic reasoning is performed through a large language model to generate a "trusted action cone". The large model will assess whether the current environment is in an overload, warning or idle range, and generate a set of permissible actions accordingly.

[0013] Furthermore, the action set is output in a structured representation and the corresponding boundary meanings are indicated. When the downstream controller generates an action, it will first verify the action cone constraint. Actions that exceed the boundary will be rejected or trigger a policy update mechanism.

[0014] Furthermore, in step S4, the safety reinforcement learning controller receives the current state vector and performs policy search in the action space defined by the trusted action cone. The control strategy aims to reduce cooling energy consumption, keep the equipment temperature within a safe range, and maintain the SLA satisfaction rate. The controller dynamically calculates the optimal control path and optimizes the action selection based on long-term cumulative rewards to form a control capability to cope with different loads and hot spot situations. Control commands are sent in real time to the cooling equipment control system, the fan speed regulation module, and the load scheduling interface.

[0015] Furthermore, in step S5, after the control execution is completed, the control response index and multimodal data of the next state will be collected, and the state transition trajectory will be constructed and fed back to the controller. If the system experiences hotspots not being eliminated, equipment startup failure, or SLA deviation after execution, the diagnostic mechanism will be triggered and the current strategy module will be frozen for analysis. During continuous control cycles, the system maintains sliding window indicators for periodic strategy fine-tuning.

[0016] A data center multimodal thermal load safety control device includes: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute a data center multimodal thermal load safety control method.

[0017] Compared with the prior art, the data center multimodal thermal load safety control method and device of the present invention have the following outstanding advantages: This invention establishes a semantically driven, action-constrained, and secure intelligent scheduling mechanism by deeply integrating a large language model with data center thermal load regulation. This effectively overcomes the problems of response lag, local overheating, and policy rigidity in traditional rule-based or local perception-based cooling control methods.

[0018] Compared to existing solutions that rely solely on physical modeling or shallow machine learning, this invention, based on unified multimodal embedding and a "trusted action cone" constraint mechanism, ensures the robustness and boundary safety of the control strategy in complex scenarios. Simultaneously, it introduces the contextual understanding capabilities of a large language model, endowing the control system with a deep understanding and rapid response capability to semantic anomalies (such as alarm logs, maintenance instructions, and operational text). By constructing a joint control closed loop of reinforcement learning strategies and a large model, the system can achieve comprehensive optimization of multiple objectives, including hot and cold throughput, load paths, and SLA priorities. It exhibits significant performance improvements in energy consumption control, resource utilization, and risk warning, making it suitable for large-scale data center environments with high-density deployment and high reliability requirements. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for safe control of multimodal thermal load in data centers. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely 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.

[0022] The following is a preferred embodiment: like Figure 1 As shown in this embodiment, a data center multimodal thermal load safety control method includes the following steps: S1. Data Acquisition and Preprocessing; The system first uses a multimodal data access module to continuously collect various data from temperature sensors, wind speed field cameras, PDU power metering devices, cooling equipment status, and maintenance log systems. The collection frequency is dynamically allocated according to the type of equipment, with key hot zone temperature probes and power consumption data synchronized at the second level, while log data is written asynchronously through a trigger mechanism.

[0023] After collection, the data underwent a preprocessing procedure for standardization. Time-series data were filled with missing values ​​using sliding window resampling and linear interpolation; image data underwent edge enhancement and extraction of salient thermal regions; and text logs were processed using regular expression filtering and keyword highlighting to construct semantic cues. All data were tagged with a unified timestamp and source label to facilitate subsequent modality alignment.

[0024] S2, Multimodal semantic representation generation; The preprocessed multimodal data enters the unified semantic representation module, where it is transformed into a unified high-dimensional embedding vector through a large language model and a multimodal fusion structure. Text logs are input into the large language model to obtain summary vectors containing contextual risk semantics; wind field images are input into a visual Transformer to extract spatial structural features such as local vortices and anomalous flow directions.

[0025] Simultaneously, time-series data such as temperature and power consumption are used to construct sliding window vectors before input, employing a time-coding mechanism to enhance the perception of trends and abrupt changes. Multimodal embeddings are ultimately fused into a unified state vector, which is then fed into the next action constraint generation module. This step achieves a comprehensive understanding of multi-source information, laying a semantic foundation for safety control.

[0026] S3. Construct a trustworthy action cone; The system utilizes the current unified semantic representation, combined with equipment control limits, industry safety standards, and service level agreements, to perform semantic reasoning through a large language model and generate "trusted action cones." The large model assesses whether the current environment is in an overload, warning, or idle range, and generates a set of permissible actions accordingly, such as the maximum adjustable speed range of the fan, whether the chiller can be delayed in starting and stopping, and the migration range of the total PDU load.

[0027] These control actions are output in a structured format, with corresponding boundary meanings indicated, such as "current airflow increase not exceeding 20%" and "migration requests must avoid SLA level A clusters." When a downstream controller generates an action, the system will first perform action cone constraint verification on it; actions exceeding the boundary will be rejected or trigger a policy update mechanism.

[0028] S4. Security policy generation and execution; The safety reinforcement learning controller receives the current state vector and searches for a policy within the action space defined by the trusted action cone. The control policy focuses on reducing cooling energy consumption, maintaining equipment temperature within a safe range, and ensuring SLA (Safety Level Assessment) satisfaction. The controller dynamically calculates the optimal control path. Based on long-term cumulative rewards, the controller optimizes action selection, forming a flexible control capability to respond to different loads and hotspot conditions.

[0029] Control commands are sent in real time to the cooling equipment control system, fan speed regulation module, and load scheduling interface. During strategy execution, the system monitors key indicators such as local temperature rise rate, air pressure fluctuation, and delay changes after load migration in real time to ensure that the control results meet expectations.

[0030] S5, Control Feedback and Strategy Iteration; After control execution is complete, the system will collect control response indicators and multimodal data of the next state, construct a state transition trajectory, and feed it back to the controller. If the system encounters anomalies such as unresolved hotspots, equipment startup failures, or SLA deviations after execution, the system will trigger a rapid diagnostic mechanism and freeze the current strategy module for analysis.

[0031] During continuous control cycles, the system maintains sliding window metrics (such as PUE rolling average and temperature rise response delay) for periodic policy fine-tuning. Furthermore, the trusted action cone is incrementally adjusted by the large language model as the context changes, enabling continuous evolution of knowledge updates and control security, thus constructing a truly closed-loop intelligent safety control system.

[0032] The multimodal data access module is responsible for collecting multi-source operational data from the data center infrastructure, including temperature and humidity probe sampling, wind speed and direction field images, PDU power consumption curves, chiller operating status, as well as log alarm texts and scheduling history records. The module converts heterogeneous data into structured time-series samples through edge device access and standardized protocol conversion.

[0033] Furthermore, to support high-frequency control responses, this module features real-time data caching and missing value imputation mechanisms, ensuring the completeness and timeliness of input acquired by the control unit. Log data is collected by incorporating key timestamps, host mapping relationships, and natural language preprocessing to provide a foundation for subsequent semantic analysis.

[0034] The unified semantic representation module uses a pre-trained large language model as the core encoder to map multimodal input data (such as images, time series data, and text) to a unified embedding vector space. A sliding time window mechanism is used for inputting temperature and power consumption time series data; a visual Transformer is used for local-global attention modeling of wind field images; and context summarization and risk factor extraction are performed on operation and maintenance log text.

[0035] Through joint training or distillation, all modalities are ultimately represented in a unified vector space, preserving key semantic information (such as heat source trends, abnormal alarm words, and air conditioning cooling efficiency degradation patterns). This unified representation provides high-dimensional semantic support for subsequent action generation and policy evaluation, while also enhancing the model's ability to capture cross-modal causal relationships.

[0036] The generation module uses a large language model to analyze the current system state and, in conjunction with equipment operating limits, industry standards, scheduling policies, and SLA clauses, constructs a set of "acceptable control actions," called the "credible action cone." This cone-shaped space semantically corresponds to a "soft constraint domain," which actually limits the range of control variables such as fan speed range, chiller start-up and shutdown conditions, and load shift ratio.

[0037] The large model comprehensively evaluates whether the current scenario allows for high-frequency fan acceleration and whether the chiller start-up and shutdown can be delayed through a chain-of-thought inference chain, and outputs the action cone boundary in the form of text-to-structured control rules. Once the controller attempts to exceed the boundary, this module will forcibly cut off the policy output and feed this behavior back into the control policy update process.

[0038] The safety reinforcement learning controller is a dynamic control core that executes under the constraint of a trusted action cone. It uses a reinforcement learning framework to construct a state-action-reward model. The state space is derived from a unified embedding representation, the action space is confined to the action cone, and the reward function comprehensively considers temperature control objectives, power consumption costs, and SLA compliance rates.

[0039] This controller employs a policy gradient-based or soft Actor-Critic structure to continuously optimize its control behavior without violating boundary constraints. The system can be pre-trained in a simulation environment during initial deployment, and an online fine-tuning mechanism can be gradually introduced into the production environment. This design ensures that the controller possesses both optimality and safety / interpretability.

[0040] Based on the above method, a data center multimodal thermal load safety control device in this embodiment includes: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute a data center multimodal thermal load safety control method.

[0041] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0042] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for safe control of multimodal thermal load in a data center, characterized in that, It has the following steps: S1. Data Acquisition and Preprocessing; S2, Multimodal semantic representation generation; S3. Construct a trustworthy action cone; S4. Security policy generation and execution; S5, Control Feedback and Strategy Iteration.

2. The data center multimodal thermal load safety control method according to claim 1, characterized in that, In step S1, the multimodal data access module first continuously collects data from temperature sensors, wind speed field cameras, PDU power metering devices, cooling equipment status, and operation and maintenance log systems. The collection frequency is dynamically allocated according to the type of equipment, and the log data is written asynchronously through a trigger mechanism. After collection, the data undergoes a preprocessing process to standardize and transform it. Time-series data is filled with missing values ​​using sliding window resampling and linear interpolation. Image data is enhanced with edge enhancement and thermal field saliency region extraction. Text logs are constructed with semantic prompts through regular expression filtering and key term highlighting. All data is tagged with a unified timestamp and source label.

3. The data center multimodal thermal load safety control method according to claim 2, characterized in that, In step S2, the preprocessed multimodal data enters the unified semantic representation module, and is converted into a unified high-dimensional embedding vector through the large language model and the multimodal fusion structure. The text log is input into the large language model to obtain a summary vector containing contextual risk semantics. The wind field image is input into the visual Transformer to extract local vortex and abnormal flow direction spatial structure features. Meanwhile, temperature and power consumption time-series data are used to construct sliding window vectors before input, and a time coding mechanism is used to enhance the perception of trends and abrupt changes. Multimodal embedding is finally fused into a unified state vector and sent to the next action constraint generation module.

4. The data center multimodal thermal load safety control method according to claim 3, characterized in that, In step S3, using the current unified semantic representation, combined with the device control limit, industry safety standards and service level agreements, semantic reasoning is performed through a large language model to generate a "trusted action cone". The large model will assess whether the current environment is in an overload, warning or idle range, and generate a set of permissible actions accordingly.

5. A data center multimodal thermal load safety control method according to claim 4, characterized in that, The action set is output in a structured representation and the corresponding boundary meaning is marked. When the downstream controller generates an action, it will first verify the action cone constraint. Actions that exceed the boundary will be rejected or trigger the policy update mechanism.

6. A method for safe control of multimodal thermal load in a data center according to claim 5, characterized in that, In step S4, the safety reinforcement learning controller receives the current state vector and performs policy search in the action space defined by the trusted action cone. The control strategy aims to reduce cooling energy consumption, keep the equipment temperature within a safe range, and maintain the SLA satisfaction rate. The controller dynamically calculates the optimal control path and optimizes the action selection based on long-term cumulative rewards to form the control capability to cope with different loads and hot spot situations. Control commands are sent in real time to the cooling equipment control system, the fan speed regulation module, and the load scheduling interface.

7. A method for safe control of multimodal thermal load in a data center according to claim 6, characterized in that, In step S5, after the control execution is completed, the control response index and multimodal data of the next state will be collected, and the state transition trajectory will be constructed and fed back to the controller. If the system experiences hotspots not being eliminated, equipment startup failure, or SLA deviation after execution, the diagnostic mechanism will be triggered and the current strategy module will be frozen for analysis. During continuous control cycles, the system maintains sliding window indicators for periodic strategy fine-tuning.

8. A data center multimodal thermal load safety control device, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to perform the method according to any one of claims 1 to 7.

Images