Intelligent Control Method for Data Center Cooling Systems Based on Machine Learning

Abstract

This invention belongs to the interdisciplinary field of machine learning and data center thermal management, specifically disclosing a machine learning-based intelligent control method for data center cooling systems. This method configures each cooling unit as an intelligent agent with perception and decision-making capabilities, achieving distributed collaboration through a digital pheromone communication mechanism. It autonomously adjusts the opening of air and water valves based on pheromone concentration gradients and optimizes energy consumption and temperature deviation using a local reinforcement learning model. In the event of a fault, neighboring agents can automatically identify the anomaly and take over the cooling task. Through the above technical solution, this invention achieves sub-second real-time control, highly reliable self-healing operation, and a significant improvement in cooling energy efficiency.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of machine learning and data center thermal management, specifically relating to a machine learning-based intelligent control method for data center cooling systems. Background Technology

[0002] With the continuous evolution of internet infrastructure, the large-scale construction of data centers places stringent demands on thermal management systems. As a critical infrastructure for maintaining the stability of server hardware operation, the energy efficiency of the cooling system directly determines the operational efficiency of the data center. In high-density deployment scenarios, fluctuations in the cooling environment exhibit high complexity, requiring the control system to perform rapid and precise resource scheduling based on real-time load changes.

[0003] Machine learning-based intelligent control methods attempt to achieve optimal parameter matching for cooling equipment by constructing a mapping model between environmental parameters and actuators. This technological direction focuses on achieving coordinated control of equipment such as cooling towers, precision air conditioners, and water pumps through the processing of multi-dimensional sensing data. In the pursuit of optimal energy efficiency, the computational efficiency of the control scheme and the overall reliability of the system become core indicators for evaluating the quality of the technology.

[0004] However, traditional centralized optimization models face computational challenges when dealing with high-dimensional state spaces with massive control variables, resulting in excessively long decision generation cycles that cannot meet the real-time control requirements of dynamic loads. Because the system logic is highly coupled to the central control unit, the risk of single-point failure is extremely high. Once the master control node experiences a logical anomaly or a communication link interruption, the entire cooling network faces the risk of collapse, lacking effective adaptive compensation and local self-healing capabilities. Furthermore, existing technologies are inadequate in handling nonlinear thermal coupling and multi-device collaborative interaction, leading to the coexistence of localized overheating or resource redundancy, impacting the operational security of the data center.

[0005] Therefore, a machine learning-based intelligent control method for data center cooling systems is desired. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent control method for data center cooling systems based on machine learning, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by this invention is an intelligent control method for data center cooling systems based on machine learning, comprising the following specific steps:

[0008] Step 1: Configure each cooling unit in the data center as an intelligent agent with independent perception and decision-making capabilities. The cooling unit includes a precision air conditioner, a chiller, and a cooling tower. Each intelligent agent collects temperature and pressure data of its local area in real time and obtains status information of other cooling units that are physically adjacent to it.

[0009] Step 2: Construct a digital pheromone communication mechanism to enable each agent to conduct distributed negotiation based on the local environment state and the neighbor state without the intervention of a central controller. The digital pheromone is used to characterize the local heat load intensity and the degree of cooling resource demand, and is dynamically propagated and updated among agents.

[0010] Step 3: Based on the concentration gradient of digital pheromones, each intelligent agent autonomously adjusts the working parameters of its actuators, which include air valves and water valves. The adjustment process follows the biomimetic group cooperation rule, so that multiple adjacent intelligent agents automatically form a cooperative alliance when they sense a local overheating event and coordinate the allocation of cooling resources.

[0011] Step 4: Introduce a machine learning model to train on historical running data to generate local decision-making strategies applicable to each agent. The machine learning model is built on a reinforcement learning framework and uses the reward function with local energy consumption and temperature deviation as optimization objectives to continuously iterate and update the behavioral strategies of each agent.

[0012] Step 5: Establish a fault self-healing mechanism. When a cooling unit experiences an operational abnormality or communication interruption, its neighboring intelligent agents automatically identify the abnormal state through changes in digital pheromone signals and dynamically take over the cooling tasks it originally undertook, ensuring that the local thermal environment is maintained within a safe threshold.

[0013] Preferably, in step 1, each cooling unit is equipped with an embedded computing module and a multi-source sensor array. The multi-source sensor array is used to synchronously collect temperature, pressure and airflow velocity data. The embedded computing module is responsible for performing local state assessment and neighbor information fusion processing. All data processing is completed on the device side and does not rely on a remote server.

[0014] Preferably, in step 2, the digital pheromone communication mechanism adopts a lightweight message broadcast protocol. Each agent periodically broadcasts its current digital pheromone value to other agents in its physical neighborhood, while receiving and weighting pheromone signals from its neighbors. The fusion weight is dynamically adjusted according to the physical distance and thermal coupling strength to ensure the accuracy and timeliness of information dissemination.

[0015] Preferably, in step 3, the opening adjustment of the air valve and water valve follows a nonlinear response rule. This rule is output by a local machine learning model. The adjustment range is positively correlated with the digital pheromone concentration gradient. Furthermore, a smoothing inhibition factor is introduced when the concentration gradient changes drastically to prevent mechanical wear and system oscillation caused by frequent actuator movements.

[0016] Preferably, the training data for the machine learning model in step 4 comes from the historical operation logs of the data center, including load change curves, environmental temperature and humidity sequences, and cooling equipment energy consumption records. During the model training process, an experience replay mechanism is used to store high-value state-action pairs, and the local decision function is optimized through a policy gradient algorithm to ensure policy convergence and generalization ability.

[0017] Preferably, the fault self-healing mechanism in step 5 includes an anomaly detection submodule and a task redistribution submodule. The anomaly detection submodule determines whether a neighboring unit has failed by monitoring the continuous absence or deviation of the digital pheromone signal from a preset threshold. The task redistribution submodule re-plans the cooling resource allocation scheme based on the capacity margin of the remaining available cooling units and the heat conduction path to ensure a balanced heat load.

[0018] Preferably, the communication topology between the agents is a dynamically reconfigurable network. The network connection relationship is adjusted in real time according to the physical layout and operating status of the cooling unit. When a cooling unit is added or removed, the system automatically triggers the topology discovery protocol to update the neighbor list of each agent, maintaining the integrity and robustness of distributed collaboration.

[0019] Preferably, the input feature vector of the machine learning model includes local temperature deviation, historical energy consumption trend, neighbor average pheromone concentration and current actuator status, and the output is the target opening degree of the air valve and water valve in the next control cycle. The model inference process is executed on the embedded computing module at fixed time intervals to ensure the real-time generation of control commands.

[0020] Preferably, the attenuation coefficient of the digital pheromone is dynamically set according to the environmental thermal diffusion characteristics. The attenuation process simulates the natural volatilization law of biological pheromones, ensuring that the effective range of the signal in space is limited to a specific area affected by thermal influence, and avoiding inaccurate control caused by long-distance interference.

[0021] Preferably, the formation process of the cooperative alliance adopts a threshold triggering mechanism. When the comprehensive heat load index of a local area exceeds a preset threshold, all intelligent agents in the area automatically enter the cooperative mode, share their respective cooling capacity information, and assign primary and secondary roles according to preset priority rules to lead the execution and coordination of cooling tasks.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. By constructing each cooling unit as an intelligent agent with perception, decision-making and communication capabilities, and introducing a distributed collaborative mechanism of biomimetic swarm intelligence, the traditional centralized architecture's dependence on a single control node is eliminated, solving the problem of excessive computational latency in high-dimensional state spaces, and achieving real-time control response at the second or even sub-second level.

[0024] 2. The system logic is distributed across all nodes in the network. A failure of a single cooling unit will not cause a global paralysis. Surrounding intelligent agents can automatically detect anomalies and dynamically fill in the gaps through a digital pheromone mechanism, which improves the operational reliability and self-healing ability of the cooling system.

[0025] 3. The decision-making strategy based on the local machine learning model can adapt to complex nonlinear thermo-coupling relationships, avoid the coexistence of resource redundancy and local overheating, improve the cooling energy efficiency ratio, and reduce the overall operating energy consumption of the data center. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the core principle framework of the digital pheromone distributed negotiation and information fusion mechanism in this invention;

[0028] Figure 3 This is a logical flowchart of the intelligent agent's autonomous regulation and group cooperative alliance in this invention;

[0029] Figure 4 This is a schematic diagram illustrating the principle framework of local decision-making strategy training and inference based on reinforcement learning in this invention.

[0030] Figure 5 This is a logical flowchart of the cooling unit fault identification and dynamic task redistribution self-healing mechanism in this invention. Detailed Implementation

[0031] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0032] The machine learning-based intelligent control method for data center cooling systems first executes step 1, configuring each cooling unit within the data center as an intelligent agent with independent perception and decision-making capabilities. The cooling units specifically include precision air conditioners deployed inside the server room, chillers providing cooling capacity, and cooling towers deployed on the building roof. Each cooling unit integrates an embedded computing module and a multi-source sensor array during the factory or installation phase. The multi-source sensor array is connected to the analog-to-digital converter interface of the embedded computing module via hard-wired connections. During operation, each intelligent agent collects real-time physical environmental parameters of its local area through the multi-source sensor array, including but not limited to inlet air temperature, outlet air temperature, chilled water inlet pressure, return water pressure, and airflow velocity data at the server rack exhaust end.

[0033] In the specific implementation of step 1, the embedded computing module has a pre-set high-frequency sampling logic with a sampling period set between 10 and 100 milliseconds to ensure rapid capture of changes in the thermal field of the computer room. The raw current or voltage signals output by the sensor array are converted into digital quantities by an analog-to-digital converter circuit and then enter the local state evaluation unit of the embedded computing module. This evaluation unit performs moving average filtering on the temperature and pressure data to eliminate sensor noise. Each agent obtains the state information of other cooling units that are physically adjacent to it through its integrated communication interface using a dynamic discovery protocol. Physical proximity is defined based on the pipe connection relationship or airflow organization path within the data center. Each agent maintains a neighbor list, which records the unique identifier, geographical coordinates, cooling capacity rating, and current operating load rate of neighboring agents. All primary data processing and feature extraction are completed within the embedded computing module on the device side, without involving large-scale data exchange with a remote central server.

[0034] After initial sensing is completed, the process proceeds to step 2, where a digital pheromone communication mechanism is constructed. Each agent utilizes digital pheromones to characterize the local heat load intensity and cooling resource demand. Digital pheromones are defined as a multi-dimensional data structure that includes the temperature deviation, pressure anomaly gradient, remaining cooling capacity, and energy efficiency factor currently sensed by the agent.

[0035] Remaining cooling capacity: refers to the proportion of the maximum additional cooling capacity currently available to the agent to the rated cooling capacity, and its calculation formula is as follows:

[0036] Remaining cooling capacity = (rated cooling capacity - current actual output cooling capacity) / rated cooling capacity;

[0037] The actual output cooling capacity is calculated by multiplying the temperature difference between the inlet and outlet of the chilled water with the real-time flow rate. The rated cooling capacity is the maximum cooling capacity specified by the manufacturer for the cooling unit. The value range of this parameter is [0,1]. The larger the value, the stronger the additional cooling support capability that the intelligent agent can provide.

[0038] Energy efficiency factor: refers to the ratio of the actual energy efficiency ratio of the agent under the current operating conditions to the optimal energy efficiency ratio under the rated operating conditions. Its calculation formula is as follows:

[0039] Energy efficiency factor = Actual cooling energy efficiency ratio / Rated optimal energy efficiency ratio;

[0040] Among them, the actual cooling energy efficiency ratio is the ratio of the current actual output cooling capacity to the current total input power of the cooling unit, and the rated optimal energy efficiency ratio is the highest energy efficiency ratio obtained by the cooling unit under standard operating conditions. The value range of this parameter is (0,1], and the closer the value is to 1, the higher the energy efficiency level of the current operating condition of the intelligent agent.

[0041] This digital pheromone communication mechanism enables distributed negotiation among agents without the need for a central controller. The mechanism employs a lightweight message broadcast protocol, where each agent broadcasts its current digital pheromone value to members of its neighbor list at preset time steps.

[0042] In step 2, the dynamic propagation and updating of digital pheromones follow the volatilization and diffusion logic of simulated biological characteristics. The concentration of digital pheromones decreases with increasing propagation distance, and the attenuation coefficient is dynamically set according to the environmental thermal diffusion characteristics. Specifically, if the physical distance between the receiving agent and the source agent increases, the effective pheromone concentration propagated to the receiving end decreases according to the square of the distance.

[0043] The attenuation coefficients of digital pheromones are divided into two categories: spatial propagation attenuation coefficients and time-dimensional evaporation coefficients. Their values ​​are determined based on the following rules:

[0044] Spatial propagation attenuation coefficient: This coefficient characterizes the attenuation of pheromone concentration over propagation distance. Its value is strongly correlated with the airflow organization within the data center, and is determined based on measured values ​​of the air thermal diffusivity within the data center. The specific setting rules are as follows:

[0045] For conventional air-cooled server rooms employing a bottom-supply, top-return airflow system, the baseline value for the spatial propagation attenuation coefficient is set to 0.8 to 1.2 m. -2 ;

[0046] For computer rooms employing an upward supply and return airflow system, the baseline value for the spatial propagation attenuation coefficient is set at 1.5 to 2.0m. -2 ;

[0047] For liquid-cooled system computer rooms, the baseline value for the spatial propagation attenuation coefficient is set at 0.3 to 0.6m. -2 ;

[0048] In practical applications, the thermal diffusion can be dynamically adjusted within a range of ±30% of the baseline value based on the measured values ​​of thermal diffusion in different areas of the computer room.

[0049] Time-dimensional evaporation coefficient: This characterizes the time decay properties of old pheromone data, i.e., the retention coefficient within each calculation cycle. Its value is determined based on the sampling period and the dynamic rate of change of the data center's thermal environment. Specific setting rules are as follows:

[0050] Retention factor = 1 - (Sampling period / Thermal environment time constant);

[0051] The sampling period is 10 to 100 ms as set in step 1. The thermal environment time constant is the minimum time scale at which the thermal environment of the computer room changes significantly. The value is 5 to 10 s for air-cooled computer rooms and 2 to 5 s for liquid-cooled computer rooms. The retention coefficient is fixed at 0.9 to 0.999 to ensure the temporal continuity of the pheromone signal and avoid interference from outdated data on current decisions.

[0052] Meanwhile, digital pheromones also decay over time; that is, older pheromone data is multiplied by a retention factor less than 1 in each computation cycle to ensure the system's sensitivity to instantaneous load changes. Each agent performs a weighted fusion process after receiving pheromone signals from its neighbors. The fusion weights are determined by two variables: physical distance and thermal coupling strength. Thermal coupling strength is characterized by the temperature correlation coefficient between the two devices in historical operational data.

[0053] The specific calculation method for the thermal coupling coefficient corresponding to the thermal coupling strength is as follows:

[0054] Select continuous Historical operational data for each sampling period The value range is from 1000 to 10000 sampling periods, based on the outlet temperature sequence of agent i. The rack-side temperature monitoring sequence of intelligent agent j Based on this, the temperature correlation between the two is calculated using the Pearson correlation coefficient. The calculation formula is as follows:

[0055]

[0056] In the formula, for and covariance, for Standard deviation of the sequence for Standard deviation of the sequence;

[0057] The calculated Pearson correlation coefficient Normalization is performed to obtain the thermal coupling coefficient. That is, only positively correlated coupling relationships are retained, and the value range is [0,1]. The closer the value is to 1, the stronger the agent. The operational status of the intelligent agent The stronger the influence of the thermal environment in the region, the higher the degree of thermal coupling between the two.

[0058] Based on the aforementioned thermal coupling coefficient and physical distance, the fusion weights are... The final calculation formula is:

[0059]

[0060] In the formula, For intelligent agents With intelligent agents The physical straight-line distance between them, in meters; This is a local constant with a value of 1×10^-6, used to avoid calculation errors when the denominator is 0.

[0061] The fused pheromone field provides the agent with a profile of the overall thermal load in its surrounding area, enabling it to perceive potential thermal threats outside its own area of ​​responsibility.

[0062] Step 3 is then executed, where each agent autonomously adjusts the operating parameters of its actuators based on the concentration gradient of the digital pheromones. These actuators include air valves that regulate the airflow of cold air and water valves that regulate the flow of chilled water. The adjustment process strictly follows biomimetic swarm collaboration rules. Specifically, the control logic within each agent calculates the difference between the local pheromone concentration and the surrounding pheromone concentration, i.e., the concentration gradient. When an overheating event occurs in a localized area, the pheromone concentration at that point rapidly increases, forming a high-concentration center. Upon sensing this concentration gradient, surrounding agents automatically determine whether they possess the capability to provide support. If the remaining cooling capacity of a surrounding agent exceeds a preset threshold, that agent automatically joins the collaborative alliance.

[0063] In the formation of the cooperative alliance in step 3, the system employs a threshold triggering mechanism. When the comprehensive heat load index of a local area—that is, the weighted sum of the local temperature deviation and the average pheromone concentration of its neighbors—exceeds a preset cooperative triggering threshold, all agents within that area automatically switch from independent operation mode to cooperative mode. In cooperative mode, each agent shares its real-time cooling output capacity and is assigned primary and secondary roles according to preset priority rules. The priority rules consider the energy efficiency ratio and response speed of the devices; devices with higher energy efficiency ratios and closer to the hotspot are assigned the primary role, responsible for the main incremental cooling output, while other devices provide redundant support as auxiliary roles. The opening adjustment of air valves and water valves follows a nonlinear response rule, which is calculated and output by a local machine learning model. The adjustment magnitude is positively correlated with the digital pheromone concentration gradient; that is, the larger the gradient, the larger the opening adjustment. To prevent system oscillation, when the rate of change of the concentration gradient exceeds a preset protection threshold, the system introduces a smoothing suppression factor to limit the maximum displacement of the actuators per unit time, thus avoiding mechanical wear and hydraulic imbalance.

[0064] Step 4 involves training a machine learning model on historical operational data. This model, built on a reinforcement learning framework, uses a reward function with local energy consumption and temperature deviation as optimization objectives to continuously iterate and update the behavioral strategies of each agent, running within the embedded computing modules of each agent. The training data for the machine learning model comes from log files accumulated over long-term operation in the data center, including temperature distribution curves under different IT loads, different ambient humidity and temperature sequences, and power consumption records of various cooling devices under different operating conditions. The model employs an experience replay mechanism, storing high-value state-action-reward tuples in local memory.

[0065] In the specific training process of step 4, the local decision function is continuously optimized through the policy gradient algorithm. The optimization objective is set as the weighted minimization of local energy consumption and temperature deviation. Specifically, the design logic of the reward function is as follows: when the regional temperature is maintained within the set range and the energy efficiency ratio is improved, a positive reward value is given; when the temperature exceeds the safety threshold or the actuator swings frequently, a negative reward value is given.

[0066] To ensure that agents learn distributed cooperative strategies rather than individual competitive strategies, and to guarantee the stable convergence of the multi-agent system, the reward function... The design employs a three-tiered structure: local individual rewards, global collaborative rewards, and penalties for competitive behavior. The specific formula is as follows:

[0067]

[0068] Among them: Local individual rewards This is used to drive the intelligent agent to achieve local cooling and energy efficiency optimization goals. The calculation formula is:

[0069]

[0070] In the formula, As a temperature control reward, R_temp is set to 10 when the area temperature is within the set safe range; when the temperature exceeds the safe range, ... , This is the actual temperature. This is the upper limit of the safe range;

[0071] As a reward for energy efficiency optimization, when the agent's current energy efficiency ratio is higher than that of the previous cycle, The value is 5; otherwise, the value is 0. , These are the weighting coefficients. ;

[0072] ②Global Collaborative Rewards This is used to drive agents to form cooperative strategies and avoid global inaccuracies caused by individual optimality. The calculation formula is:

[0073]

[0074] In the formula, A temperature reward is given to neighboring agents when the temperatures of all neighboring agents' jurisdictions are within a safe range. The value is 8; otherwise, the value is 0. To provide a coordinated regional energy efficiency reward, when the average energy efficiency ratio of all agents within the collaborative alliance increases, The value is 5; otherwise, the value is 0. , These are the weighting coefficients. This reward ensures that the agent's decision-making not only focuses on its own region, but also takes into account the state of the surrounding collaborative regions, thus guiding collaborative behavior through the reward mechanism.

[0075] Punishment for competitive behavior It is used to suppress malicious competitive behavior of agents and avoid policy conflicts between multiple agents. The calculation formula is:

[0076]

[0077] In the formula, As a penalty for action conflicts, when an agent's opening adjustment direction is opposite to that of the dominant agent in the cooperative alliance, and this leads to an increase in the temperature gradient of the cooperative region, The value is 10, otherwise it is 0; As a penalty for action frequency, when the number of actions performed by the actuator per unit time exceeds a preset threshold, , This represents the number of actions that exceed the threshold; otherwise, the value is 0. , These are the weighting coefficients. ;

[0078] During model training, all agents share the computational logic of the collaborative reward term. At the same time, a training framework of centralized training and distributed execution is adopted. In the offline training stage, the policy network is optimized based on global state information, and in the online inference stage, the decision is output based only on the locally encoded state space. This ensures that the distributed system can stably converge to the globally optimal collaborative policy, rather than the individually optimal competitive policy.

[0079] The input feature vector of the machine learning model includes local temperature deviation, historical energy consumption trends, neighboring average pheromone concentration, and the current state of the air and water valves. The output is the target opening degree of the actuators for the next control cycle.

[0080] The specific implementation rules for encoding digital pheromone multidimensional data into the reinforcement learning state space are as follows:

[0081] The state space S of reinforcement learning consists of local state components. pheromone fusion state components Neighbor cooperative state components It consists of three parts, with a fixed dimension of 18, and the specific encoding method is as follows:

[0082] Local state components The dimension is 6-dimensional, including the normalized local temperature deviation, local pressure anomaly gradient, current air valve opening, current water valve opening, current energy consumption power, and remaining cooling capacity. All parameters are mapped to the [0,1] interval through min-max normalization.

[0083] Pheromon fusion state components The dimensions are 8-dimensional. The fused multidimensional pheromone data is encoded, including: local pheromone baseline concentration, neighboring pheromone weighted average concentration, pheromone concentration gradient magnitude, pheromone concentration gradient direction, energy efficiency factor weighted average, pressure sensitivity factor maximum value, temperature deviation maximum value, and heat load change rate. All parameters are normalized to the [-1,1] interval through z-score.

[0084] Neighbor Cooperative State Components The dimension is 4-dimensional, including the number of neighboring agents, the average remaining cooling capacity of neighbors, the minimum distance between neighbors, and the maximum thermal coupling coefficient between neighbors. All parameters are normalized to the [0,1] interval.

[0085] The encoded state space S serves as the sole input to the reinforcement learning policy network, enabling a deep fusion of the digital pheromone mechanism and the local reinforcement learning model. This ensures that the agent's decisions take into account both local state and global collaborative information.

[0086] The model inference process is executed at fixed time intervals, such as generating a control command every 500 milliseconds. Since training is performed offline or in the background, the online inference process involves only simple matrix operations, ensuring the real-time generation of control commands. As runtime accumulates, the model can continuously learn the complex nonlinear thermodynamic coupling laws within the data center, achieving more accurate predictive control than traditional PID algorithms.

[0087] Finally, step 5 is executed to establish a fault self-healing mechanism. This mechanism consists of an anomaly detection submodule and a task redistribution submodule. The anomaly detection submodule continuously monitors digital pheromone signals from neighboring agents. When a cooling unit experiences a hardware failure, power outage, or communication link failure, the digital pheromone signals it sends will show continuous gaps or values ​​that significantly deviate from the physically possible preset threshold. Once a neighboring agent does not receive a heartbeat signal or pheromone update from a specific neighbor within three consecutive sampling periods, it is determined that the neighboring unit is in a failed state.

[0088] In the task reassignment phase of step 5, the system no longer relies on manual intervention. The task reassignment submodule automatically calculates based on the capacity margin of the remaining available cooling units. The calculation logic follows the principle of optimal heat conduction path, that is, prioritizing the scheduling of other cooling devices on the same airflow loop as the failed unit. The cooling task originally undertaken by the failed unit, i.e., its corresponding heat load share, is dynamically mapped to the decision space of the surrounding agents. The surrounding agents automatically increase their pheromone perception sensitivity coefficient, generating greater cooling output in their local decision-making strategies to fill the cooling gap caused by the failure. This self-healing process is completed within seconds, ensuring that the local thermal environment of the data center is always maintained within a safe threshold, preventing localized overheating of the rack due to a single point of failure.

[0089] During the implementation of the above method, the communication topology between the agents is not fixed but configured as a dynamically reconfigurable network. When the data center is expanded, i.e., a new cooling unit is added, the new device automatically triggers the topology discovery protocol after connecting to the network. The new device broadcasts its physical coordinates and capability parameters to the entire network. Upon receiving the information, existing agents in its vicinity automatically update their neighbor lists and include the new device in the propagation path of digital pheromones. Similarly, when a device is taken offline for maintenance, the system automatically reconfigures the communication weights to ensure the integrity and robustness of the distributed collaborative logic.

[0090] To further improve system performance, each agent in Example 1 will also make adjustments based on real-time collected outdoor environmental parameters during the adjustment process in step 3. When the outdoor temperature is lower than the preset natural cooling threshold, the agent sends a higher-priority call request to the cooling tower agent via digital pheromones to prioritize the use of external natural cooling sources, while simultaneously reducing the compressor power consumption of the chiller by adjusting the water valve opening. This biomimetic collaborative mechanism enables the entire cooling system to exhibit self-organizing characteristics similar to a biological community, spontaneously seeking the operating state with the lowest energy consumption and the highest environmental stability.

[0091] At the data processing level, the internal memory of the embedded computing module is divided into a real-time data area, a historical buffer, and a policy model area. The real-time data area stores the raw, millisecond-level data from the current sensor feedback; the historical buffer stores key state features from the past 24 hours for online fine-tuning of the local machine learning model; and the policy model area stores the trained neural network weight parameters. All this data is exchanged via an internal high-speed bus, ensuring that the total latency from perception to decision-making to execution is kept within 200 milliseconds.

[0092] Example 2: Building upon Example 1, Example 2 provides a more targeted intelligent control solution for data centers with large-scale liquid cooling systems. In Example 2, the cooling unit is further expanded to include a liquid cooling distribution unit (CDU) and a secondary-side circulation pump. The data collected in Step 1 includes coolant flow rate, inlet and outlet temperature difference, and the status of the leak detection sensor.

[0093] In the construction of the digital pheromone in step 2, Example 2 introduces the concept of "pressure balance pheromone". Because the liquid cooling system is extremely sensitive to the pressure balance of the piping network, a pressure sensitivity factor is added to the digital pheromone data structure. When an increase in demand in a certain liquid cooling branch causes fluctuations in the main pipe pressure, this information is quickly synchronized among all CDU agents through the digital pheromone mechanism.

[0094] In Example 2, the actuator adjustment in step 3 is manifested as the coordinated linkage between the variable frequency pump speed and the proportional-integral control valve. A compensation coefficient for fluid resistance is added to the nonlinear response rule. When the digital pheromone concentration gradient indicates a sudden increase in computing load in a certain area, the collaborative alliance includes not only adjacent cooling units but also upstream cold source-side agents. By pre-increasing the secondary-side flow rate and fine-tuning the primary-side heat exchange temperature difference under the guidance of a local machine learning model, rapid heat transfer is achieved.

[0095] In step 4 of Example 2, the reward function of the machine learning model is further refined. In addition to energy consumption and temperature deviation, a system lifespan prediction factor is added. By monitoring the operating frequency of the actuators and the vibration data of the pump group, the machine learning model learns an action sequence that can reduce mechanical fatigue. For example, under the premise of meeting cooling requirements, the model will tend to choose a smoother frequency adjustment curve rather than a step-like switching action.

[0096] Regarding the fault self-healing mechanism, Example 2 adds automatic hydraulic balance recovery logic in step 5. When a pump failure is detected in a liquid-cooled loop, the task redistribution submodule calculates the hydrodynamic distribution of the entire network and adjusts the valve openings of other branches, using pressure difference changes to guide the coolant to automatically compensate the affected area. This self-healing method, based on a combination of physical laws and biomimetic logic, improves the operational safety of the liquid-cooled data center.

[0097] Regarding the communication topology, Example 2 employs a dual-redundant ring network structure, with digital pheromone signals transmitted on two mutually redundant physical channels. When an anomaly is detected in a communication branch, the agent automatically switches to the backup channel and updates the topology weights to ensure that distributed negotiation is not interrupted due to physical layer failures.

[0098] Example 3: Example 3 focuses on the application of the present invention in an edge computing data center scenario. Edge centers are typically small in scale, but their environments change drastically and they lack professional operation and maintenance personnel.

[0099] In step 1 of Example 3, the agent integrates the function of monitoring ambient humidity and incorporates the humidity data into the perception vector. In step 2, the propagation mechanism of digital pheromones adds adaptive adjustment to the communication bandwidth. When the edge network environment is poor, the agent automatically compresses the pheromone data packets, retaining only the concentration gradient and key alarm bits to reduce network load.

[0100] Step 3 in Example 3 specifically emphasizes the agent's response to extreme weather. When the outdoor temperature sensor detects a sharp rise in ambient temperature, a cooperative alliance is formed in advance, and all agents enter a pre-cooling mode. Through a digital pheromone mechanism, a cooling margin is reserved in advance to prevent temperature overshoot caused by thermal inertia of IT equipment.

[0101] In step 4, considering the limited computing resources at the edge, a lightweight quantized neural network was adopted for the machine learning model. The model parameters were reduced from 32-bit floating-point numbers to 8-bit integers, reducing storage space usage while maintaining inference accuracy above 95%. The online learning process of reinforcement learning was changed to periodic triggering, for example, iteratively updating the model weights during periods of low IT business activity.

[0102] In step 5 of Example 3, a one-click emergency mode is added to the fault self-healing mechanism. When multiple neighboring agents simultaneously determine that a certain area has failed and cannot be compensated for by existing resources, the system will send a signal to the IT layer through digital pheromones to trigger virtual machine migration or computing task speed limiting, actively reducing the heat load from the heat source side, and realizing a deep-level linkage between the cooling system and the IT system.

[0103] In a specific application example, we deployed the above method in a medium-sized data center with 500 racks. The system includes 20 precision air conditioners, 4 chiller units, and 4 cooling towers. Each device is equipped with an embedded computing module based on the ARM architecture.

[0104] During the initial operation phase, each agent established a topology map containing physical coordinates by executing step 1. The sensor array updated local temperature and pressure data every 20 milliseconds. When simulating a local hotspot event caused by high-density server computing, the digital pheromone concentration generated by the precision air conditioner in the affected area through step 2 rapidly increased, rising from a baseline value of 0.2 to 0.85.

[0105] Upon receiving the pheromone signal, three precision air conditioners within a 5-meter radius triggered the coordination rules in step 3. These three devices automatically calculated their remaining cooling capacities to be 30%, 25%, and 40%, respectively. According to the priority rules, the device closest and with the largest remaining capacity was selected as the dominant agent, and its air valve opening increased from 50% to 85% within 3 seconds. The auxiliary agent simultaneously fine-tuned its water valve opening to increase the chilled water flow.

[0106] During operation, the machine learning model in step 4 continuously monitors energy efficiency indicators. By comparing historical data, the system found that after adopting the distributed collaborative mechanism, the total energy consumption of the cooling system was reduced by 18% under the same load. The uniformity of temperature distribution was improved, and the deviation between the local maximum temperature and the average temperature was reduced by 2.3 degrees Celsius.

[0107] When a simulated precision air conditioner goes offline due to a power module failure, step 5 is executed. The two adjacent intelligent agents detect the missing pheromone signal within 60 milliseconds. The task redistribution submodule immediately calculates a new control strategy, and neighboring devices compensate for the cooling in the affected area by increasing airflow. The entire self-healing process has no negative impact on the operation of the IT equipment, and local temperature fluctuations are controlled within 1.5 degrees Celsius.

[0108] Through the detailed descriptions of the three embodiments above, it can be seen that the present invention achieves decentralized control of data center cooling systems by simulating biological swarm intelligence. Each cooling unit is no longer a passive actuator, but an intelligent node capable of autonomously sensing, making decisions, and cooperating with its surroundings. This architecture solves the computational latency problem in large-scale systems and also endows the system with strong survivability and energy efficiency optimization potential. Digital pheromones, as information carriers, cleverly transform the complex global optimization problem into a simple local gradient following problem, reducing the algorithm's requirements for computing resources.

[0109] In terms of specific implementation details, the embedded computing module adopts a multi-threaded architecture. High-priority threads handle sensor data acquisition and actuator control, medium-priority threads handle the parsing and fusion of digital pheromones, and low-priority threads are responsible for the online learning of the machine learning model. This task scheduling method ensures that the most critical cooling task receives immediate response under any operating condition. At the communication level, carrier sense multiple access (CMA) technology with collision detection is used to guarantee channel utilization when a large number of agents are broadcasting pheromone signals.

[0110] Furthermore, this invention supports protocol integration with data center environmental monitoring systems. Although this method does not rely on a central controller, each intelligent system periodically sends compressed operational summaries to the monitoring backend for macro-level control and report generation by administrators. This "distributed control, centralized monitoring" model retains system flexibility while also ensuring manageability.

[0111] In terms of machine learning model optimization, in addition to reinforcement learning, Long Short-Term Memory (LSTM) networks can be integrated to predict future heat load trends based on actual needs. The prediction results are encoded into the early warning field of digital pheromones, enabling surrounding agents to adjust the pre-cooling amount in advance before the hotspot actually forms, further eliminating regulatory lag.

[0112] Regarding the regulation of cooling towers and chillers, this invention considers the coupling of the hydraulic system. When multiple chillers operate in parallel, the intelligent agents of each chiller maintain a common supply water temperature through pheromone negotiation. If a chiller operates in a low-efficiency range, its pheromones will exhibit "repulsive" properties, prompting other high-efficiency chillers to take on more load, thereby achieving optimal overall energy efficiency.

[0113] For the nonlinear control of air and water valves, this invention introduces the textual logic of a fuzzy logic controller into the local strategy. When the temperature deviation is within a small range, a fine-tuning mode is adopted, and the stepper motor's increment is set to the minimum value; when the deviation enters the warning range, a fast response mode is switched to, increasing the adjustment step size. This multimodal switching logic is integrated into the output mapping of the machine learning model.

[0114] During fault identification, the system also has the ability to self-diagnose sensor drift. If the temperature data sensed by a certain agent has a long-term and fixed significant deviation from the temperature field represented by the average pheromones of its neighbors, the anomaly detection submodule will determine that the sensor has experienced accuracy drift and automatically start a calibration procedure or reduce the weight of the sensor data in subsequent decisions.

[0115] In summary, this invention deeply integrates biological swarm intelligence with modern machine learning algorithms to construct a data center cooling system with self-organizing, self-learning, and self-healing capabilities. Its technical features encompass a complete and highly reliable technical solution, ranging from underlying multi-source sensing and mid-level pheromone communication mechanisms to top-level reinforcement learning strategy optimization, as well as extended implementation methods for different application scenarios.

[0116] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent regulation of a data center cooling system based on machine learning, characterized in that, Includes the following steps: Step 1: Configure each cooling unit in the data center as an intelligent agent with independent perception and decision-making capabilities. Use a multi-source sensor array integrated in the cooling unit to collect physical environment parameters of its local area in real time, and obtain the status information of other cooling units that are physically adjacent to it. Step 2: Construct a digital pheromone communication mechanism to enable each agent to conduct distributed negotiation based on the local environment state and the neighbor state without the intervention of a central controller. The digital pheromone is used to characterize the local heat load intensity and the degree of cooling resource demand, and is dynamically propagated and updated among agents. Step 3: Based on the concentration gradient of digital pheromones, each agent autonomously adjusts the working parameters of its actuators. The adjustment process follows the biomimetic group cooperation rules, so that multiple neighboring agents automatically form a cooperative alliance when they sense a local overheating event and coordinate the allocation of cooling resources. Step 4: Introduce a machine learning model to train on historical operating data and generate local decision-making strategies applicable to each agent. The machine learning model is built on a reinforcement learning framework and uses a reward function with local energy consumption and temperature deviation as optimization objectives to continuously iterate and update the behavioral strategies of each agent. Step 5: Establish a fault self-healing mechanism. When a cooling unit experiences an operational abnormality or communication interruption, its neighboring intelligent agents automatically identify the abnormal state through changes in digital pheromone signals and dynamically take over the cooling tasks it originally undertook, ensuring that the local thermal environment is maintained within a safe threshold. In step 2, each agent simultaneously receives and weights and fuses pheromone signals from its neighbors; The fusion weight is dynamically adjusted based on physical distance and thermal coupling strength. The thermal coupling strength is characterized by the temperature correlation coefficient between the two devices in historical operating data; The fault self-healing mechanism in step 5 includes dynamically mapping the heat load share originally borne by the failed unit to the decision space of the surrounding intelligent agent. The surrounding intelligent agent automatically increases its pheromone perception sensitivity coefficient and generates a larger cooling output in the local decision strategy to fill the cooling gap caused by the fault. In step 3, the adjustment range of the working parameters of the actuator is positively correlated with the digital pheromone concentration gradient, that is, the larger the concentration gradient, the larger the opening adjustment. When the rate of change of concentration gradient exceeds the preset protection threshold, the system introduces a smoothing inhibition factor to limit the maximum displacement of the actuator per unit time, thereby avoiding mechanical wear and hydraulic imbalance.

2. The machine learning based intelligent regulation method for data center cooling system according to claim 1, wherein: In step 1, each cooling unit is equipped with an embedded computing module and a multi-source sensor array; The multi-source sensor array is used to synchronously collect temperature, pressure, and airflow velocity data; the physical environment parameters include inlet temperature, outlet temperature, chilled water inlet pressure, return water pressure, and airflow velocity data at the cabinet exhaust end; The embedded computing module has a pre-set high-frequency sampling logic with a sampling period set between 10 milliseconds and 100 milliseconds. The raw current or voltage signal output by the sensor array is converted into a digital quantity by an analog-to-digital converter circuit, and then enters the local state evaluation unit of the embedded computing module. The local condition assessment unit performs a moving average filter on the temperature and pressure data; At the same time, each agent obtains the status information of other cooling units that are physically adjacent to it through its integrated communication interface using a dynamic discovery protocol; The status information is recorded in the neighbor list, which includes the unique identifier of the neighbor agent, its geographical coordinates, cooling capacity rating, and current operating load rate.

3. The machine learning based intelligent regulation method for data center cooling system according to claim 1, wherein: In step 2, the digital pheromone communication mechanism adopts a lightweight message broadcasting protocol, in which each agent periodically broadcasts its current digital pheromone value to other agents in its physical neighborhood. The digital pheromone is defined as a multidimensional data structure that includes the temperature deviation, pressure anomaly gradient, remaining cooling capacity, and energy efficiency factor currently perceived by the agent. The fused pheromone field provides the agent with a profile of the overall thermal load in its surrounding area, enabling it to perceive potential thermal threats outside its own area of ​​responsibility. In the weighted fusion process, the received neighbor pheromone values ​​are multiplied by the corresponding fusion weights, and the product results of all neighbors are summed to serve as the basis for correcting the local pheromone field.

4. The intelligent control method for data center cooling systems based on machine learning according to claim 3, characterized in that: The dynamic propagation and updating of the digital pheromones follow the volatilization and diffusion logic that simulates biological characteristics; The concentration value of the digital pheromone decreases as the propagation distance increases, and the attenuation coefficient is dynamically set according to the environmental thermal diffusion characteristics. The attenuation law is described as follows: if the physical distance between the receiving agent and the source agent increases, the effective pheromone concentration propagated to the receiving end decreases according to the square of the distance. Digital pheromones also decay over time; old pheromone data is multiplied by a retention factor of less than 1 in each calculation cycle. The decay process of digital pheromones simulates the natural volatilization of biological pheromones, ensuring that the effective range of the signal in space is limited to the thermally affected area, thus avoiding inaccurate regulation caused by long-distance interference.

5. The intelligent control method for data center cooling systems based on machine learning according to claim 1, characterized in that: The formation process of the collaborative alliance in step 3 adopts a threshold triggering mechanism; When the comprehensive heat load index of a local area exceeds the preset collaborative trigger threshold, all intelligent agents in the area automatically switch from independent operation mode to collaborative mode. The comprehensive heat load index is a weighted sum of local temperature deviation and neighboring average pheromone concentration; In collaborative mode, each agent shares its real-time cooling output capability and is assigned primary and secondary roles according to preset priority rules. The priority rule takes into account the energy efficiency ratio and response speed of the devices, assigning devices with higher energy efficiency ratios and closer to the hotspot as the leading devices, responsible for the main incremental cooling output, while other devices serve as auxiliary devices to provide redundant support. Each intelligent agent communicates its remaining cooling capacity with each other via digital pheromones. When the remaining cooling capacity of a neighboring intelligent agent exceeds a preset threshold, that intelligent agent automatically joins the cooperative alliance.

6. The intelligent control method for data center cooling systems based on machine learning according to claim 1, characterized in that: The process of adjusting the working parameters of the actuator in step 3 is to calculate the concentration gradient between the local pheromone concentration and the surrounding pheromone concentration. The actuator includes an air valve and a water valve. The opening adjustment of the air valve and the water valve follows a nonlinear response rule, which is output by a local machine learning model. The agent will also make corrections based on real-time collected outdoor environmental parameters. When the outdoor temperature is lower than the preset natural cooling threshold, the agent will send a higher priority call request to the cooling tower agent through digital pheromones to make priority use of external natural cooling sources.

7. The intelligent control method for data center cooling systems based on machine learning according to claim 1, characterized in that: The training data for the machine learning model in step 4 comes from the historical operation logs of the data center, including load change curves, environmental temperature and humidity sequences, and cooling equipment energy consumption records. During model training, an experience replay mechanism is used to store triples consisting of high-value states, actions, and rewards in local memory, and the local decision function is optimized through a policy gradient algorithm. The design logic of the reward function is as follows: when the regional temperature is maintained within the set range and the energy efficiency ratio is improved, a positive reward value is given; when the temperature exceeds the safety threshold or the actuator operation frequency exceeds the preset frequency threshold, a negative reward value is given. The optimization objective is set as the weighted minimization of local energy consumption and temperature deviation. By continuously adjusting the weight parameters of the decision function, the expected cumulative reward value is maximized.

8. The intelligent control method for data center cooling systems based on machine learning according to claim 7, characterized in that: The input feature vector of the machine learning model includes local temperature deviation, historical energy consumption trend, neighbor average pheromone concentration, and current actuator status; the actuator status includes the current air valve opening and water valve opening. The output is the target opening degree of the air valve and water valve for the next control cycle; The model inference process is executed on the embedded computing module at fixed time intervals to ensure the real-time generation of control commands; The machine learning model uses a lightweight quantized neural network, reducing the model parameters from 32-bit floating-point numbers to 8-bit integers, and transforming the computational logic involved in the inference process into fixed-point matrix multiplication. At the data processing level, the memory inside the embedded computing module is divided into a real-time data area, a historical buffer, and a policy model area. Data is exchanged through an internal high-speed bus to ensure that the total latency from perception to decision-making to execution is controlled within 200 milliseconds.

9. The intelligent control method for data center cooling systems based on machine learning according to claim 1, characterized in that: The fault self-healing mechanism in step 5 includes an anomaly detection submodule and a task reassignment submodule. The anomaly detection submodule determines whether a neighboring cell has failed by monitoring the continuous absence or significant deviation of the digital pheromone signal from a preset threshold. If a neighboring agent does not receive a heartbeat signal or pheromone update from its neighbor within three consecutive sampling periods, the neighboring unit is determined to be in a failed state. The task redistribution submodule re-plans the cooling resource allocation scheme based on the capacity margin of the remaining available cooling units and the heat conduction path to ensure a balanced heat load. The task redistribution process follows the principle of optimal heat conduction path, prioritizing the scheduling of other cooling devices that are on the same airflow loop as the faulty unit.

10. The intelligent control method for data center cooling systems based on machine learning according to claim 1, characterized in that: The communication topology between the various intelligent agents is a dynamically reconfigurable network, and the network connection relationship is adjusted in real time according to the physical layout and operating status of the cooling unit. When a cooling unit is added or removed, the system automatically triggers the topology discovery protocol to update the neighbor list of each agent, maintaining the integrity and robustness of distributed collaboration. In the application scenario of liquid cooling system, the cooling unit includes a liquid cooling distribution unit and a secondary side circulation pump, and the collected data includes coolant flow rate, inlet and outlet temperature difference and the status of leakage detection sensor; A pressure-sensitive factor is added to the data structure of the digital pheromone. When the demand of a certain liquid cooling branch increases, causing the main pipe pressure to fluctuate, the status information is synchronized among all liquid cooling distribution unit agents through the digital pheromone mechanism. The fault self-healing mechanism also includes hydraulic balance automatic recovery logic. When a pump failure is detected in a certain liquid cooling loop, the task redistribution submodule calculates the hydrodynamic distribution of the entire network and uses pressure difference changes to guide the coolant to automatically compensate the affected area by adjusting the valve opening of other branches.

Images