A prefabricated data center cooling control method and system based on AI

Abstract

This invention relates to the field of automatic control technology, and more particularly to an AI-based prefabricated data center cooling control method and system. The method includes the following steps: acquiring a texture map and a depth map of the rack under test; analyzing the material information of the texture map to determine the theoretical heat capacity benchmark of the equipment and the geometric features of the depth map to determine the equivalent volumetric heat capacity of the rack under test; calculating the tortuosity index of the rack under test based on the spatial gradient features of the depth map to characterize heat exchange efficiency; constructing a feedforward cooling power curve by combining the equivalent volumetric heat capacity, the tortuosity index, the geometric volume of the rack under test, and the initial heat dissipation temperature difference; and controlling the operation of a precision air conditioning unit based on this curve. This invention, through intelligent perception of the appearance of the rack under test, predicts its actual heat load and heat dissipation characteristics, realizing a shift from passive feedback to active feedforward temperature control mode, improving temperature control accuracy and operating efficiency.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and in particular to an AI-based method and system for cooling control of prefabricated data centers. Background Technology

[0002] In the operation of modern data center digital monitoring and management systems, the automated control efficiency of the cooling system is directly related to the operational safety of IT equipment and the overall energy efficiency of the data center. The racks and IT equipment deployed within prefabricated server rooms exhibit extremely high dynamism and thermal inertia differences. Different server densities and their internal cooling aisle distribution can have complex nonlinear effects on heat load. Therefore, developing intelligent systems with load sensing and feedforward control capabilities is crucial for achieving supply-demand balance and optimizing energy efficiency.

[0003] In related technologies, a feedback regulation mechanism based on an environmental monitoring platform is adopted. Parameters are collected by deploying ambient temperature sensors at key locations in the hot and cold aisles. When temperature fluctuations occur or hot spot temperature rises exceed limits, a bus-triggered algorithm is used to instruct the air conditioning system to adjust its output. This primarily utilizes temperature sensing deviation feedback to drive operating condition switching, providing basic stability under steady-state heat dissipation requirements and ensuring the computer room environment remains within the preset process requirements.

[0004] However, this conventional control strategy based on temperature feedback has limitations in dealing with the rapid transient changes in high-density loads in prefabricated data centers: existing control logic generally treats the computer room as a uniform thermal field, lacking an analysis of the physical form and real-time thermal characteristics of the equipment, making it difficult to distinguish the essential differences in thermal response speed and heat dissipation performance among different business units. Due to the unavoidable time lag between equipment heating and ambient temperature perception, and the lack of means to assess heat transfer efficiency in complex environments, control decisions can only be made based on the consequences of temperature changes. This adjustment struggles to accurately match cooling output with the real-time thermal demands of the load, and is prone to response delays due to inaccurate identification of the thermal inertia of high-density nodes, leading to oscillations in the computer room temperature field, increased energy consumption of the air conditioning system, and potential risks of localized overheating of computing nodes. Summary of the Invention

[0005] To address the technical problems of existing control schemes that fail to accurately identify the thermal inertia differences of heterogeneous devices, resulting in delayed cooling regulation response and mismatch between output power and real-time heat load, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an AI-based prefabricated data center cooling control method, the method comprising the steps of:

[0007] The process involves acquiring texture and depth maps of the rack under test within a prefabricated data center; analyzing the material information of the texture map to determine the theoretical heat capacity benchmark for the equipment within the rack; modifying the theoretical heat capacity benchmark by combining gradient parameters characterizing the geometric undulations of the depth map and texture feature parameters characterizing the complexity of the texture map to obtain the equivalent volumetric heat capacity of the rack under test; calculating the three-dimensional surface area of ​​the rack under test based on the spatial gradient features of the depth map, and recording the ratio of the three-dimensional surface area to the projected area of ​​the depth map as the tortuosity index of the rack under test; determining the total heat load amplitude of the rack under test by combining the equivalent volumetric heat capacity, the geometric volume of the rack under test, and the initial heat dissipation temperature difference; determining the thermal response time constant characterizing the heat exchange rate based on the tortuosity index; constructing a feedforward cooling power curve based on the total heat load amplitude and the thermal response time constant; and controlling the operation of the precision air conditioning unit in the prefabricated data center based on the feedforward cooling power curve.

[0008] This invention acquires texture and depth maps characterizing the rack under test in the cold aisle of the data center. Based on material analysis, a theoretical heat capacity benchmark is determined, and this benchmark is corrected using gradient parameters reflecting geometric undulations and texture feature parameters reflecting pattern complexity. This allows for the evaluation of the equivalent volumetric heat capacity, which accurately reflects the actual density of the hardware deployment within the rack. By calculating the ratio of the area of ​​a three-dimensional curved surface based on spatial gradient features to the projected area, a tortuosity index is obtained. This invention can accurately assess the effective contact area size for heat exchange between the rack's physical structure and airflow. The total heat load amplitude is determined by combining the equivalent volumetric heat capacity, geometric volume, and initial heat dissipation temperature difference. Based on the tortuosity index, the thermal response time constant is determined, and a feedforward cooling power curve conforming to the equipment's heat dissipation law is constructed. This allows for the instruction of precision air conditioning units to operate on demand before the actual temperature rise in the data center caused by rack load variations. This improves the refined management level and energy efficiency of the data center's digital monitoring and control system.

[0009] Preferably, the acquisition of material information of the texture map includes: inputting the texture map into a pre-trained deep learning classification model to output the classification probability of the device in the rack under test belonging to each material category in the preset material library, and selecting the category with the highest probability as the material information of the device; the gradient parameter is the normalized value of the gradient variance of the depth map; the texture feature is the normalized value of the surface texture entropy of the texture map.

[0010] This invention inputs texture maps into a pre-trained deep learning classification model, outputting the classification probability of the device belonging to each material category in a preset material library, and selecting the category with the highest probability as the material information. Compared with manual input, this improves the automation and accuracy of material recognition. By obtaining the normalized value of the gradient variance of the depth map and the normalized value of the surface texture entropy of the texture map, this invention unifies the features of the geometric depth dimension and the surface pattern dimension to the same scale, reducing the error in physical property fusion calculation caused by differences in data volume, providing standardized input parameters for the subsequent thermal capacity correction model, and ensuring a more objective and reliable comprehensive evaluation of the material and structural characteristics of the rack under test.

[0011] Preferably, the equivalent volumetric heat capacity satisfies the following relationship:

[0012] ;

[0013] in, It is the equivalent volumetric heat capacity of the rack under test; This is the total number of material categories in the preset material library; It is the number of the equipment in the rack under test. The probability of classifying a material; It is the first The heat capacity reference of this material; It is the natural exponential function; It is the porosity sensitivity coefficient; It is the normalized value of the gradient variance of the depth map; It is the normalized value of the surface texture entropy of the texture map; It is a preset micro value.

[0014] This invention employs a specific relational formula to calculate the equivalent volumetric heat capacity. It uses the ratio of the normalized value of the gradient variance of the depth map to the normalized value of the surface texture entropy of the texture map as an exponential decay term to nonlinearly correct the weighted theoretical heat capacity benchmark. This calculation logic reflects the physical relationship between the porosity of the rack hardware deployment and its heat storage capacity. Specifically, it reduces the heat capacity of incompletely filled areas of heat sink fins or slots with high geometric undulations and low texture complexity. This makes the calculated equivalent volumetric heat capacity closer to the true physical state of the rack under test, thereby reducing the waste of cooling capacity caused by overestimating the heat capacity of non-solid structures and improving the physical accuracy of heat load assessment.

[0015] Preferably, the step of calculating the three-dimensional surface area of ​​the rack under test based on the spatial gradient features of the depth map includes: calculating the vector sum of the horizontal gradient magnitude and the vertical gradient magnitude of each pixel in the depth map to obtain the spatial gradient magnitude; taking the square root of the square of the spatial gradient magnitude plus 1 to obtain the surface area micro-element of the corresponding pixel; and accumulating the surface area micro-element of all pixels in the depth map to obtain the three-dimensional surface area of ​​the rack under test.

[0016] This invention calculates the vector sum of the horizontal and vertical gradient magnitudes of each pixel in the depth map to obtain the spatial gradient magnitude, and uses the square root of the square plus one of the spatial gradient magnitude to obtain the surface area micro-element. By accumulating the surface area micro-element of all pixels, the three-dimensional curved surface area of ​​the rack under test is obtained. This invention uses the principle of calculus to evaluate the microscopic geometric extension of the air inlet side surface of the rack, providing accurate data support, which helps to accurately evaluate the convective heat transfer efficiency between airflow and equipment, and provides a reliable geometric basis for determining the thermal response time constant.

[0017] Preferably, obtaining the total heat load amplitude of the rack under test includes: recording the product of the equivalent volumetric heat capacity, the geometric volume of the rack under test, and the initial heat dissipation temperature difference as the total heat load amplitude.

[0018] Preferably, in the feedforward cooling power curve, the first... The feedforward cooling power at any given time satisfies the following relationship:

[0019] ;

[0020] in, It is the first time after the rack under test has been running Feedforward cooling power at any given time; It is the equivalent volumetric heat capacity of the rack under test; It is the geometric volume of the frame under test; It is the initial heat dissipation temperature difference of the rack under test; It is the thermal response time constant; It is the natural exponential function; It is the duration of operation of the rack under test.

[0021] This invention determines the load change monitoring trigger after a specific relationship is established. The feedforward cooling power at any given moment is precisely controlled by adjusting the exponential decay rate of the power using the thermal response time constant. This curve strictly conforms to the physical laws of unsteady-state heat conduction and the monotonically decreasing trend of sensible heat release from IT equipment operation over time, enabling the cooling output of the precision air conditioning unit to achieve dynamic synchronization with the actual heat release rate of the rack under test throughout the entire time domain. This not only overcomes the inherent time lag of traditional feedback control and avoids room temperature overshoot due to insufficient cooling supply in the initial stage of load changes, but also avoids energy waste caused by excessive cooling in the later stages of cooling, achieving a high-precision, low-energy-consumption constant temperature control process.

[0022] Preferably, obtaining the geometric volume of the rack under test includes: extracting foreground pixels from the depth map based on a preset depth threshold; obtaining the intrinsic parameter matrix of the image acquisition device used to acquire the depth map; back-projecting the foreground pixels in the depth map into a three-dimensional spatial point cloud based on the depth map and the intrinsic parameter matrix; and performing volume integration or bounding box calculation on the three-dimensional spatial point cloud to obtain the geometric volume of the rack under test.

[0023] Preferably, the acquisition of the texture map and depth map of the rack under test in the prefabricated data center includes: deploying an image acquisition device in the cold aisle of the computer room to acquire an original color image and an original depth image containing the rack under test and the background; performing background subtraction processing on the original depth image to obtain a foreground region; and using the foreground region as a mask to apply to the original color image and the original depth image respectively to obtain the texture map and depth map of the rack under test.

[0024] Preferably, the precision air conditioning unit includes a fixed-speed compressor and a variable-frequency compressor connected in parallel; the control of the operation of the precision air conditioning unit of the prefabricated data center based on the feedforward cooling power curve includes: converting the value of the feedforward cooling power curve into the target operating frequency of the precision air conditioning unit; when the target operating frequency exceeds a set threshold, starting the fixed-speed compressor and the variable-frequency compressor to operate together.

[0025] In a second aspect, the present invention provides an AI-based prefabricated data center cooling control system, which includes a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the AI-based prefabricated data center cooling control method of the first aspect of the present invention is implemented.

[0026] By adopting the above technical solution, a computer program is generated from the AI-based prefabricated data center cooling control method of the first aspect of the present invention, and stored in a memory for loading and execution by a processor, thereby creating a terminal device based on the memory and processor for convenient use.

[0027] The beneficial effects of this invention are as follows: This invention integrates the texture and depth geometric features of the rack under test to invert the equivalent volumetric heat capacity and tortuosity index, establishing a mapping relationship between visual appearance and physical thermal properties. This allows the system to distinguish between high-density compact racks and low-density non-solid structure racks without contacting the equipment, correcting the theoretical heat capacity benchmark, reducing the heat load estimation deviation caused by ignoring the heterogeneity of the internal space of the equipment in traditional methods, and improving the adaptability of the temperature control system to IT equipment with different load characteristics. This invention constructs a time-varying heat conduction model based on physical parameters such as the total heat load amplitude and the thermal response time constant, generating a feedforward cooling power curve. By issuing control commands the instant the load change is sensed, the heat release process is predicted, reducing the lag in temperature control, ensuring that the cooling output is synchronized with the rack heat load release, and maintaining the stability of the temperature environment in the data center. This invention controls the operating frequency and combination mode of precision air conditioning units based on the calculated feedforward cooling power curve. This achieves an accurate match between cooling supply and cooling demand, preventing temperature fluctuations due to insufficient cooling for high-density computing nodes, and reducing the overall energy consumption of data center operations while ensuring the safe operation of computing resources. Attached Figure Description

[0028] Figure 1 A flowchart illustrating an AI-based prefabricated data center cooling control method provided in an embodiment of the present invention;

[0029] Figure 2 This is a feature space mapping diagram of high-density computing nodes and low-density storage nodes in a data center scenario provided by an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram illustrating the comparative analysis of the volumetric thermal capacity of high-density computing nodes and low-density storage nodes provided in an embodiment of the present invention.

[0031] Figure 4 The following are the feedforward cooling power dynamic response curves for two different rack types, high-density computing nodes and low-density storage nodes, provided in embodiments of the present invention.

[0032] Figure 5 This is a structural block diagram of an AI-based prefabricated data center cooling and control system provided in an embodiment of the present invention. Detailed Implementation

[0033] The first aspect of this invention provides an AI-based method for cooling control in prefabricated data centers, such as... Figure 1 As shown, the method includes steps S100-S400:

[0034] Step S100: Obtain the texture map and depth map of the rack to be tested in the prefabricated data center.

[0035] It should be noted that in the actual monitoring scenario of a prefabricated data center, the controlled environment typically includes background interference such as floors, cable trays, maintenance aisles, and mobile maintenance tools. If the full-frame image is analyzed directly, this non-cooling-related background noise will significantly interfere with subsequent calculations of rack stacking density and heat exchange channel geometry. This invention is used to accurately extract pixel data of the rack or IT equipment under test from the complex background of the data center, providing clean data input for subsequent thermal response time constant and heat capacity inversion. Considering the relatively constant lighting environment within the data center, environmental interference mainly manifests as spatial obstruction by non-target objects and background redundancy.

[0036] First, an image acquisition device, preferably an RGB-D camera, is deployed above the cold aisle of the prefabricated data center module, with the camera lens tilted at a preset angle towards the air intake side of the rack under test. Under the illumination of LED lights on the server room ceiling, color data and distance data corresponding to each pixel are simultaneously acquired. When the energy-saving control platform senses changes in equipment load or receives an inspection command, it uses the image acquisition device to capture color texture images and depth topology images. The color texture image records the color and ventilation hole pattern information of the rack panel and IT equipment surface, while the depth topology image records the physical distance from each point on the equipment surface to the image acquisition device.

[0037] Secondly, a depth map pre-collected in the rack's undeployed state is used as the background image. Pixel-level difference operations are performed between the depth topology image and the background image to obtain a depth difference map. A safe displacement threshold is set, for example, a vertical height difference of 10 cm relative to the background plane. Areas in the depth difference map whose height change relative to the background plane exceeds this threshold are retained and marked as rack foreground.

[0038] Then, considering the physical positional deviation between the color lens and depth lens of the RGB-D camera, the color texture image and the depth topology image are spatially registered using the intrinsic and extrinsic parameter matrices of the image acquisition device, ensuring that the same coordinate point in the two images corresponds to the same point in the real world. Finally, the foreground of the gantry is used as a mask to cover the registered color texture image and depth topology image respectively, retaining only the pixel values ​​within the mask area and filtering out background noise outside the mask to obtain the texture map and depth map.

[0039] At this point, we have obtained the texture map and depth map with background interference removed.

[0040] Step S200: Analyze the material information of the texture map to determine the theoretical thermal capacity reference of the equipment in the rack under test. Combine the gradient parameters that characterize the geometric undulations of the depth map and the texture feature parameters that characterize the complexity of the texture map to correct the theoretical thermal capacity reference and obtain the equivalent volumetric thermal capacity of the rack under test.

[0041] It should be noted that in the actual monitoring scenarios of prefabricated data centers, visual texture alone is insufficient to accurately determine the true thermal inertia of IT equipment combinations, leading to interference from mismatches between visual features and physical properties. For example, server panels with complex labels or dense arrays of ventilation holes may have high surface texture entropy, but their internal structure may be relatively flat and compact. Conversely, areas with numerous complex-shaped heat sinks or partially filled rack slots may have a relatively regular visual texture, but contain a large number of air layers. If heat capacity is determined solely based on texture identification, racks with numerous airflow gaps may be misjudged as high-density solid heat sources, resulting in an overestimation of the calculated cooling load feedforward. Therefore, this invention introduces spatial geometric features in the depth dimension, evaluating the physical undulations of the equipment surface to correct conclusions based on texture judgment, thereby deriving an equivalent volumetric heat capacity that more closely approximates the actual physical state of the data center.

[0042] Specifically, firstly, a material recognition model for equipment is constructed and trained to obtain benchmark data for heat capacity calculation. This involves collecting historical equipment images accumulated during data center inspections and constructing a labeled dataset containing various material categories of data center components, such as cold-rolled steel casings, aluminum alloy heat sinks, engineering plastic panels, and copper heat pipes. Mainstream convolutional neural network architectures, such as ResNet or VGG, are selected as the backbone network. Supervised training is then conducted using the labeled dataset. Through iterative optimization, the network learns the texture features and visual patterns of different data center hardware surfaces until the model converges, resulting in a deep learning classification network with material recognition capabilities. How to train a deep learning classification network is an existing technology and will not be elaborated upon here.

[0043] Secondly, the texture map is input into the deep learning classification network for inference, outputting the classification probability of each material category of the IT equipment in the rack under test. The category with the highest probability is selected as the material information of the equipment, and the corresponding theoretical heat capacity benchmark is retrieved from a pre-built material library that records standard physical properties. It should be noted that the material library is a database pre-built based on the historical asset registration records of the data center and the thermophysical property manual of general industrial materials. The library stores a variety of common hardware materials in data centers and records the standard physical property values ​​corresponding to each material category, including theoretical density and specific heat capacity.

[0044] Then, the gradient variance and surface texture entropy of the test frame surface are calculated. Considering the difference in their dimensions, the min-max normalization method is used to map them to... The interval is used to obtain the normalized values ​​of the gradient variance and the surface texture entropy.

[0045] Finally, the equivalent volumetric heat capacity of the rack under test is calculated. It should be noted that, to construct an accurate heat capacity inversion model, this invention establishes a logical assumption based on structural correction: the true heat capacity of an IT equipment assembly should be its theoretical baseline heat capacity multiplied by a correction coefficient reflecting the density of equipment deployment and heat dissipation characteristics. This correction coefficient should be negatively correlated with the spatial geometric unevenness of the equipment surface. Considering that although structures such as heat sink fins increase the geometric gradient, they occupy a relatively small proportion of the solid volume, this invention constructs a mathematical model with the theoretical heat capacity as the base and the ratio of depth features to texture features as the exponential decay term. The monotonically decreasing characteristic of the exponential function is used to reduce the physical heat capacity of non-solid hardware structures.

[0046] Based on the above logic, the equivalent volumetric heat capacity satisfies the following relationship:

[0047] ;

[0048] in, It is the equivalent volumetric heat capacity of the rack under test; This is the total number of material categories in the preset material library; It is the number of the equipment in the rack under test. The probability of classifying a material; It is the first The heat capacity reference of this material; It is the natural exponential function; It is the porosity sensitivity coefficient; It is the normalized value of the gradient variance of the depth map; It is the normalized value of the surface texture entropy of the texture map; It is a preset microvalue used to prevent It should be 0, or it can be set to 0.001.

[0049] In this relation, This is the basic heat capacity term, which uses the classification confidence level to perform a weighted average of the theoretical heat capacity of different materials in order to eliminate the risk of misjudgment that may exist in a single classification. This is the structural correction term, the fractional part of which is used to evaluate the mapping weight of the spatial heterogeneity of IT equipment relative to visual features. If the equipment is a storage array with a flat surface and complex labels, although the normalized value of the texture entropy is high, the normalized value of the gradient variance approaches zero, causing the exponent to approach zero, and this term approaches 1, that is, the theoretical heat capacity baseline is preserved. Conversely, if the equipment has a complex heat sink structure or empty deployment area in the rack slots, the normalized value of the gradient variance approaches 1, causing the exponent to become a large negative number, and the structural correction term decays rapidly, thereby physically reducing the heat capacity.

[0050] It should be noted that the porosity sensitivity coefficient should be set according to the equipment density of the data center. For data centers mainly equipped with high-density blade servers, the coefficient should be set to a smaller value, such as 0.5, to maintain a higher thermal inertia baseline. For data centers with network switching equipment or low slot fill rate, the coefficient should be set to a larger value, such as 2, to enhance the sensitivity to the reduction of internal gaps. In this embodiment, it is preferred to set it to 1.2.

[0051] like Figure 2 The figure shows the feature space mapping of high-density computing nodes and low-density storage nodes in a data center scenario. The horizontal axis represents the normalized value of surface texture entropy, and the vertical axis represents the normalized value of gradient variance. The data distribution in the figure shows that the two types of devices exhibit distinct regional characteristics in the feature space: high-density computing nodes are concentrated in regions with high texture entropy and low gradient variance, and their corresponding structure correction term approaches 1, indicating that they are judged to be approximately solid and compact structures; while low-density storage nodes are concentrated in regions with low texture entropy and high gradient variance, and their corresponding structure correction term is significantly reduced, indicating that a large number of air layers exist inside due to incomplete filling.

[0052] like Figure 3 The diagram illustrates a comparative analysis of the volumetric heat capacity of high-density computing nodes and low-density storage nodes. One set of data represents the theoretical heat capacity baseline obtained solely based on material identification, while the other set represents the equivalent volumetric heat capacity after correction using the algorithm of this invention. The comparison results show that for high-density computing nodes, the heat capacity values ​​before and after correction are highly consistent, indicating that the cooling response weight for high-density computing nodes is correctly maintained. However, for low-density storage nodes, the corrected equivalent volumetric heat capacity is significantly lower than the theoretical heat capacity baseline. This numerical difference demonstrates that this invention can effectively eliminate false heat loads caused by rack gaps and complex heat dissipation structures, avoiding excessive cooling output due to simple estimation based on geometric volume.

[0053] Thus, the equivalent volumetric heat capacity, which reflects the true physical state inside the IT equipment, has been obtained.

[0054] Step S300: Calculate the three-dimensional surface area of ​​the rack under test based on the spatial gradient features of the depth map, and record the ratio of the three-dimensional surface area to the projected area of ​​the depth map as the tortuosity index of the rack under test.

[0055] It should be noted that for a test rack of the same physical volume, the more complex its surface physical structure and the denser the distribution of heat dissipation pores, the larger its unfolded true three-dimensional surface area, the higher its airflow exchange efficiency, and the faster its heat dissipation rate. To accurately evaluate this physical characteristic, this invention uses the concept of surface area integration from calculus to construct an evaluation index. The core of this idea is to treat discrete depth data as a continuous geometric surface, and approximate the real surface area of ​​the object by integrating the area of ​​small local tangent planes. Its advantage lies in its ability to keenly capture continuous surface undulations and obtain area ratios with clear physical meaning, rather than merely statistically analyzing the degree of numerical dispersion. This method used in this invention can more accurately reflect the size of the effective heat exchange interface between the cooling airflow and the rack hardware.

[0056] Specifically, for each pixel in the depth map, the gradient magnitude in the horizontal and vertical directions is calculated using the Sobel or Prewitt operators respectively, and the vector sum of the gradient magnitudes in the two directions is taken as the spatial gradient magnitude of the pixel.

[0057] Secondly, the surface area element is constructed based on the principle of differential geometry: It should be noted that since the gradient magnitude represents the degree of inclination of the pixel relative to the projection plane, the greater the inclination, the greater the increase in the actual surface area relative to the unit projected area. Therefore, the surface area element of each pixel can be solved by the functional relationship of the gradient.

[0058] Specifically, the spatial gradient magnitude of each pixel is used to calculate the area of ​​the tiny tangent plane corresponding to that pixel in three-dimensional space.

[0059] Then, the tortuosity index of the rack under test is calculated based on the surface area elements of all pixels and the total projected area of ​​the rack. It should be noted that this ratio is essentially an area expansion factor, used to measure the degree of extension of an object's surface relative to its two-dimensional projection. A ratio closer to 1 indicates a smoother, flatter surface; a larger ratio indicates that the surface has more folded or rolled-up physical surfaces within a unit projection area, i.e., a more complex and rugged surface structure. This aligns with the concept of surface tortuosity in topology. Therefore, this ratio can accurately characterize the topological tortuosity of the physical structure on the air inlet side of the rack under test.

[0060] Based on the above logic, the tortuosity index of the rack under test satisfies the following relationship:

[0061] ;

[0062] in, It is the tortuosity index of the frame under test; It represents the total number of pixels in the depth map; It is the first in the depth map The gradient magnitude of each pixel.

[0063] In this formula, the numerator represents the actual total unfolded area of ​​the air inlet side of the rack under test, and the denominator represents... This represents the total projected area of ​​the rack under test, and the entire formula is the ratio of the two. (Square Root term) The surface area element is a infinitesimal element, and its geometric principle originates from the generalization of the Pythagorean theorem to three-dimensional surfaces: that is, assuming that the unit area of ​​a single pixel on the projection plane is 1, while... This represents the rate of change or slope of that location along the depth direction. According to the Pythagorean theorem, the area of ​​the inclined tangent plane corresponding to that pixel is equal to... When the surface of the test frame is flat, the spatial gradient modulus approaches 0, the surface area of ​​the infinitesimal element approaches 1, and the total value after accumulating the numerators is approximately equal to the denominator. The calculation result at this time A value close to 1 indicates a small heat dissipation surface area. When the surface of the test frame has a large number of heat dissipation fins or heat dissipation pore structures, the spatial gradient modulus is large, the surface area micro-element is significantly greater than 1, and the total value after accumulating the numerators is significantly greater than the denominator. The calculation result at this time A value significantly greater than 1 indicates the presence of numerous lateral surfaces and pores for airflow and heat exchange.

[0064] It should be further explained that the tortuosity exponent serves as a time constant correction factor in the thermal response model. In practical applications, considering that depth sensors are prone to flying point noise at object edges, causing the calculated local gradient magnitude to tend towards infinity, the final calculated exponent is artificially inflated. To eliminate such noise interference, this invention introduces a gradient magnitude limiting mechanism. Before performing area element integration, a physically reasonable threshold is set for the spatial gradient magnitude, filtering out abnormal gradient values ​​that exceed the normal geometric deformation range, thus ensuring the numerical stability of the tortuosity exponent.

[0065] Specifically, the spatial gradient modulus is numerically equivalent to the tangent of the tilt angle of the object's surface relative to the imaging plane. Considering that in actual rack deployment scenarios, whether it is the physical slope of the heat sink fins or the geometric contour of the rack panel, the effective observation tilt angle in a single-view depth map usually has a physical upper limit, such as 75 degrees. Depth variations exceeding this angle are usually caused by edge occlusion or measurement flypoints. Therefore, the tangent value corresponding to this upper limit of tilt angle can be used as a reasonable threshold. In this embodiment, this threshold is preferably set to 3, corresponding to a tilt angle of approximately 71.6 degrees, which can both retain most of the real surface steepness features and effectively eliminate non-physical high-frequency noise caused by depth discontinuities.

[0066] Thus, the tortuosity index, which characterizes the heat dissipation capability of the rack under test, was obtained.

[0067] Step S400: Combine the equivalent volumetric heat capacity, the geometric volume of the rack under test, and the initial heat dissipation temperature difference to determine the total heat load amplitude of the rack under test. Based on the tortuosity index, determine the thermal response time constant characterizing the heat exchange rate. Based on the total heat load amplitude and the thermal response time constant, construct a feedforward cooling power curve. Based on the feedforward cooling power curve, control the operation of the precision air conditioning unit of the prefabricated data center.

[0068] It should be noted that, in order to achieve zero-hysteresis constant temperature control, this invention uses the physical parameters obtained in the aforementioned steps to construct a time-varying heat conduction model, calculates in advance the heat release power of the test rack as it changes over time after load operation, and performs feedforward control accordingly.

[0069] Specifically, the basic variables required for the calculation are first determined: based on a preset depth distance threshold, the depth map is segmented, and pixels smaller than the depth threshold are extracted as foreground pixels; based on the depth map and the intrinsic parameter matrix of the image acquisition device, the foreground pixels in the depth map are back-projected into a three-dimensional spatial point cloud, and volume integration or bounding box calculation is performed on the three-dimensional spatial point cloud to obtain the geometric volume of the rack under test. Simultaneously, the current surface temperature of the rack under test and the set cooling temperature of the server room are obtained, and the initial heat dissipation temperature difference is calculated. Secondly, based on basic thermodynamic principles, the equivalent volumetric heat capacity, the geometric volume of the rack under test, and the initial heat dissipation temperature difference are multiplied to calculate the total enthalpy change required for complete cooling of the rack under test, i.e., the amount of heat energy to be removed.

[0070] In one feasible implementation, to adapt to detection needs at different installation heights or in complex environments, the depth threshold is not a fixed value, but rather adaptively obtained through statistical analysis of the current frame depth map. The specific steps are as follows: The depth value distribution of all valid pixels in the depth map is statistically analyzed to construct a depth histogram. The depth histogram exhibits a bimodal distribution, with one peak corresponding to the background (e.g., the depth distribution of the ground or wall) and the other peak corresponding to the foreground (e.g., the depth distribution of the device under test). The optimal segmentation threshold between the two peaks is calculated using the maximum inter-class variance method. This algorithm iterates through possible thresholds, searching for the value that maximizes the inter-class variance between the foreground and background pixel classes, and determines this value as the current depth threshold. Implementers can also set this threshold according to their needs.

[0071] Then, the feedforward cooling power curve is calculated. It should be noted that, to construct a mathematical model that accurately matches the heat dissipation characteristics of IT equipment, this invention uses Newton's law of cooling and the principle of unsteady-state heat conduction for mechanism analysis and model construction: First, the total cooling energy of the rack under test is determined by its volume, temperature difference, and equivalent volumetric heat capacity; second, the energy release is not instantaneous but exhibits a dynamic decay process influenced by surface area. The larger the surface area, i.e., the higher the tortuosity index, the higher the heat exchange efficiency of the airflow, and the shorter the characteristic time of heat release. Based on this, this invention constructs an exponential response model with the total energy divided by the characteristic time as the amplitude and the ratio of time to the characteristic time as the independent variable, to objectively map the dynamic physical process of heat load gradually penetrating from the moment of operation and affecting the environment.

[0072] Based on the above logic, the first... The feedforward cooling power at any given time satisfies the following relationship:

[0073] ;

[0074] in, The first after the load fluctuation of the rack under test Feedforward cooling power at any given time; It is the equivalent volumetric heat capacity of the rack under test; It is the geometric volume of the frame under test; It is the initial heat dissipation temperature difference of the rack under test; It is the thermal response time constant, which is obtained by dividing the basic time constant factor by the tortuosity index of the frame under test; It is the natural exponential function; It is the duration after load change monitoring is triggered.

[0075] In this relation, As the amplitude coefficient, it defines the peak cooling power required for the instantaneous change in the thermal load of the rack under test. The denominator represents the total enthalpy that needs to be removed to cool the test rack from its current state to the set temperature; This reflects the rate of heat release, because The thermal response time constant is inversely proportional to the tortuosity index of the test frame. The rougher and looser the surface of the test frame, i.e., the larger the tortuosity index, the longer the thermal response time constant. The smaller the value, the faster the heat exchange rate, the more heat is released per unit time, and therefore the greater the instantaneous cooling power required. Describe heat load over time The dynamic response process exhibits monotonically decaying behavior, particularly during the initial stages of a load surge. When the value is low, this value approaches 1, indicating that the maximum cooling power is output the moment a thermal risk is detected to effectively suppress the initial thermal shock caused by the maximum temperature difference. As time goes on, the temperature difference between the test rack and the environment decreases, and this value decreases exponentially. The commanded cooling power then smoothly drops back down, thereby achieving real-time synchronization between the cooling output and the actual heat dissipation rate of the equipment.

[0076] It should be noted that the basic time constant factor This is a preset physical quantity characterizing the basic heat exchange capacity of a precision air conditioning refrigeration system, representing the thermal response inertia of a standard flat-surface cabinet under the current wind speed. Its value needs to be set in conjunction with the airflow circulation capacity of the computer room: if the computer room fans have a large air volume and fast circulation... It can be set to a smaller value, such as 600 seconds, allowing the system to output more aggressive power commands; if the wind circulation is weak, It should be set to a relatively large value, such as 1200 seconds, to avoid localized overcooling. In this embodiment, The preferred setting is 900 seconds.

[0077] In actual control processes, time variables are continuously updated using a time step, such as 1 second. A series of continuous calculations were performed by iterating through the feedforward cooling power relationship. The value is used to generate a complete feedforward cooling power curve. Based on the current value of this curve, the collaborative control logic of the precision air conditioning unit is as follows:

[0078] The calculated instantaneous power value The controller converts the target operating frequency of the precision air conditioning unit in real time. When the calculated target frequency increment exceeds the adjustment range of a single variable frequency precision air conditioner, the system automatically commands the standby fixed speed precision air conditioner to start and uses the variable frequency precision air conditioner to undertake the remaining non-integer multiple cooling capacity adjustment task.

[0079] like Figure 4 The figure shows the dynamic response curves of feedforward cooling power for two different rack types: high-density computing nodes and low-density storage nodes. The horizontal axis represents the duration after load change monitoring is triggered, and the vertical axis represents the feedforward cooling power. The power curve of the high-density computing node shows a clear attenuation trend and a high amplitude, indicating that the system outputs strong cooling power to suppress thermal shock at the moment of load initiation, and then automatically reduces it as the heat load is released. In contrast, the power curve of the corresponding low-density storage node responds quickly but has a lower amplitude, indicating that it recognizes that the actual thermal shock of this type of rack is small, thus automatically maintaining low-power operation.

[0080] The second aspect of this embodiment provides an AI-based prefabricated data center cooling and control system, such as... Figure 5As shown, the AI-based prefabricated data center cooling control system includes a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an AI-based prefabricated data center cooling control method according to the first aspect of the present invention is implemented.

[0081] The AI-based prefabricated data center cooling and control system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their setup and functions are known in the art and will not be described in detail here.

[0082] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (DRAM), high-bandwidth memory, hybrid memory cube, etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

[0083] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A prefabricated data center cooling control method based on AI, characterized in that, include: Obtain the texture and depth maps of the rack under test within the prefabricated data center; The material information of the texture map is analyzed to determine the theoretical heat capacity benchmark of the equipment in the rack under test. The theoretical heat capacity benchmark is corrected by combining the gradient parameters that characterize the geometric undulations of the depth map and the texture feature parameters that characterize the complexity of the texture map, so as to obtain the equivalent volumetric heat capacity of the rack under test. The three-dimensional surface area of ​​the rack under test is calculated based on the spatial gradient features of the depth map, and the ratio of the three-dimensional surface area to the projected area of ​​the depth map is recorded as the tortuosity index of the rack under test. The total heat load amplitude of the rack under test is determined by combining the equivalent volume heat capacity, the geometric volume of the rack under test, and the initial heat dissipation temperature difference. The thermal response time constant, which characterizes the heat exchange rate, is determined based on the tortuosity index. Based on the total heat load amplitude and the thermal response time constant, a feedforward cooling power curve is constructed. The operation of the precision air conditioning unit of the prefabricated data center is controlled based on the feedforward cooling power curve. The equivalent volumetric heat capacity satisfies the following relationship: ; It is the equivalent volumetric heat capacity of the rack under test; This represents the total number of material categories in the preset material library; It is the number of the equipment in the rack under test. The probability of classifying a type of material; It is the first The heat capacity reference of this material; It is a natural exponential function; It is the porosity sensitivity coefficient; It is the normalized value of the gradient variance of the depth map; It is the normalized value of the surface texture entropy of the texture map; It is a preset micro value.

2. The AI-based prefabricated data center cooling control method according to claim 1, characterized in that, The acquisition of the material information of the texture map includes: The texture map is input into a pre-trained deep learning classification model to output the classification probability of the device in the rack under test belonging to each material category in the preset material library. The category with the highest probability is selected as the material information of the device. The gradient parameter is the normalized value of the gradient variance of the depth map; Texture features are the normalized values ​​of the surface texture entropy of the texture map.

3. The AI-based prefabricated data center cooling control method according to claim 1, characterized in that, The calculation of the three-dimensional surface area of ​​the rack under test based on the spatial gradient features of the depth map includes: Calculate the vector sum of the horizontal and vertical gradient magnitudes of each pixel in the depth map to obtain the spatial gradient magnitude; The surface area element of the corresponding pixel is obtained by taking the square root of the square of the spatial gradient magnitude plus 1. The surface area of ​​all pixels in the depth map is accumulated to obtain the three-dimensional surface area of ​​the frame under test.

4. The AI-based prefabricated data center cooling control method according to claim 1, characterized in that, The acquisition of the total heat load amplitude of the rack under test includes: recording the product of the equivalent volumetric heat capacity, the geometric volume of the rack under test, and the initial heat dissipation temperature difference as the total heat load amplitude.

5. The AI-based prefabricated data center cooling control method according to claim 1 or 4, characterized in that, The feedforward cooling power curve in the first The feedforward cooling power at any given time satisfies the following relationship: ； in, It is the first time after the rack under test has been running Feedforward cooling power at any given time; It is the equivalent volumetric heat capacity of the rack under test; It is the geometric volume of the frame under test; It is the initial heat dissipation temperature difference of the rack under test; It is the thermal response time constant; It is a natural exponential function; It is the time variable corresponding to the feedforward cooling power curve.

6. The AI-based prefabricated data center cooling control method according to claim 1, characterized in that, Obtaining the geometric volume of the test frame includes: Extract foreground pixels from the depth map based on a preset depth threshold; Obtain the intrinsic parameter matrix of the image acquisition device used to acquire depth maps; Based on the depth map and intrinsic parameter matrix, the foreground pixels in the depth map are back-projected into a three-dimensional spatial point cloud. The volume integral or bounding box calculation is performed on the three-dimensional spatial point cloud to obtain the geometric volume of the rack under test.

7. The AI-based prefabricated data center cooling control method according to claim 1, characterized in that, The acquisition of texture and depth maps of the rack under test within the prefabricated data center includes: Image acquisition equipment is deployed in the cold aisle of the computer room to acquire raw color images and raw depth images containing the rack under test and the background. Background subtraction is performed on the original depth image to obtain the foreground region; The foreground region is used as a mask and applied to the original color image and the original depth image respectively to obtain the texture map and depth map of the gantry under test.

8. The AI-based prefabricated data center cooling control method according to claim 1, characterized in that, The precision air conditioning unit includes a fixed-speed compressor and a variable-frequency compressor connected in parallel; the operation of the precision air conditioning unit in the prefabricated data center is controlled based on the feedforward cooling power curve, including: converting the value of the feedforward cooling power curve into the target operating frequency of the precision air conditioning unit; when the target operating frequency exceeds a set threshold, starting the fixed-speed compressor and the variable-frequency compressor to operate together.

9. A prefabricated data center cooling and control system based on AI, characterized in that, The AI-based prefabricated data center cooling control system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an AI-based prefabricated data center cooling control method according to any one of claims 1-8.

Images