Fan area flow field prediction method and electronic device

Abstract

The application discloses a fan area flow field prediction method and electronic equipment, and relates to the technical field of servers. Embodiments of the application simplify fan area flow field simulation of an air-cooled server into a prediction problem of an inlet boundary condition, and construct a fast prediction model through an AI algorithm, so that the fan prediction flow field can be efficiently predicted. Through a double-model intelligent scheduling strategy based on a rotating speed interval, high-precision prediction of the fan outlet flow field can be realized in the full operating condition range. A nonlinear fitting model is used for low-speed steady flow to ensure calculation efficiency, and a space-time modeling model is used for high-speed unsteady flow to accurately capture dynamic characteristics. The method not only overcomes the defects of insufficient prediction accuracy of traditional single models under different operating conditions, but also avoids huge calculation overhead caused by complex numerical simulation under high rotating speed operating conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of servers, and particularly relates to a fan area flow field prediction method and an electronic device. BACKGROUND

[0002] Air-cooled servers are the core heat dissipation solutions for data centers and high-performance computing clusters. The flow field characteristics generated by the internal fans directly determine the heat dissipation efficiency and device reliability. The physical quantity distribution of the fan outlet section as a key boundary condition has a decisive influence on the overall thermal management design.

[0003] In the related art, the computational fluid dynamics method is usually used to simulate the fan flow field, but there are problems of low calculation efficiency and insufficient prediction accuracy. SUMMARY

[0004] The present application provides a fan area flow field prediction method and an electronic device to at least solve the problem of low prediction efficiency and low accuracy in the related art.

[0005] The present application provides a fan area flow field prediction method, comprising:

[0006] Obtaining input data corresponding to a fan in a server to be predicted; the input data comprising working condition parameters and geometric parameters of the fan; the working condition parameters at least comprising a fan speed;

[0007] If the fan speed belongs to a first speed interval, inputting the input data into a first model to obtain a first prediction result of the physical quantity distribution of the fan outlet section of the fan; the maximum fan speed of the first speed interval being less than a first preset threshold;

[0008] If the fan speed belongs to a second speed interval, inputting the input data and time characteristics into a second model to obtain a second prediction result of the physical quantity distribution; the minimum fan speed of the second speed interval being greater than a second preset threshold; the second preset threshold being greater than the first preset threshold; the physical quantity comprising at least one of the following: velocity, pressure, and temperature.

[0009] The present application also provides a fan area flow field prediction device, comprising:

[0010] An obtaining module is configured to obtain input data corresponding to a fan in a server to be predicted; the input data comprising working condition parameters and geometric parameters of the fan; the working condition parameters at least comprising a fan speed;

[0011] An inputting module is configured to, if the fan speed belongs to a first speed interval, input the input data into a first model to obtain a first prediction result of the physical quantity distribution of the fan outlet section of the fan; the maximum fan speed of the first speed interval being less than a first preset threshold;

[0012] The input module is further configured to input the input data and the time feature into a second model to obtain a second prediction result of the physical quantity distribution if the fan rotating speed belongs to a second rotating speed interval, wherein a minimum fan rotating speed of the second rotating speed interval is greater than a second preset threshold, and the second preset threshold is greater than the first preset threshold; and the physical quantity includes at least one of the following: speed, pressure, and temperature.

[0013] The application further provides an electronic device, comprising a memory configured to store a computer program, and a processor configured to execute the computer program to implement the steps of any of the fan region flow field prediction methods.

[0014] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of any of the fan region flow field prediction methods.

[0015] The application further provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the steps of any of the fan region flow field prediction methods.

[0016] By the application, the fan region flow field simulation of the air-cooled server is simplified into a prediction problem of an inlet boundary condition, and a rapid prediction model is constructed through an artificial intelligence algorithm, so that the fan prediction flow field can be efficiently predicted, and through the double-model intelligent scheduling strategy based on the rotating speed interval, high-precision prediction of the fan outlet flow field can be realized in the full operating condition range, the nonlinear fitting model is used for low-speed steady flow to ensure the calculation efficiency, and the space-time modeling model is used for high-speed unsteady flow to accurately capture the dynamic characteristics, so that the defects of the traditional single model in the prediction accuracy under different operating conditions are overcome, and the huge calculation overhead caused by the complex numerical simulation under the high-speed operating condition is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the application, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 The application further provides a fan region flow field prediction method provided by the application, and a schematic diagram of an application scenario of the fan region flow field prediction method is shown in FIG. 1.

[0019] Figure 2 The application further provides a fan region flow field prediction method provided by the application, and a schematic diagram of a flow process of the fan region flow field prediction method is shown in FIG. 2.

[0020] Figure 3 The application further provides a fan region flow field prediction method provided by the application, and a schematic diagram of a flow process of the training method of the first model and the second model is shown in FIG. 3.

[0021] Figure 4 A flowchart of a first model training method provided for an embodiment of the present application is shown in FIG. 1A.

[0022] Figure 5 An architecture diagram of the first model provided for an embodiment of the present application is shown in FIG. 1B.

[0023] Figure 6 A flowchart of a second model training method provided for an embodiment of the present application is shown in FIG. 2A.

[0024] Figure 7 An architecture diagram of the second model provided for an embodiment of the present application is shown in FIG. 2B.

[0025] Figure 8 A structure diagram of a fan area flow field prediction device provided for an embodiment of the present application is shown in FIG. 3.

[0026] Figure 9 A structure diagram of an electronic device provided for the present application is shown in FIG. 4. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, any other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0028] It should be noted that, in the description of the present application, the terms “comprise”, “contain” or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. The terms “first”, “second” and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence.

[0029] Air-cooled servers are widely used in high-density electronic device cooling scenarios such as data centers and high-performance computing clusters. The prediction of the internal flow field is a core link of the cooling design, which runs through the whole life cycle from the conceptual design to the product landing. The electronic components (such as central processing units (CPUs), graphics processing units (GPUs), memory modules, etc.) inside the server will generate a large amount of heat during operation, which needs to be quickly discharged through the airflow driven by the fan to maintain the stable operation of the device.

[0030] The flow field characteristics of the fan region directly affect the distribution of internal airflow, the uniformity of the temperature field, and the heat dissipation efficiency of the server. The flow field state (such as velocity, pressure, and temperature distribution) of the fan outlet section is the inlet boundary condition of the internal flow field simulation of the server, and the prediction accuracy and efficiency thereof have a decisive role in the reliability and cost control of the overall heat dissipation design.

[0031] In the related art, numerical simulation of the internal flow field of the air-cooled server is usually performed by relying on commercial simulation software or open-source software. Specifically, for the simulation of the fan region, when the fan speed is low and the flow is in a laminar state, the rotation effect of the fan is usually simulated by a multi-reference frame or a sliding mesh method, and a source term is introduced into the Navier-Stokes equation to represent the driving action of the fan. When the fan speed is high and the flow enters a turbulent state, a turbulence model or large eddy simulation method needs to be used, but these methods have extremely high demands on computing resources and are difficult to handle the complexity of unsteady flow.

[0032] In another way, to reduce the computing cost, some designers simplify the fan region as a duct, and replace the detailed simulation of the fan with empirical formulas or simplified boundary conditions (such as setting the inlet velocity or pressure). However, when the fan speed is high and the flow is unsteady, the prediction results of such methods deviate significantly, and cannot meet the high-precision design requirements.

[0033] However, in the above-mentioned ways, the solution of the turbulent flow simulation and the unsteady flow requires extremely high computing resources, resulting in long simulation period and high hardware cost. The simplified model cannot accurately capture the flow field characteristics of the fan outlet section under complex working conditions (such as high speed and unsteady flow), leading to distortion of the prediction of the internal temperature field of the server. Moreover, it is difficult to quickly adapt to the prediction requirements of different fan geometric parameters (such as blade shape and number) or speed changes. Therefore, how to provide a high-efficiency and high-precision fan region flow field prediction method to improve the efficiency and reliability of the air-cooled server heat dissipation design is a technical problem to be solved at present.

[0034] To solve the above technical problems, the present inventors have found that the high-efficiency and high-precision prediction of the fan region flow field can be realized by introducing an AI algorithm (such as a neural network model or a diffusion model), thereby improving the efficiency and reliability of the air-cooled server heat dissipation design.

[0035] In order for those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0036] In combination with the specific application environment architecture or specific hardware architecture on which the execution of the fan region flow field prediction method depends, the specific application environment architecture or specific hardware architecture is described herein. For reference Figure 1 ,Figure 1 An application scenario diagram of the fan area flow field prediction method provided by the present application is shown. As shown in Figure 1 When designing an air-cooled server, the entire flow field of the air-cooled server can be divided into a fan area flow field and a server internal flow field, with the fan outlet section as the boundary.

[0037] In the specific implementation process, the terminal device or the server can obtain input data corresponding to the fan in the server to be predicted (air-cooled server), and the input data includes working condition parameters and geometric parameters of the fan. The working condition parameters at least include the fan speed. The input data is input into the target model (first model and second model), and the prediction result of the physical quantity distribution of the fan outlet section of the fan is obtained; the physical quantity includes at least one of the following: velocity, pressure, and temperature; so that the prediction result can be used as the inlet boundary condition of the server internal flow field of the server to be predicted, and the server internal flow field is simulated to obtain a simulation result. The fan prediction flow field prediction method provided by the present application simplifies the simulation of the fan area flow field of the air-cooled server into a prediction problem of the inlet boundary condition, and constructs a fast prediction model through an artificial intelligence (Artificial Intelligence, AI) algorithm, so that the fan prediction flow field can be efficiently and accurately predicted, thereby improving the efficiency and reliability of the air-cooled server heat dissipation design, and significantly reducing the calculation cost.

[0038] Figure 2 A flowchart of the fan area flow field prediction method provided by the present application is shown in Figure 2 As shown in the figure, the embodiment of the present application provides a fan area flow field prediction method, which is described in detail as follows:

[0039] 201, obtaining input data corresponding to the fan in the server to be predicted; the input data includes working condition parameters and geometric parameters of the fan; the working condition parameters at least include the fan speed.

[0040] The execution subject of the present embodiment is a terminal device or a server.

[0041] The server to be predicted refers to an air-cooled server device whose heat dissipation performance is in the design, verification or optimization stage. That is, the object and target served by the present embodiment. The server can be a physical entity prototype or a digital prototype that has not yet been manufactured.

[0042] The fan refers to a rotating mechanical device used to drive air flow and forcibly convect heat for internal components of the server (such as CPU, GPU, hard disk, memory, etc.). It is usually an axial fan (Axial Fan), and its performance (air volume, air pressure) is mainly determined by the speed and blade geometry. It is the core power source of the server heat dissipation system.

[0043] Input data refers to the set of parameters required to be input into the target model (first model and second model) for prediction. These data are the basis for the model to perform forward calculation and output the prediction results.

[0044] Operating condition parameters describe the variables of fan operating state and environmental conditions. These parameters determine the working point of the fan and directly affect its airflow output characteristics. At least include fan speed. Fan speed is the number of rotations of the fan rotor per unit time. It is the most important and direct parameter to control the fan air volume and air pressure. The change of speed will significantly change the velocity and pressure distribution of the flow field.

[0045] Other possible operating condition parameters can be environmental pressure (atmospheric pressure of the environment where the server is located, affecting air density), environmental temperature (air temperature at the fan inlet, affecting air properties (such as density, viscosity)), system resistance (fan required to overcome the server duct resistance, which is usually reflected in the working point of the fan, that is, the air volume and static pressure value at a certain speed).

[0046] Geometric parameters describe the characteristic data of fan physical shape and structural size. These parameters determine the inherent aerodynamic performance of the fan. For example: global parameters can include fan outer diameter, hub diameter, number of blades. Section parameters can include: airfoil section shape of the blade at different radii, such as: installation angle / pitch angle (the angle between the chord line of the blade and the rotation plane), chord length (the straight line distance between the leading edge and the trailing edge of the airfoil), camber (the maximum distance between the camber line and the chord line of the airfoil), thickness (the maximum thickness of the airfoil). Three-dimensional parameters can include: twist law of the blade, inclination / sweep angle.

[0047] Specifically, by obtaining input data, the target model is provided with sufficient necessary information to uniquely determine the outlet flow field characteristics of a fan under a specific operating condition.

[0048] 202、If the fan speed belongs to the first speed interval, input the input data into the first model to obtain the first prediction result of the physical quantity distribution of the fan outlet section of the fan. The first model has a nonlinear function approximation fitting capability. The maximum fan speed of the first speed interval is less than the first preset threshold.

[0049] 203、If the fan speed belongs to the second speed interval, input the input data and time characteristics into the second model to obtain the second prediction result of the physical quantity distribution of the fan outlet section of the fan. The second model has a space-time modeling capability. The second preset threshold is greater than the first preset threshold. The minimum fan speed of the second speed interval is greater than the second preset threshold. The physical quantity includes at least one of the following: velocity, pressure, temperature.

[0050] The prediction result is used as an inlet boundary condition of a server internal flow field of a server to be predicted, the server internal flow field is simulated, and a simulation result is obtained.

[0051] Specifically, the working conditions and geometric parameters of the fan are input into a pre-trained AI model (i.e., a target model), the model is quickly inferred, and a detailed distribution prediction result of the flow field physical quantity on the fan outlet section is output. The output physical quantity at least includes one of velocity, pressure, and temperature, and the result is a distribution field on the entire two-dimensional section, rather than a single value. The core use of the prediction result is to provide a high-fidelity inlet boundary condition directly to the downstream server internal flow field simulation. This makes it unnecessary to simulate the complex fan rotating area in the internal simulation, and only the calculation from this accurately known inlet condition is needed, thereby greatly reducing the calculation cost and time of subsequent simulation under the premise of ensuring accuracy, and finally efficiently obtaining reliable simulation results of the server internal temperature field, flow field, and the like.

[0052] When the fan speed is in a first speed range, the maximum speed is less than a first preset threshold, at this time, the fluid flow is generally stable, the nonlinear effect is dominant, but the time-varying characteristic is not significant. The system inputs the input data, including the geometric parameters and the working condition parameters, into a first model with strong nonlinear function approximation capability, for example, a deep feedforward neural network or a fully connected neural network. The model has learned the mapping relationship in the stable flow field state through offline training, and can efficiently output high-precision steady-state prediction results of the velocity, pressure, and temperature distribution on the fan outlet section.

[0053] When the fan speed is increased to a second speed range, that is, the minimum speed is greater than a second preset threshold, which is higher than the first preset threshold, the fan region flow often presents strong unsteadiness and complex vortex structure, and the spatial distribution and time evolution characteristics need to be captured. In this case, the system inputs the input data and additional time characteristics into a second model designed for such scenarios. The second model has excellent spatiotemporal modeling capability, which can use long short-term memory networks, time series convolution networks, or transformer Transformer architecture based on attention mechanism, and can effectively handle sequence dependence and predict flow field dynamic evolution. Through the inference of the model, a second prediction result of the physical quantity on the fan outlet section changing with time can be obtained, thereby providing a high-fidelity unsteady inlet boundary condition for the transient simulation of the server internal flow field. This intelligent model scheduling mechanism based on physical working conditions ensures that the prediction has high efficiency and high reliability in the full speed range.

[0054] For example, for the case of low fan speed. Back Propagation (BP) neural network algorithm can be used to directly construct the mapping between the coordinates at the fan outlet section and the physical quantity distribution at the fan outlet interface, complete the model training and optimization, and realize the rapid prediction of the physical quantity at the fan outlet. For the case of high fan speed. A diffusion model coupled with the Transformer algorithm is used to construct a rapid prediction model of the fan outlet flow field distribution under unsteady turbulent flow conditions. At the same time, the residual of the control equation is added to the loss function of the Diffusion in Transformer (DiT) model to improve the accuracy and generalization of the model prediction results.

[0055] The embodiments of the present application can achieve high-precision prediction of the fan outlet flow field in the full operating condition range by adopting a double-model intelligent scheduling strategy based on the speed interval. For low-speed steady flow, a nonlinear fitting model is used to ensure calculation efficiency, and for high-speed unsteady flow, a space-time modeling model is used to accurately capture dynamic characteristics. This not only overcomes the defects of insufficient prediction accuracy of traditional single models under different operating conditions, but also avoids the huge computational overhead caused by complex numerical simulation under high-speed operating conditions, significantly improving the efficiency and reliability of the cooling server design.

[0056] In some embodiments, the method further comprises: if the fan speed belongs to a third speed interval, inputting the input data into the first model to obtain a third prediction result of the physical distribution of the fan outlet section of the fan, inputting the input data and the time feature into the second model to obtain a fourth prediction result of the physical quantity distribution of the fan outlet section of the fan, and fusing the third prediction result and the fourth prediction result to obtain a fifth prediction result of the physical quantity distribution of the fan outlet section of the fan. The minimum fan speed in the third speed interval is greater than or equal to the first preset threshold and the maximum fan speed is less than or equal to the second preset threshold. This embodiment uses a double-model fusion prediction strategy in the transition interval (third speed interval) between the first and second speed intervals, effectively solving the prediction mutation problem at the model switching boundary. By combining the steady-state accuracy of the static model and the transient characteristics of the dynamic model, the prediction smoothness and accuracy under the transition operating condition are significantly improved, ensuring the continuity and reliability of the flow field prediction in the full speed range, and providing more stable boundary condition input for server cooling design.

[0057] In some embodiments, the third prediction result and the fourth prediction result are fused to obtain the fifth prediction result of the physical quantity distribution of the fan outlet section of the fan, comprising:

[0058] The current fan speed is input into a preset model to obtain a fusion weight. The preset model is obtained by training simulation data or experimental data of the fan in the third speed interval.

[0059] The third prediction result and the fourth prediction result are weighted and averaged based on a fusion weight to obtain a fifth prediction result.

[0060] The embodiment realizes adaptive optimal fusion of the double-model prediction results in the transition interval by introducing a weight prediction model trained based on measured or simulated data. The weighting coefficient is dynamically adjusted according to the change in the rotating speed, which not only retains the accuracy advantage of the first model in the near steady-state operating condition, but also fuses the capturing ability of the second model for the unsteady characteristics, so that the prediction results in the transition interval are both smooth and continuous and accurately reflect the dynamic evolution of the flow field, significantly improving the physical authenticity and numerical stability of the boundary condition data.

[0061] In some embodiments, the server to be predicted includes a sensor. The method further includes: acquiring sensing data collected by the sensor.

[0062] If the error between the sensing data and the first prediction result is greater than a preset value and the duration is greater than a preset duration, the first model is adjusted online.

[0063] If the error between the sensing data and the second prediction result is greater than a preset value and the duration is greater than a preset duration, the second model is adjusted online.

[0064] The embodiment introduces a closed-loop verification mechanism of the sensor data and the prediction result, and constructs an intelligent prediction system with online self-correction capability. When it is monitored that there is a persistent deviation between the sensing data and the prediction result, the system automatically triggers the online model adjustment function, so that the model can adapt to the performance drift problem caused by device aging, dust accumulation, installation difference and the like in the actual operating environment in real time. This dynamic correction mechanism significantly improves the long-term prediction accuracy and robustness of the model in complex real scenarios, ensuring the continuous reliability of the boundary conditions of the heat dissipation simulation, and providing protection for the thermal management of the server throughout its life cycle.

[0065] Figure 3 A flowchart of a training method of the first model and the second model provided in the embodiment of the present application is shown. As shown in Figure 3 , the method includes:

[0066] 301. Obtain the physical quantity distribution of the fan outlet section under different operating parameters and geometric parameters of the fan.

[0067] Specifically, the distribution data of physical quantities at the fan outlet section can be obtained through numerical simulation or experimental measurement methods. For example, computational fluid dynamics simulation software can be used to establish a three-dimensional flow field model for different fan geometric parameters (including but not limited to the number of blades, blade inclination angle, and hub ratio) and different operating parameters (including but not limited to rotational speed, inlet flow rate, and ambient temperature). Detailed distribution data of velocity, pressure, and temperature fields at the fan outlet section can be obtained by solving the Navier-Stokes equations. Alternatively, corresponding physical quantity distribution data can be collected in actual wind tunnel tests by deploying experimental equipment such as particle image velocimetry systems, five-hole probe arrays, and thermal imagers.

[0068] In some embodiments, the operating parameters include at least one of the following: fan speed, ambient pressure, and ambient temperature. Obtaining the physical quantity distribution of the fan outlet cross-section under different operating parameters and geometric parameters includes: calculating the physical quantity distribution of the fan outlet cross-section under different operating parameters and geometric parameters using a preset fluid simulation tool; and / or obtaining the physical components of the fan outlet cross-section under different operating parameters and geometric parameters based on particle image velocimetry experiments or laser Doppler velocimetry experiments.

[0069] In this embodiment, a complete training dataset covering a multi-dimensional parameter space is constructed by combining high-fidelity numerical simulation with high-precision experimental measurements. Numerical simulation can efficiently generate a large amount of data covering extreme conditions and special geometric configurations, while experimental methods such as particle image velocimetry provide benchmark data with verification significance in a real physical environment. This data acquisition strategy that combines virtual and real elements ensures both the scale and diversity of the dataset, as well as its physical authenticity and accuracy, laying a solid data foundation for the subsequent training of target models (the first model and the second model) with strong generalization ability and high prediction accuracy.

[0070] 302. Construct the first dataset and the second dataset based on the distribution of physical quantities.

[0071] Specifically, the acquired physical quantity distribution data is standardized and features are extracted to construct a structured dataset. For example, the original data is dimensionless to eliminate the influence of dimensions, and the fan geometric feature parameters and operating condition parameters are extracted as input features. The corresponding flow field physical quantity distribution data are used as output labels and divided into training set, validation set and test set according to a preset ratio to form a standard dataset for model training.

[0072] 303. Train the first model corresponding to the first model based on the first dataset to obtain the first model.

[0073] 304. Train the second model corresponding to the second model based on the second dataset to obtain the second model.

[0074] Specifically, the first training model can be trained by using the first data set in a supervised learning manner, and the second training model can be trained by using the second data set. In the respective training processes, the model parameters can be iteratively optimized by using a back propagation algorithm, so that a loss function between a prediction distribution output by the model and a real distribution in the data set is minimized, and finally a target model (the first model and the second model) meeting a precision requirement is obtained. In the training process, a validation set is used for hyperparameter tuning, and a test set is used for evaluating the generalization performance of the model.

[0075] The embodiments of the application ensure the completeness and representativeness of the model training data by systematically obtaining the flow field data in the multi-dimensional parameter space and constructing the high-quality data set. The model is trained in an end-to-end manner by using the supervised learning manner, so that the model can deeply learn the complex nonlinear mapping relationship between the fan geometric parameters, the working condition parameters and the outlet flow field distribution, and finally an intelligent model with high-precision prediction capability is obtained, thereby providing a reliable data-driven solution for the subsequent precise heat dissipation design and real-time control of the air cooling system.

[0076] In some embodiments, the physical quantity distribution of the fan outlet section under different working condition parameters and geometric parameters of the fan is obtained, including: obtaining a first physical quantity distribution of the fan outlet section under different working condition parameters and geometric parameters of the fan in a first rotational speed range. The maximum fan rotational speed in the first rotational speed range is less than a first preset threshold.

[0077] The first training model is trained according to the first data set to obtain the first model, including: constructing the first training model based on a neural network model algorithm. The first training model includes an input layer, a plurality of hidden layers and an output layer. Each layer of the hidden layers includes a plurality of neurons. The first hyperparameters of the first training model are determined. The first hyperparameters include an activation function, an optimizer and a first loss function. The first training model is trained according to the first hyperparameters based on the first data set to obtain the first model.

[0078] The embodiments of the application focus the training data of the first model on the low rotational speed working condition characteristics by limiting the first rotational speed range and constructing the first data set. The deep neural network structure combined with the specially optimized hyperparameter combination can effectively capture the steady state characteristics and nonlinear laws of the flow field distribution under low rotational speed. This targeted training strategy significantly improves the prediction accuracy and calculation efficiency of the first model in its specific working range, lays a solid foundation for subsequent implementation of partition modeling and fusion prediction, and enhances the expression ability of the model for specific flow state characteristics.

[0079] In some embodiments, the first loss function comprises a first distance term between the predicted value and the true value. Based on the first data set, the first to-be-trained model is trained according to the first hyperparameter to obtain the first model, comprising: inputting the first sample in the first data set into the first to-be-trained model of the last round to obtain the corresponding first predicted value.

[0080] Based on the first loss function, the first loss value corresponding to the first sample is calculated. The first loss value comprises a first distance corresponding to the first distance term.

[0081] Based on the optimizer and the activation function, the target parameter of the first to-be-trained model of the last round is adjusted according to the first loss value to obtain the first to-be-trained model of the current round. The target parameter comprises at least one of the following: the number of hidden layer, the number of neurons of each layer of hidden layer, and the initial learning rate.

[0082] In the embodiments of the present application, the customized loss function containing the distance measure is adopted, and the iterative parameter optimization mechanism is combined to realize the fine control of the training process of the first model. The distance term directly constrains the deviation between the predicted value and the true value, and ensures that the model optimization direction is highly consistent with the accuracy target of the physical quantity distribution. Through the dynamic adjustment of the optimizer and the activation function to the key parameters such as the hidden layer structure and the number of neurons, the model can adaptively learn the special mode of the flow field distribution under the low speed condition, effectively improving the convergence speed and the prediction accuracy of the model, and enhancing the feature capturing ability of the model to the training data.

[0083] As shown in the example of FIG. 1, the training process of the first model comprises the following steps: Figure 4

[0084] 401, based on the simulation tool, the flow field distribution of different fans and blade geometric parameters under lower speed and the fan speed is simulated.

[0085] Specifically, the fluid simulation software is used to calculate the physical quantity distribution at the fan outlet under different speeds and different blade geometric parameters: the commonly used simulation software is used to calculate the flow field around the fan under different speeds and different geometric parameters. When the fan speed is low, the flow field around the fan is in a laminar state, so the physical quantity distribution at the fan outlet under different speeds or different blade geometric parameters is basically constant, i.e. does not change with time or only changes periodically with time.

[0086] 402, collect the velocity, pressure and temperature distribution at the fan outlet under different simulation conditions to construct the data set required for model training.

[0087] Specifically, the velocity, pressure and temperature data at the fan outlet cross section can be intercepted to construct the data set required for model training and verification.

[0088] ​403. Determine the input data, number of hidden layers, and output data for model training, and complete the construction of the BP neural network model (i.e., the first model to be trained).

[0089] A fast prediction model can be built using a backpropagation (BP) neural network algorithm. The model's input is the physical coordinates of the fan outlet interface, with multiple hidden layers, each containing several neurons. The outlet parameters are the velocity, pressure, and temperature distribution at the fan outlet interface. A schematic diagram of the neural network model architecture is shown below. Figure 5 As shown.

[0090] 404. Determine the various hyperparameters of the neural network model, such as activation functions, optimizers, and loss functions.

[0091] The hyperparameters of the neural network are set as follows:

[0092] Activation function: Leaky Rectified Linear Unit (LeakyReLu) can be used. , where x i The input value for the activation function is specifically the weighted input of a neuron in a certain layer of the neural network; y i This is the output value of the activation function, that is, the signal that is passed to the next layer of neurons after being processed by the activation function; for Fixed parameters within the interval.

[0093] Optimizer: An Adaptive Moment Estimation (Adam) optimizer can be used. ,in, This represents the value of the parameter to be optimized (such as the weights or biases in a neural network) after the t-th iteration update; The value of this parameter is represented at the (t-1)th iteration, which is the basis for the current update; η is the learning rate, which is a hyperparameter that controls the step size of parameter updates, and its size directly affects the convergence speed and stability of the model. It is the first moment estimate of the gradient, which is an exponentially weighted moving average of the current and historical gradients, used to smooth gradient fluctuations and reduce the impact of noise. It is the second-moment estimate of the gradient, which is an exponentially weighted moving average of the squares of the current and historical gradients, reflecting the magnitude of gradient changes. Adam is used to normalize gradients, making the updates of parameters in different dimensions more balanced and reasonable. It is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process, and iteratively updates the weights of neural networks based on training data.

[0094] Loss function: mean squared error wherein N represents the total number of samples participating in the calculation; ∑ is the summation symbol, meaning that the correlation items from the 1st to the Nth sample are summed up; represents the predicted value of the i-th sample; represents the true value of the i-th sample. The formula is used to calculate the Euclidean distance between the predicted value and the true value. The closer the predicted value and the true value, the smaller the mean square error of the two.

[0095] In some embodiments, the first to-be-trained model can adopt the framework of a Physics-Informed Neural Network (PINN). In the first loss function, not only the data loss (Mean Squared Error (MSE) of the predicted value and the Computational Fluid Dynamics (CFD) value) is included, but also a physical loss (the predicted result is substituted into the Navier-Stokes equation to calculate the residual error) is added. Through the simultaneous introduction of data loss and physical loss in the first loss function, the strong learning ability of the data-driven method and the physical law of the first principle can be combined. The core pain points faced by pure data-driven or pure physical simulation methods in the flow field prediction of the air-cooled server are fundamentally solved. The model can learn the universal physical law from limited data, accurately predict unknown working conditions, reduce the dependence on expensive CFD data, reduce the total cost of model construction, ensure that the prediction result strictly follows the physical conservation law, eliminate non-physical interpretation, take the physical law as an internal constraint, and increase the transparency and engineering reliability of the model.

[0096] 405、Based on the error between the test result and the true result, the hyperparameters of the neural network are adjusted to complete the optimization of the model prediction result.

[0097] During the model training process, the number of layers of the neural network hidden layer, the number of neurons in each layer of the hidden layer, the initial learning rate and other parameters are adjusted based on the accuracy of the model prediction result, and the optimization of the model prediction ability is completed.

[0098] In some embodiments, the physical quantity distribution of the fan outlet section under different working condition parameters and geometric parameters of the fan is obtained, including: obtaining the second physical quantity distribution of the fan outlet section at different time points of the fan under different working condition parameters and geometric parameters in a second rotating speed interval. The minimum fan rotating speed of the second rotating speed interval is greater than a second preset threshold.

[0099] The second target model is obtained by training the second to-be-trained model according to a second data set, including: constructing the second to-be-trained model based on a diffusion model algorithm of a self-attention mechanism. The second to-be-trained model includes a conditional encoder, a diffusion module based on self-attention, and a function space decoder. A second loss function of the second to-be-trained model is determined. The second loss function includes a second distance term and a physical constraint term. The second to-be-trained model is trained according to the second loss function based on the second data set, and a second model is obtained.

[0100] The embodiments of the present application innovatively use a diffusion model architecture based on a self-attention mechanism, and combine a loss function containing physical constraints, to achieve high-precision modeling of transient flow fields in view of the significant characteristics of flow field dynamic characteristics in the high-speed range. The conditional encoder effectively fuses high-speed operating parameters, the self-attention mechanism captures the spatiotemporal correlation characteristics of the flow field, the diffusion module gradually learns the data distribution, and finally the function space decoder outputs the physical quantity distribution. The distance term in the second loss function ensures data-driven accuracy, and the physical constraint term embeds the basic principles of fluid mechanics (such as mass conservation, momentum conservation), enhancing the physical reasonableness and generalization ability of the prediction results. This scheme significantly improves the accuracy and reliability of flow field prediction under high-speed and transient operating conditions, and provides an effective solution for intelligent simulation of complex flow phenomena.

[0101] In some embodiments, the physical constraint term includes a residual of a fluid mechanics control equation or a residual of a boundary condition. The residual of the fluid mechanics control equation includes residuals of continuity equations, momentum equations, and energy equations. The residual of the boundary condition includes an inlet velocity residual, an outlet pressure residual, and a blade surface temperature residual.

[0102] The embodiments of the present application embed the residuals of the fluid mechanics control equation and the boundary condition as the physical constraint term into the loss function, realizing the deep integration of physical priori knowledge and data-driven modeling. This physical information embedding mechanism effectively constrains the optimization direction of the model, ensuring that the prediction results strictly follow the basic physical laws such as mass conservation, momentum conservation, and energy conservation, while meeting the boundary condition constraints of actual engineering problems. This not only significantly improves the generalization ability of the model in the sparse region of the training data, but also enhances the physical reasonableness and reliability of the prediction results, especially suitable for accurate prediction of complex flow phenomena under high-speed operating conditions, and provides an innovative solution for constructing intelligent fluid simulation models with physical consistency.

[0103] In some embodiments, the second model is obtained by training the second to-be-trained model according to the second loss function based on the second data set, including: sampling multiple sample data from the second data set, which includes the real distribution of physical quantities at the fan outlet cross section under specific operating conditions, geometric parameters, and time points.

[0104] Based on a preset noise scheduling strategy, Gaussian noise is added to the second sample in the second data set to obtain noisy data.

[0105] The noisy data and the corresponding time are input into the second trained model of the previous round, and the diffusion module based on the self-attention mechanism in the second trained model of the previous round is used for forward calculation to output predicted noise corresponding to the Gaussian noise.

[0106] Based on the Gaussian noise and the predicted noise, a second loss value corresponding to the second sample is determined based on a second loss function. The second loss value includes a second distance corresponding to a second distance term and a residual corresponding to a physical constraint term. The second distance is the distance between the Gaussian noise and the predicted noise. The residual is the residual of the fluid mechanics control equation.

[0107] Based on the gradient descent algorithm and the second loss value, the parameters of the second trained model are adjusted to obtain the second trained model of the current round.

[0108] The embodiments of the present application introduce a noise prediction training paradigm of the diffusion model, and combine a loss function with physical constraints to achieve efficient and physically consistent modeling of high-speed transient flow fields. By gradually adding Gaussian noise through a noise scheduling strategy, the model learns the inverse process of recovering the true flow field distribution from noise; the self-attention mechanism is used to effectively capture the spatiotemporal correlation characteristics of the flow field; the distance loss between the predicted noise and the true noise ensures the accuracy of the denoising process; at the same time, the residual of the fluid mechanics control equation is embedded as a physical constraint term in the optimization objective, forcing the model to output results that satisfy the basic physical laws. This training strategy not only improves the model's ability to capture details of complex transient flow and prediction accuracy, but also ensures that the generated flow field has strict physical reasonableness, significantly enhancing the model's generalization performance and reliability in data sparse areas, providing an innovative solution for high-fidelity flow field intelligent prediction.

[0109] As shown in Figure 6 The training process of the second model can include the following steps:

[0110] 601. Based on a simulation tool, the flow field distribution under different fan and blade geometric parameters and fan rotation speeds at high rotation speeds is simulated.

[0111] Specifically, a fluid simulation software can be used to calculate the physical quantity distribution at the fan outlet under different rotation speeds and different blade geometric parameters at high rotation speeds: a commonly used simulation software is used to calculate the flow field around the fan under different rotation speeds and different geometric parameters at high rotation speeds. When the fan rotation speed is high, the flow field around the fan will enter a turbulent state, and the physical quantity distribution at the fan outlet will change over time.

[0112] 602、Collect the speed, pressure and temperature distribution at the fan outlet under different simulation conditions and at different times under the same condition to build the data set required for model training.

[0113] Specifically, unlike the embodiment shown in Figure 4 , in this embodiment, in addition to collecting the physical quantity distribution at the fan outlet cross section under different simulation conditions, the physical quantity distribution at the fan outlet cross section at different times under the same simulation condition also needs to be collected, and the data set used for model training is constructed based on the collected data.

[0114] 603、Based on the collected data set, the input data and output data of the model are arranged, and a fast prediction model (i.e., the second to-be-trained model) is constructed using a diffusion algorithm based on attention.

[0115] Specifically, based on the obtained simulation data, a diffusion algorithm based on attention (DiffusionTransformer) can be used to construct a prediction model of the physical quantity at the fan outlet cross section, as shown in Figure 7 , the specific architecture of the DiT model is as follows:

[0116] The entire architecture consists of three main components: a conditional encoder, a DiT module, and a function space decoder. The conditional decoder is used to process the input data to generate an initial latent representation. This encoded information is then input into the DiT block, where noise is added and a series of linear transformations, reshaping, layer normalization, and attention mechanisms are performed to effectively model the underlying data distribution. The DiT block uses a multi-head self-attention mechanism, point-wise feedforward networks, and residual connections to capture complex relationships between spatial variables. Finally, the function decoder decodes the learned representation into meaningful output data, such as the speed, pressure, and temperature distribution at the fan outlet.

[0117] 604、Couple the residual of the control equation with the residual of the boundary condition with the loss function of the DiT model to obtain the final loss function of the prediction model.

[0118] Increase the physical constraints in the residual of the DiT model:

[0119] The residual of the DiT model itself is obtained by the following formula:

[0120] (1)

[0121] Where t represents the time step, and represent the initial state and the flow field variables such as velocity, pressure, etc. corresponding to the t time step, respectively. is the input condition, including the coordinates of the fan outlet cross section and the time, etc. is the Gaussian noise added in the forward diffusion process, Representative model learns optimal parameters, Then the code mathematical expectation.

[0122] The basic control equations are Navier-Stokes equations, which consist of continuity equation , momentum equation and energy equation. The final residual of physical constraints is composed of the residual of control equation and the residual of boundary condition, and the residual of control equation is as follows:

[0123] (2)

[0124] Wherein, is the residual of continuity equation, which is used to measure the satisfaction degree of continuity equation; represents the change rate of density p with time t; is the divergence of the product of density p and velocity vector V, which describes the transmission and change of fluid mass in space.

[0125] (3)

[0126] Wherein, is the residual of momentum equation, which is used to measure the satisfaction degree of momentum equation; p is the density of fluid; represents the change rate of velocity vector V with time t; is the dot product of velocity vector V and velocity gradient , which reflects the convection effect of velocity field; is the gradient of pressure p, which describes the influence of pressure change in space on fluid momentum; is the stress tensor, which reflects the influence of internal viscous stress of fluid on momentum; represents the volume force (such as gravity) per unit volume, and f is the vector form of volume force.

[0127] (4)

[0128] Wherein, is the residual of energy equation, which is used to measure the satisfaction degree of energy equation; p is the density of fluid; represents the change rate of specific internal energy e with time t; is the dot product of velocity vector V and specific internal energy gradient , which reflects the convection transfer of specific internal energy; is the divergence of heat flux density vector q, which describes the heat transfer process such as heat conduction; involves the product of pressure p and velocity divergence , which reflects the influence of fluid volume change on energy; is the stress tensor τ and velocity gradient The double dot product represents the effect of viscous dissipation and other factors on energy. The term representing the volumetric heat source per unit volume of fluid, It is the heat source intensity per unit mass of fluid.

[0129] Let be the viscous stress tensor, and its expression is:

[0130] (5)

[0131] Where τ is the stress tensor, used to describe mechanical effects such as viscous stress inside the fluid; μ is the dynamic viscosity of the fluid, reflecting the fluid's ability to resist shear deformation. It is the gradient of the velocity vector V, which reflects the change of velocity in space; The sum of the velocity gradient and its transpose is used to construct the symmetric tensor part related to viscous stress; middle, It is the divergence of the velocity vector, describing the rate of change of fluid volume, where I is the unit tensor. This term is used to account for the effect of fluid volume change on viscous stress. It is the correlation coefficient.

[0132] The three residuals coupled above can be used to obtain the residuals of the physical equations, namely:

[0133] (6)

[0134] Regarding boundary conditions, the residuals for the fan inlet boundary conditions, outlet boundary conditions, and fan blade wall boundary conditions need to be provided. The residual for the inlet boundary conditions is: ,in This represents the number of grid points at the entrance boundary. Let be the velocity at any point on the entrance boundary. This is the pre-set inlet velocity. The residual of the outlet boundary condition is: ,in This represents the number of grid points at the export boundary. For the pressure at the grid points at the export boundary, The pre-set outlet boundary pressure value. The final residual of the blade surface boundary conditions is: .in This represents the number of grid points on the blade surface. The velocity of the grid points on the blade surface. For the temperature of the blade surface and A temperature preset for the blade surface.

[0135] The final total loss function can be expressed as:

[0136] (7)

[0137] in, These are the weighting coefficients corresponding to different residuals.

[0138] 605. Begin training the DiT model, complete the forward diffusion and reverse denoising processes, and obtain the final fast prediction model.

[0139] Model training primarily involves two key steps: forward diffusion and backward denoising. In the forward diffusion step, Gaussian noise is progressively added to the input flow field, gradually transforming it into a noise distribution. Then, the backward denoising process iteratively reconstructs the original flow field, using Bayesian inference to model the transition probabilities. Finally, these processes enable the model to effectively approximate the fundamental distribution and generate accurate predictions.

[0140] The method provided in this application optimizes the problem of excessively high prediction costs for the flow field in the fan region of air-cooled servers. It simplifies the flow field in the fan region to the inlet boundary conditions of the flow field within the server and constructs a fast prediction model based on different fan speed ranges. For lower fan speeds, a BP neural network algorithm is used to construct the mapping relationship between the geometric coordinates at the fan outlet section and the distribution of physical quantities such as velocity, pressure, and temperature under different blade geometric parameters, enabling rapid prediction of the physical quantity distribution at the fan outlet. For higher fan speeds, the flow around the fan enters a turbulent state, exhibiting high unsteadiness. An attention-based diffusion algorithm, which has strong predictive capabilities for nonlinear and multi-scale flows, is used to construct the prediction model. Physical equation residuals are added to the model loss function to improve the accuracy of the model prediction results, achieving rapid prediction of the physical quantity distribution at the fan outlet under high speed conditions.

[0141] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0142] Figure 8 This is a schematic diagram of the structure of the fan region flow field prediction device provided in an embodiment of this application. Figure 8 As shown, embodiments of this application also provide a fan region flow field prediction device 80, including:

[0143] The acquisition module 801 is used to acquire the input data corresponding to the fan in the server to be predicted; the input data includes the fan's operating parameters and geometric parameters; the operating parameters include at least the fan speed.

[0144] The input module 802 is used to input the input data into the first model if the fan speed is within the first speed range, and obtain the first prediction result of the physical quantity distribution of the fan outlet section; the maximum fan speed in the first speed range is less than the first preset threshold.

[0145] Input module 802 is further configured to input the input data and time features into the second model to obtain a second prediction result of the physical quantity distribution if the fan speed belongs to the second speed range; the minimum fan speed in the second speed range is greater than a second preset threshold; the second preset threshold is greater than a first preset threshold; the physical quantity includes at least one of the following: speed, pressure, and temperature.

[0146] The fan region flow field prediction device provided in this application simplifies the simulation of the flow field in the fan region of an air-cooled server into the prediction problem of the inlet boundary conditions, and constructs a fast prediction model through AI algorithms, thereby enabling efficient prediction of the fan flow field. Furthermore, by adopting a dual-model intelligent scheduling strategy based on the rotational speed range, it can achieve high-precision prediction of the fan outlet flow field across the entire operating range. For low-speed steady-state flow, a nonlinear fitting model is used to ensure computational efficiency, while for high-speed unsteady flow, a spatiotemporal modeling model is used to accurately capture dynamic characteristics. This overcomes the shortcomings of insufficient prediction accuracy of traditional single models under different operating conditions, and avoids the huge computational overhead caused by complex numerical simulations under high-speed operating conditions.

[0147] In some embodiments, the input module 802 is specifically used to: if the fan speed belongs to a third speed range, input the input data into a first model to obtain a third prediction result of the physical distribution of the fan outlet cross-section; input the input data and time features into a second model to obtain a fourth prediction result of the physical quantity distribution of the fan outlet cross-section; and fuse the third and fourth prediction results to obtain a fifth prediction result of the physical quantity distribution of the fan outlet cross-section. The minimum fan speed in the third speed range is greater than or equal to a first preset threshold and the maximum fan speed is less than or equal to a second preset threshold.

[0148] In some embodiments, the input module 802 is specifically used to: input the current fan speed into a preset model to obtain fusion weights. The preset model is trained based on simulation data or experimental data of the fan in the third speed range.

[0149] Based on the fusion weights, the third and fourth prediction results are weighted and averaged to obtain the fifth prediction result.

[0150] In some embodiments, the server to be predicted includes sensors. The method further includes:

[0151] Acquire sensor data collected by the sensor.

[0152] If the error between the sensor data and the second prediction result is greater than a preset value and the duration is longer than a preset duration, the second model will be adjusted online.

[0153] In some embodiments, the device 80 further includes a training module 803, which is specifically used to: acquire the physical quantity distribution of the fan outlet cross-section under different operating conditions and geometric parameters; construct a first dataset and a second dataset based on the physical quantity distribution; train a first model to be trained corresponding to the first model based on the first dataset to obtain a first model; and train a second model to be trained corresponding to the second model based on the second dataset to obtain a second model.

[0154] In some embodiments, the operating parameters include at least one of the following: fan speed, ambient pressure, and ambient temperature. The training module 803 is specifically used to: calculate the physical quantity distribution of the fan outlet cross-section under different operating parameters and geometric parameters based on a preset fluid simulation tool; and / or, obtain the physical components of the fan outlet cross-section under different operating parameters and geometric parameters based on particle image velocimetry experiments or laser Doppler velocimetry experiments.

[0155] In some embodiments, the training module 803 is specifically used to: obtain the first physical quantity distribution of the fan outlet cross-section under different operating conditions and geometric parameters within a first speed range. The maximum fan speed in the first speed range is less than a first preset threshold.

[0156] A first training model is constructed based on a neural network model algorithm. The first training model includes an input layer, multiple hidden layers, and an output layer. Each hidden layer contains multiple neurons.

[0157] Determine the first hyperparameters of the first model to be trained. The first hyperparameters include the activation function, optimizer, and first loss function.

[0158] Based on the first dataset, the first model to be trained is trained according to the first hyperparameters to obtain the first model.

[0159] In some embodiments, the first loss function includes a first distance term between the predicted value and the true value. The training module 803 is specifically configured to: input the first sample from the first dataset into the first model to be trained in the previous round to obtain the corresponding first predicted value.

[0160] Based on the first loss function, calculate the first loss value corresponding to the first sample. The first loss value includes the first distance corresponding to the first distance term.

[0161] Based on the optimizer and activation function, the target parameters of the first model to be trained in the previous round are adjusted according to the first loss value to obtain the first model to be trained in the current round. The target parameters include at least one of the following: the number of hidden layers, the number of neurons in each hidden layer, and the initial learning rate.

[0162] In some embodiments, the training module 803 is specifically used to: obtain the second physical quantity distribution of the fan outlet cross-section at different time points under different operating conditions and geometric parameters within the second speed range. The minimum fan speed in the second speed range is greater than a second preset threshold.

[0163] A second training model is constructed using a diffusion model algorithm based on a self-attention mechanism. The second training model includes a conditional encoder, a self-attention-based diffusion module, and a function space decoder.

[0164] Determine the second loss function for the second model to be trained. The second loss function includes a second distance term and a physical constraint term.

[0165] Based on the second dataset, the second training model is trained according to the second loss function to obtain the second model.

[0166] In some embodiments, physical constraint terms include residuals of the fluid dynamics governing equations or residuals of the boundary conditions. The residuals of the fluid dynamics governing equations include residuals of the continuity equation, momentum equation, and energy equation. The residuals of the boundary conditions include inlet velocity residuals, outlet pressure residuals, and blade surface temperature residuals.

[0167] In some embodiments, the training module 803 is specifically used to: sample multiple sample data from a second dataset, which includes the true distribution of physical quantities of the fan outlet section under specific operating conditions, geometric parameters and time points.

[0168] Based on a preset noise scheduling strategy, Gaussian noise is added to the second sample in the second dataset to obtain the noisy data.

[0169] The noisy data and the corresponding time are input into the second training model of the previous round. The forward calculation is performed by the diffusion module based on the self-attention mechanism in the second training model of the previous round, and the predicted noise corresponding to the Gaussian noise is output.

[0170] Based on the Gaussian noise and the predicted noise, a second loss value is determined for the second sample using a second loss function. The second loss value includes the second distance corresponding to the second distance term and the residual corresponding to the physical constraint term. The second distance is the distance between the Gaussian noise and the predicted noise. The residual is the residual from the fluid dynamics governing equations.

[0171] Based on the gradient descent algorithm and the second loss value, the parameters of the second model to be trained are adjusted to obtain the second model to be trained in the current round.

[0172] For a description of the features in the embodiment corresponding to the fan region flow field prediction device, please refer to the relevant description of the embodiment corresponding to the fan region flow field prediction method, which will not be repeated here.

[0173] Figure 9 A schematic diagram of the structure of the electronic device provided in this application. Figure 9 As shown, the electronic device 90 provided in this embodiment includes at least one processor 901 and a memory 902. Optionally, the electronic device 90 further includes a communication component 903. The processor 901, memory 902, and communication component 903 are connected via a bus.

[0174] In the specific implementation process, at least one processor 901 executes computer execution instructions stored in memory 902, causing at least one processor 901 to execute the above-described embodiment of the fan region flow field prediction method.

[0175] The specific implementation process of processor 901 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0176] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0177] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0178] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0179] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described XX method embodiments when it is run.

[0180] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0181] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the fan region flow field prediction method.

[0182] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the fan region flow field prediction method.

[0183] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0184] The flow field prediction method for a fan region provided in this application has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only intended to help understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A method for predicting the flow field in a fan region, characterized in that, include: Obtain the input data corresponding to the fan in the server to be predicted; the input data includes the operating parameters and geometric parameters of the fan; the operating parameters include at least the fan speed; If the fan speed is within a first speed range, the input data is input into the first model to obtain a first prediction result of the physical quantity distribution of the fan outlet section; the maximum fan speed in the first speed range is less than a first preset threshold; the first model is a nonlinear fitting model, which is a neural network model; the first model has learned the mapping relationship under a stable flow field state through offline training, and outputs high-precision steady-state prediction results of velocity, pressure and temperature distribution on the fan outlet section; If the fan speed belongs to the second speed range, the input data and time features are input into the second model to obtain the second prediction result of the physical quantity distribution; the minimum fan speed in the second speed range is greater than the second preset threshold. The second preset threshold is greater than the first preset threshold; the physical quantity includes at least one of the following: velocity, pressure, temperature; the second model is a spatiotemporal modeling model, which is a diffusion model based on a self-attention mechanism; the second model adopts a long short-term memory network, a temporal convolutional network, or a Transformer architecture based on an attention mechanism to handle sequential dependencies and predict the dynamic evolution of the flow field; and, through the second model inference, the second prediction result of the change of the physical quantity of the fan outlet section with time is obtained, providing high-fidelity unsteady inlet boundary conditions for transient simulation of the internal flow field of the server; The method further includes: Obtain the distribution of physical quantities at the fan outlet cross-section under different operating parameters and geometric parameters; Construct a first dataset and a second dataset based on the distribution of the physical quantities; The first model is obtained by training the first model corresponding to the first model based on the first dataset. The second model is obtained by training the second model corresponding to the second model based on the second dataset; The acquisition of the physical quantity distribution of the fan outlet cross-section under different operating parameters and geometric parameters includes: The distribution of the second physical quantity of the fan outlet section at different time points under different operating conditions and geometric parameters within the second speed range is obtained; the minimum fan speed in the second speed range is greater than a second preset threshold. The step of training the second model corresponding to the second model based on the second dataset to obtain the second model includes: The second training model is constructed using a diffusion model algorithm based on a self-attention mechanism; the second training model includes a conditional encoder, a diffusion module based on self-attention, and a function space decoder. Determine the second loss function for the second model to be trained; the second loss function includes a second distance term and a physical constraint term. Based on the second dataset, the second model to be trained is trained according to the second loss function to obtain the second model.

2. The method according to claim 1, characterized in that, The method further includes: If the fan speed belongs to the third speed range, the input data is input into the first model to obtain the third prediction result of the physical distribution of the fan outlet section. The input data and time features are input into the second model to obtain the fourth prediction result of the physical quantity distribution of the fan outlet section. The third prediction result and the fourth prediction result are fused to obtain the fifth prediction result of the physical quantity distribution of the fan outlet section. The minimum fan speed in the third speed range is greater than or equal to the first preset threshold and the maximum fan speed is less than or equal to the second preset threshold.

3. The method according to claim 1, characterized in that, The server to be predicted includes sensors; the method further includes: Acquire the sensing data collected by the sensor; If the error between the sensing data and the second prediction result is greater than a preset value and the duration is greater than a preset duration, then the second model is adjusted online.

4. The method according to any one of claims 1-3, characterized in that, The operating parameters include at least one of the following: fan speed, ambient pressure, and ambient temperature; The acquisition of the physical quantity distribution of the fan outlet cross-section under different operating parameters and geometric parameters includes: Based on the preset fluid simulation tool, the physical quantity distribution of the fan outlet section under different operating parameters and geometric parameters is calculated; And / or, Based on particle image velocimetry experiments or laser Doppler velocimetry experiments, the physical components of the fan outlet cross section under different operating conditions and geometric parameters are obtained.

5. The method according to any one of claims 1-3, characterized in that, The physical constraint terms include residuals of the fluid dynamics governing equations or residuals of the boundary conditions; the residuals of the fluid dynamics governing equations include residuals of the continuity equation, momentum equation, and energy equation; the residuals of the boundary conditions include inlet velocity residuals, outlet pressure residuals, and blade surface temperature residuals.

6. The method according to any one of claims 1-3, characterized in that, The step of training the second model based on the second dataset and according to the second loss function to obtain the second model includes: Multiple sample data are sampled from the second dataset, which includes the true distribution of physical quantities of the fan outlet section under specific operating conditions, geometric parameters and time points; Based on a preset noise scheduling strategy, Gaussian noise is added to the second sample in the second dataset to obtain the noisy data. The noisy data and the corresponding time are input into the second training model of the previous round. Forward calculation is performed through the diffusion module based on the self-attention mechanism in the second training model of the previous round, and the predicted noise corresponding to the Gaussian noise is output. Based on the Gaussian noise and the predicted noise, a second loss value corresponding to the second sample is determined using the second loss function; the second loss value includes a second distance corresponding to the second distance term and a residual corresponding to the physical constraint term; the second distance is the distance between the Gaussian noise and the predicted noise; the residual is the residual of the fluid dynamics control equation. Based on the gradient descent algorithm and the second loss value, the parameters of the second model to be trained are adjusted to obtain the second model to be trained in the current round.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the fan region flow field prediction method as described in any one of claims 1 to 6 when executing the computer program.

Images