Training method for temperature field determination model, temperature field determination method and device

Abstract

The present invention provides a training method, a temperature field determination method and an apparatus for a temperature field determination model. The training method for the temperature field determination model includes: obtaining a sample temperature data set about a target hardware module, the sample temperature data set including a real temperature data set, a simulated temperature data set and a generated temperature data set; performing transfer learning training on a pre-trained source model based on the real temperature data set and the simulated temperature data set to obtain an intermediate model, wherein the source model includes a physical information neural network with target heat conduction equation constraints; and training the intermediate model based on the real temperature data set, the simulated temperature data set and the generated temperature data set to obtain a trained temperature field determination model.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence technology, computer technology, and equipment thermal management technology, and more specifically, to a training method for a temperature field determination model, a temperature field determination method, and a device. Background Art

[0002] In the field of computer technology, the temperature of key hardware modules within electronic devices (such as the CPU, GPU, and power supply) directly affects the stability, availability, and operating life of the device. Excessively high hardware temperatures can cause performance degradation or even permanent damage, necessitating temperature monitoring to ensure timely cooling measures. However, due to limitations such as limited internal space and cost constraints, it is often not possible to densely deploy temperature sensors within the device to monitor the entire device's temperature. To address this, a temperature field reconstruction method can be employed, reconstructing the entire temperature field using a limited amount of temperature data collected by sparsely deployed temperature sensors. This allows for global temperature field perception and thermal management control of key hardware modules.

[0003] However, related temperature field reconstruction methods have technical problems such as complex multi-physical field coupling, low computational efficiency, poor real-time performance, insufficient dynamic adaptability, and strong data dependence. Summary of the Invention

[0004] In view of the above problems, the present invention provides a training method for a temperature field determination model, a temperature field determination method, an apparatus, a device, a medium and a program product.

[0005] According to one aspect of the present invention, a training method for a temperature field determination model is provided, comprising: obtaining a sample temperature data set about a target hardware module, the sample temperature data set including a real temperature data set, a simulated temperature data set and a generated temperature data set; wherein the real temperature data set includes a plurality of first sample temperature data collected by a plurality of temperature sensors arranged on the target hardware module, the simulated temperature data set includes a plurality of second sample temperature data obtained by finite element simulation based on a simulation physical model and simulation conditions of the target hardware module, and the generated temperature data set includes a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data and the third sample temperature data all include a sample time, a sample temperature value and a sample temperature position corresponding to the sample temperature value; based on the real temperature data set and the simulated temperature data set, performing transfer learning training on a pre-trained source model to obtain an intermediate model, wherein the source model includes a physical information neural network with target heat conduction equation constraints; and based on the real temperature data set, the simulated temperature data set and the generated temperature data set, training the intermediate model to obtain a trained temperature field determination model.

[0006] Another aspect of the present invention provides a temperature field determination method, including: obtaining a historical temperature data set of a target hardware module, the historical temperature data including historical collection time, historical temperature values ​​and collection location, the collection location representing the layout position of the temperature sensor relative to the target hardware module; inputting the target time and historical temperature data into a temperature field determination model, and outputting a target temperature data set, the target temperature data set being used to represent the temperature field of the target hardware module at the target time; wherein the temperature field determination model is trained using the training method of the temperature field determination model as described above.

[0007] Another aspect of the present invention provides a training device for a temperature field determination model, comprising: a first acquisition module, configured to acquire a sample temperature data set about a target hardware module, the sample temperature data set comprising a real temperature data set, a simulated temperature data set, and a generated temperature data set; wherein the real temperature data set comprises a plurality of first sample temperature data acquired by a plurality of temperature sensors disposed on the target hardware module, the simulated temperature data set comprises a plurality of second sample temperature data acquired by finite element simulation based on a simulated physical model and simulation conditions of the target hardware module, and the generated temperature data set comprises a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data, and the third sample temperature data all comprise a sample time, a sample temperature value, and a sample temperature position corresponding to the sample temperature value; a first training module, configured to perform transfer learning training on a pre-trained source model based on the real temperature data set and the simulated temperature data set to obtain an intermediate model, wherein the source model comprises a physical information neural network constrained by a target heat conduction equation; and a second training module, configured to train the intermediate model based on the real temperature data set, the simulated temperature data set, and the generated temperature data set to obtain a trained temperature field determination model.

[0008] Another aspect of the present invention provides a temperature field determination device, including: a second acquisition module, used to obtain historical temperature data of a target hardware module, the historical temperature data including historical collection time, historical temperature value and collection location, the collection location representing the layout position of the temperature sensor relative to the target hardware module; a prediction module, used to input the target time and historical temperature data into a temperature field determination model, and output a target temperature data set, the target temperature data set being used to represent the temperature field of the target hardware module at the target time; wherein the temperature field determination model is trained using the training method of the temperature field determination model as described above.

[0009] Another aspect of the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the above method.

[0010] Another aspect of the present invention further provides a computer-readable storage medium having a computer program or instructions stored thereon, which implements the steps of the above method when the computer program or instructions are executed by a processor.

[0011] Another aspect of the present invention further provides a computer program product, comprising a computer program or instructions, which implement the steps of the above method when executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings.

[0013] Figure 1 A diagram illustrating application scenarios of a temperature field determination model training method, a temperature field determination method, an apparatus, a device, a medium, and a program product according to an embodiment of the present invention is shown.

[0014] Figure 2 A flow chart of a method for training a temperature field determination model according to an embodiment of the present invention is shown.

[0015] Figure 3 A schematic diagram showing the principles of a target generative adversarial network according to an embodiment of the present invention is shown.

[0016] Figure 4A A schematic diagram of transfer learning according to an embodiment of the present invention is shown.

[0017] Figure 4B A schematic diagram of training a physical information neural network according to an embodiment of the present invention is shown.

[0018] Figure 5 A flow chart of a temperature field determination method according to an embodiment of the present invention is shown.

[0019] Figure 6 A structural block diagram of a training device for a temperature field determination model according to an embodiment of the present invention is shown.

[0020] Figure 7 A structural block diagram of a temperature field determination device according to an embodiment of the present invention is shown.

[0021] Figure 8 A block diagram of an electronic device suitable for implementing a training method for a temperature field determination model and a temperature field determination method according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0022] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0023] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the presence of the features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0024] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0025] When expressions such as "at least one of A, B, and C, etc." are used, they should generally be interpreted in accordance with the meaning commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0026] The embodiments of the present invention relate to the fields of computer technology and device thermal management technology, and can be applied to the field of server thermal management technology, for example. To facilitate understanding, some technical terms mentioned in this article are introduced and explained here:

[0027] CPU (Central Processing Unit), as the computing and control core of the computer system, is the final execution unit for information processing and program running.

[0028] GPU (Graphics Processing Unit) is a microprocessor specifically designed to process graphics and computing tasks. Through its multi-core architecture and high-throughput design, it can efficiently handle complex graphics and computing tasks and is widely used in high-performance computing, artificial intelligence, scientific computing and other fields.

[0029] PCB (Printed Circuit Board) is the physical carrier of electronic components, used for mechanical support and electrical connection of electronic components.

[0030] The server motherboard is one of the most basic and important components of a server. It connects and coordinates all hardware devices within the server, transmits power and signals, and connects the processor, memory, storage devices, expansion slots, and other key components. For example, a server motherboard is typically composed of one or more PCBs. Server PCBs are characterized by high layer count, high density, and high transmission rates. They are typically constructed from multiple stacked metal and dielectric layers to improve electrical signal transmission quality and interference resistance.

[0031] The finite element method (FEM) is a technique for finding approximate solutions to boundary value problems of partial differential equations. It can be used to solve complex engineering and physics problems. It discretizes a continuous physical domain into a finite number of small units (called "finite elements") and approximates the problem using these units, enabling numerical simulation of complex systems. The FEM is widely used in fields such as structural mechanics, heat conduction, fluid mechanics, and electromagnetics.

[0032] Physics-Informed Neural Networks (PINNs) are machine learning models that combine deep learning and the laws of physics. They embed the laws of physics (usually expressed as partial differential equations) into the neural network's loss function, enabling the network to learn not only based on data but also subject to the constraints of physical laws. The core idea of ​​PINNs is to use neural networks to approximate the solutions to partial differential equations. For example, the input to the neural network is spacetime coordinates (such as position and time), and the output is the values ​​of physical variables (such as temperature, pressure, and velocity) at these coordinates. By adding the residual of the partial differential equation (PDE) (i.e., the left side minus the right side of the PDE) to the loss function, the network considers both data fit and compliance with physical laws during training.

[0033] Generative Adversarial Networks (GANs) are deep learning models that generate new data samples through adversarial training between a generator and a discriminator. The generator attempts to trick the discriminator into misclassifying generated samples as real samples, while the discriminator attempts to correctly distinguish between real and generated samples. This adversarial process ultimately enables the generator to produce high-quality samples.

[0034] Transfer learning is a machine learning method that leverages the knowledge of a model pre-trained on a related task (called the source model) to improve learning performance on a new task (called the target task). The core idea of ​​transfer learning is that knowledge learned from one task can be transferred to another related task, thereby reducing the amount of training data and computing resources required for the target task.

[0035] In the field of computer technology, the temperature of key hardware modules within electronic devices (such as the CPU, GPU, and power supply) directly affects the stability, availability, and operating life of the device. Excessively high hardware temperatures can cause performance degradation or even permanent damage, necessitating temperature monitoring to ensure timely cooling measures. However, due to limitations such as limited internal space and cost constraints, it is often not possible to densely deploy temperature sensors within the device to monitor the entire device's temperature. To address this, a temperature field reconstruction method can be employed, reconstructing the entire temperature field using a limited amount of temperature data collected by sparsely deployed temperature sensors. This allows for global temperature field perception and thermal management control of key hardware modules.

[0036] However, related temperature field reconstruction methods have technical problems such as complex multi-physical field coupling, low computational efficiency, poor real-time performance, insufficient dynamic adaptability, and strong data dependence.

[0037] In one embodiment, for example, in the server technology field, key server hardware modules such as the CPU, GPU, and power supply are typically installed on the server motherboard (typically composed of multiple PCB circuit boards). These hardware modules generate heat during operation and are also extremely sensitive to temperature changes, requiring effective heat dissipation measures to ensure their normal operation. To meet the heat dissipation requirements of key hardware modules such as the CPU, GPU, and power supply, for example, temperature sensors can be deployed on the server motherboard to monitor their temperature data. Once the BMC (Baseboard Management Controller) detects that the temperature of these key hardware modules exceeds a set threshold based on the temperature data collected by the temperature sensors, it will cool them down through heat dissipation measures such as air cooling or liquid cooling to ensure stable operation of the server.

[0038] However, in actual engineering practice, only a limited number of sensors can monitor temperature data at certain locations, making it impossible to monitor the temperature at any location on the server motherboard. Furthermore, temperature sensors can only monitor the surface temperature of the server motherboard, but not its internal temperature. While the temperature monitoring range can be expanded by increasing the number of temperature sensors, this also increases costs. Furthermore, the server motherboard has a limited area, and adding temperature sensors will also require adjustments or reductions to other components. To address this issue, a temperature field reconstruction method can be used to reconstruct the temperature at any location on the server motherboard using temperature data collected by a small number of temperature sensors, thus enabling full-domain temperature field perception of the server motherboard.

[0039] However, related temperature field reconstruction methods include, for example, methods based on physical models (such as the finite element method (FEM), data-driven methods (such as interpolation regression and machine learning), and hybrid methods based on physical models and data-driven methods (such as physical information neural network (PINN)). However, these methods have the following problems:

[0040] (1) Deficiencies of the Finite Element Method (FEM)

[0041] ① High computational cost: transient problems require iterative solutions, and the time step is limited (such as the stability condition of the Crank-Nicolson format); ② Dependence on precise boundary conditions: In actual engineering, boundary heat flux density is difficult to measure accurately; ③ Ill-posedness: Small measurement errors in solving inverse problems lead to violent oscillations in the solution, requiring Tikhonov regularization, but parameter selection is difficult.

[0042] (2) Shortcomings of Pure Data-Driven Approaches

[0043] Traditional neural networks require end-to-end learning to achieve temperature field mapping, which has the following problems: 1. Data dependency: Training requires tens of thousands of sets of labeled data (such as those collected by infrared thermal imagers), which is costly; 2. Lack of physical consistency: The prediction results may violate energy conservation (such as local temperature mutations exceeding the melting point of the material).

[0044] (3) Bottlenecks of the Physical Information Neural Network (PINN) method

[0045] ① Difficulty in processing complex boundaries: The sampling efficiency of residual points in unstructured grids (such as surface boundaries) is low; ② Unstable training: The difference in the magnitude of PDE residuals and data loss leads to conflicts in optimization directions.

[0046] Furthermore, the large number and complex distribution of heat-generating components such as CPUs, GPUs, and power modules installed on server motherboards contributes to the complex and variable temperature distribution and significant transient characteristics of server motherboards, due to the large number and complex distribution of heat-generating components such as CPUs, GPUs, and power modules. The heat generation of these components varies dynamically with workload, and the thermal conductivity between PCB layers varies. These factors contribute to the complex and variable temperature distribution of server motherboards, with significant transient characteristics. Related temperature field reconstruction methods are inadequate for such heat source systems with complex multi-physics field coupling and significant transient characteristics. For example, they fail to fully account for dynamic heat sources (such as the instantaneous power changes of CPUs under sudden loads) and rarely consider complex material properties (such as the multi-layer heterogeneous structure of server motherboards and the anisotropic thermal conductivity between PCB layers).

[0047] In view of this, an embodiment of the present invention provides a training method for a temperature field determination model, a temperature field determination method, an apparatus, a device, a medium and a program product, wherein the training method for the temperature field determination model comprises: obtaining a sample temperature data set about a target hardware module, the sample temperature data set comprising a real temperature data set, a simulated temperature data set and a generated temperature data set; wherein the real temperature data set comprises a plurality of first sample temperature data acquired by a plurality of temperature sensors arranged on the target hardware module, and the simulated temperature data set comprises a plurality of second sample temperature data acquired by performing finite element simulation based on a simulated physical model and simulation conditions of the target hardware module. Data, the generated temperature data set includes a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data and the third sample temperature data all include sample time, sample temperature value and sample temperature position corresponding to the sample temperature value; based on the real temperature data set and the simulated temperature data set, the pre-trained source model is transferred and trained to obtain an intermediate model, wherein the source model includes a physical information neural network with a target heat conduction equation constraint; and based on the real temperature data set, the simulated temperature data set and the generated temperature data set, the intermediate model is trained to obtain a trained temperature field determination model.

[0048] Figure 1 A diagram illustrating application scenarios of a temperature field determination model training method, a temperature field determination method, an apparatus, a device, a medium, and a program product according to an embodiment of the present invention is shown.

[0049] like Figure 1 As shown, the application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or optical fiber cables.

[0050] A user may use a first terminal device 101, a second terminal device 102, or a third terminal device 103 to interact with a server 105 via a network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, or the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (for example only).

[0051] The first terminal device 101 , the second terminal device 102 , and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0052] The server 105 may be a server that provides various services, such as a background management server (for example only) that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process received data such as user requests, and feed back processing results (e.g., web pages, information, or data obtained or generated based on user requests) to the terminal devices.

[0053] The server can be a cloud server, also known as a cloud computing server or cloud host. It is a hosting product within the cloud computing service system that addresses the management difficulties and poor scalability of traditional physical hosts and VPS services ("Virtual Private Servers" or "VPS"). The server can also be a distributed system server or a server integrated with blockchain.

[0054] It should be noted that the training method for the temperature field determination model and the temperature field determination method provided in the embodiment of the present invention can generally be executed by the server 105. Accordingly, the training device for the temperature field determination model and the temperature field determination device provided in the embodiment of the present invention can generally be set in the server 105. The training method for the temperature field determination model and the temperature field determination method provided in the embodiment of the present invention can also be executed by a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Accordingly, the training device for the temperature field determination model and the temperature field determination device provided in the embodiment of the present invention can also be set in a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105.

[0055] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0056] Figure 2 A flow chart of a method for training a temperature field determination model according to an embodiment of the present invention is shown.

[0057] like Figure 2 As shown, the method 200 includes operations S210 to S230.

[0058] In operation S210, a sample temperature data set about the target hardware module is obtained, and the sample temperature data set includes a real temperature data set, a simulated temperature data set and a generated temperature data set; wherein, the real temperature data set includes a plurality of first sample temperature data collected by a plurality of temperature sensors arranged in the target hardware module, the simulated temperature data set includes a plurality of second sample temperature data obtained by finite element simulation based on the simulation physical model and simulation conditions of the target hardware module, and the generated temperature data set includes a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data and the third sample temperature data all include a sample time, a sample temperature value and a sample temperature position corresponding to the sample temperature value.

[0059] In operation S220 , transfer learning training is performed on the pre-trained source model based on the real temperature data set and the simulated temperature data set to obtain an intermediate model, wherein the source model includes a physical information neural network with target heat conduction equation constraints.

[0060] In operation S230 , the intermediate model is trained based on the real temperature data set, the simulated temperature data set, and the generated temperature data set to obtain a trained temperature field determination model.

[0061] According to one embodiment of the present invention, the target hardware module includes at least one dynamic heat source and an anisotropic heat conductive structure thermally coupled to the at least one dynamic heat source, wherein the heat generation condition of the dynamic heat source changes in response to a workload state of the target hardware module.

[0062] In one embodiment, the target hardware module may include a server motherboard assembly, the dynamic heat source may include, for example, a CPU, GPU, power module, etc. mounted on the server motherboard, and the anisotropic heat conductive structure may include, for example, the server motherboard. The temperature field determination model may be used to determine the temperature field of the server motherboard assembly. For example, a sample temperature data set for the server motherboard assembly may be obtained, and the sample temperature data set may include a real temperature data set, a simulated temperature data set, and a generated temperature data set.

[0063] In one example, the real temperature data set includes a plurality of first sample temperature data collected by a plurality of temperature sensors arranged on the server motherboard. The first sample temperature data includes a sample time, a sample temperature value, and a sample temperature position corresponding to the sample temperature value. For example, n temperature sensors (such as thermocouples and thermal resistors) can be arranged at a plurality of key positions of the server motherboard (such as the surface of the CPU chip, near the memory chip, the edge of the PCB board, etc.) to obtain the temperature data of the server motherboard components during actual operation, and obtain a plurality of first sample temperature data. ,in, The sample time (e.g., the time from the server startup) sampling moments), is the sample temperature position (e.g. The location coordinates of the temperature sensors ), is the sample temperature value at the sample temperature position at the sample time.

[0064] In one example, the simulated temperature data set includes a plurality of second sample temperature data obtained by finite element simulation based on the simulated physical model and simulation conditions of the target hardware module. The second sample temperature data includes a sample time, a sample temperature value, and a sample temperature position corresponding to the sample temperature value. For example, a three-dimensional finite element model of a server motherboard assembly can be constructed, and the transient temperature field of the server motherboard assembly can be solved using finite element simulation software under set simulation conditions to obtain a plurality of second sample temperature data. ,in, is the sample moment (e.g., represents the p-th time step), is the sample temperature position (such as the position coordinate of the qth finite element ), is the sample temperature value at the sample temperature position at the sample time. It should be noted that the finite element method solves the transient temperature by discretizing the solution domain into a finite number of finite elements. Although the calculation result is discrete, the temperature value at any position at any time can be obtained through interpolation, thereby obtaining multiple second sample temperature data.

[0065] In one example, the generated temperature data set includes a plurality of third sample temperature data generated by a target-based generative adversarial network, wherein the third sample temperature data includes a sample time, a sample temperature value, and a sample temperature position corresponding to the sample temperature value. For example, a generative adversarial network can be trained based on temperature data about a server motherboard component to obtain a target generative adversarial network. The plurality of third sample temperature data can be generated using a generator of the target generative adversarial network. ,in, is the sample time, is the sample temperature position (such as the position coordinates ), is the sample temperature value at the sample temperature position at the sample time. Optionally, physical constraints can be embedded in the training process of the target generative adversarial network to ensure that the temperature field data generated by the generator satisfies the laws of physics.

[0066] In one embodiment, the pre-trained source model can be trained through transfer learning based on the real temperature data set and the simulated temperature data set to obtain an intermediate model. The purpose of transfer learning is to transfer knowledge from the source model in the source domain (such as data-rich and complete annotations) to the target model (such as the temperature field determination model) in the target domain (such as data-scarce and insufficiently annotated), thereby reducing the amount of training data and computing resources required for the target model, thereby significantly improving the learning efficiency and model performance of the target model. As an example, the target task can be, for example, to determine the transient temperature field of the target hardware module, and the pre-trained source model can be obtained by training using the annotated data of the source domain. For example, a source hardware structure that is similar or close to the structure of the target hardware module can be selected as the source domain, and the source model can be trained using the temperature data of the source domain to obtain a pre-trained source model.

[0067] In one embodiment, the model structure of the source model can be identical or similar to that of the temperature field determination model to better adapt to the target task. For example, the temperature field determination model can include a physical information neural network (PINN) constrained by a target heat conduction equation to ensure that the temperature field output by the temperature field determination model satisfies physical laws such as heat conduction, thereby ensuring the model's physical rationality and prediction accuracy. Accordingly, the source model also includes a physical information neural network constrained by a target heat conduction equation. As an example, the target heat conduction equation can be selected as the anisotropic transient heat conduction governing equation to better characterize complex heat source systems such as server motherboard components, which have complex multi-physics coupling and significant transient characteristics.

[0068] For example, the governing equation for anisotropic transient heat conduction can be expressed as follows:

[0069] (1)

[0070] In formula (1), is the material density , is the specific heat capacity , is the thermal conductivity , Heat source term , is the Hamiltonian operator, in three-dimensional space , used to calculate gradient and divergence. is the temperature gradient .

[0071] In one embodiment, the intermediate model may be trained based on the real temperature data set, the simulated temperature data set, and the generated temperature data set to obtain a trained temperature field determination model.

[0072] According to embodiments of the present invention, a real temperature data set can be obtained by collecting temperature data through a temperature sensor. A simulated temperature data set can be obtained by solving the transient temperature field using the finite element method (FEM). A generated temperature data set can also be generated based on a generator in a target-based generative adversarial network. The real temperature data set and the simulated temperature data set ensure the physical plausibility of the sample data, and the generated temperature data set can effectively expand the number of sample data and reduce data dependency.

[0073] According to an embodiment of the present invention, a pre-trained source model can be trained through transfer learning based on a real temperature data set and a simulated temperature data set to obtain an intermediate model. Transfer learning is beneficial to improving the performance and training efficiency of the model. On this basis, the intermediate model can be trained based on the real temperature data set, the simulated temperature data set and the generated temperature data set to obtain a temperature field determination model. Since the sample temperature data in the real temperature data set, the simulated temperature data set and the generated temperature data set all include sample time, sample temperature value and sample temperature position, the temperature field determination model obtained by training with the aforementioned sample data can output a transient temperature field. Furthermore, since the temperature field determination model includes a physical information neural network with anisotropic transient heat conduction control equations, it can be ensured that the transient temperature field output by the temperature field determination model can satisfy physical laws such as heat conduction, and can be well applied to complex heat source systems with complex multi-physical field coupling and significant transient characteristics, thereby effectively improving the physical rationality, prediction accuracy, real-time performance, and dynamic heat source adaptability of the model.

[0074] It can be understood that the training method for the temperature field determination model provided by the embodiment of the present invention innovatively combines the finite element method (FEM), physical information neural network (PINN), transfer learning, and generative adversarial network (GAN). The trained temperature field determination model can be applied to dynamic thermal monitoring and fault warning of high-density server motherboard components. It can effectively solve the problem of high-precision real-time reconstruction of the temperature field in complex geometry, transient load and data sparse scenarios, and can achieve high-precision temperature field prediction under sparse sensor data, complex geometric boundaries and dynamic transient conditions. The beneficial effects of the embodiment of the present invention may include:

[0075] (1) Applicable to complex heat source systems with complex multi-physical field coupling and significant transient characteristics: All sample data of the model include sample time, sample temperature value and sample temperature position, and the model uses anisotropic transient heat conduction control equation as physical constraint, so that the trained temperature field determination model can be better applied to heat conduction scenarios with transient and anisotropic characteristics.

[0076] (2) Reduce data dependence: By utilizing the knowledge of the source domain through transfer learning and combining it with the target adversarial generative network to generate the generated temperature data set, the demand for a large amount of labeled data in the target domain is significantly reduced, thus reducing the data collection cost.

[0077] (3) Improve reconstruction accuracy, real-time performance, and dynamic adaptability: By adopting the anisotropic transient heat conduction control equation as a physical constraint, the reconstructed temperature field is ensured to satisfy the physical laws of heat conduction and transient, anisotropic heat conduction scenarios, thereby improving the physical rationality and prediction accuracy of the model.

[0078] In an optional embodiment, the training method for the temperature field determination model provided in the embodiments of the present invention can be applied to liquid-cooled server motherboard assemblies to provide accurate temperature field reconstruction support for the thermal management of liquid-cooled servers. For example, for liquid-cooled server motherboard assemblies, the impact of coolant flow on the temperature field can be considered. By setting appropriate three-dimensional simulation models, simulation conditions, physical constraint equations, etc., a sample data set for the liquid-cooled server motherboard assembly can be constructed, and model training can then be performed based on the sample data set.

[0079] In another optional embodiment, the training method for the temperature field determination model provided in the embodiments of the present invention can also be applied to a data center comprising multiple servers, enabling overall thermal status monitoring and optimization control of the data center, thereby reducing the data center's energy consumption. For example, for a data center, the temperature field distribution of multiple servers and their mutual influence can be comprehensively considered, combined with factors such as the data center's airflow organization and the operation of the cooling system. By setting appropriate three-dimensional simulation models, simulation conditions, and physical constraint equations, a sample data set for the data center can be constructed, and model training can then be performed based on the sample data set.

[0080] According to an embodiment of the present invention, the simulation conditions include material property information, specification property information, heat source distribution information, ambient temperature information and workload status information of the target hardware module; the simulated temperature data set is determined by the following operations: based on the simulation physical model and simulation conditions, multiple second sample temperature data are solved according to the target heat conduction equation by the finite element method to obtain the simulated temperature data set.

[0081] In one embodiment, the target hardware module may be, for example, a server motherboard assembly equipped with a CPU, a GPU, a power supply, etc. A three-dimensional simulation model of the server motherboard assembly may be constructed, and based on the three-dimensional simulation model of the server motherboard assembly, finite element simulation software may be used to solve the heat conduction equation (such as equation (1) above) according to simulation conditions to output a transient temperature field.

[0082] For example, the simulation conditions may include: material properties of each component in the server motherboard assembly, such as the thermal conductivity of the CPU chip , thermal conductivity of PCB board Geometric dimensions, such as the length, width, and height of the chip; boundary conditions (such as the convection heat transfer coefficient of the air around the motherboard components, ambient temperature; heat source distribution, such as the heat rate of the CPU 、Heat rate of memory chips etc.; working conditions, such as CPU load of 20%, 50%, 80%, 100%; ambient temperature, such as 25℃, 30℃, 35℃, 40℃, etc. Among them, the temperature field data simulated and output by the finite element software covers the entire spatial area of ​​the server motherboard component, and the sampling time interval of the temperature value is .

[0083] According to an embodiment of the present invention, a target generative adversarial network includes a generator and a discriminator, and the target generative adversarial network is trained using the following operations: random sample noise is input into a candidate generator, and a sample generation temperature data set is output; a sample reference temperature data set or a sample generation temperature data set is input into a candidate discriminator, and a discrimination result is output; wherein, the sample reference temperature includes a real temperature data set and a simulated temperature data set, and the discrimination result represents the probability that the data input into the candidate discriminator is the sample reference temperature data set; the candidate generator and the candidate discriminator are trained independently and alternately until the end condition is met, thereby obtaining a generator and a discriminator; wherein, the loss function of the generator includes an adversarial loss term determined based on the discrimination result and a first physical constraint loss term determined based on the target heat conduction equation.

[0084] According to one embodiment of the present invention, the target generative adversarial network may include a generator and the discriminator As an example, the generator A convolutional neural network can be used, whose input is random noise , the output is the generated temperature field data Discriminator A convolutional neural network can be used, whose input is the real temperature field data (Including real temperature data set and simulated temperature data sets ) or generate data , the output is the probability of judging whether the data input to the discriminator is real data or .

[0085] The training goal of the target generative adversarial network is to make the data generated by the generator as close as possible to the real data distribution, and the discriminator can accurately distinguish between real data and generated data. During the training process, the generator and the discriminator can be trained alternately. The goal of the generator is to minimize , the discriminator's goal is to maximize At the same time, in order to ensure the physical rationality of the generated data, physical constraint loss can be introduced to the training of the generator, that is:

[0086] (2)

[0087] In formula (2), is the loss function of the generator, To combat the loss term, is the physical constraint weight coefficient, is the first physical constraint loss.

[0088] The calculation method is:

[0089] (3)

[0090] In formula (3), For the The sample temperature value of the generated data sample (that is, the third temperature data) is the generator Output result.

[0091] The calculation method is:

[0092] (4)

[0093] In formula (4), is the expectation operator; is the real data distribution; is the noise distribution. By minimizing , the generator and the discriminator are continuously trained in adversarial fashion, so that the generator can generate more realistic temperature field data.

[0094] Figure 3 A schematic diagram showing the principles of a target generative adversarial network according to an embodiment of the present invention is shown.

[0095] like Figure 3As shown, for example, the network structures of the candidate generator and the candidate discriminator can be designed, such as both being designed as convolutional neural networks. A real temperature data set and a simulated temperature data set can be prepared as sample real temperature field data. Random sample noise can be input into the candidate generator, and the sample generated temperature field data can be output. The sample real temperature field data or the sample generated temperature field data can be input into the candidate discriminator, and the probability that the input data is real data can be output. The candidate generator and the candidate discriminator can be trained alternately until the end condition is met to obtain the generator and the discriminator. Among them, the loss function of the generator includes an adversarial loss term and a first physical constraint loss term. After the training is completed, a large amount of synthetic data (i.e., a generated temperature data set) similar to the temperature field distribution of the real server motherboard components can be generated based on the generator to expand the training data set.

[0096] According to an embodiment of the present invention, transfer learning training is performed on a pre-trained source model based on a real temperature data set and a simulated temperature data set to obtain an intermediate model, including: migrating source model parameters of the source model to a candidate intermediate model as model parameters of the candidate intermediate model; based on the real temperature data set and the simulated temperature data set, transfer learning training is performed on the candidate intermediate model to obtain an intermediate model; wherein the loss function of the transfer learning training includes a source loss term and a target loss term; the target loss term includes a first data fitting loss term determined based on the real temperature data set and the simulated temperature data set, and a second physical constraint loss term determined based on the target heat conduction equation, and the first data fitting loss term includes a first sub-loss term determined based on the real temperature data set and a second sub-loss term determined based on the simulated temperature data set.

[0097] According to one embodiment of the present invention, a source server motherboard component with a structure similar or close to that of the target server motherboard component can be selected as the source domain, and the temperature data of the source domain can be used to train the source model to obtain a pre-trained source model. The pre-trained source model can be subjected to transfer learning training based on the real temperature data set and the simulated temperature data set to obtain an intermediate model. Exemplarily, the network structure of the source model and the intermediate model can both be selected as a physical information neural network (PINN). The structure of the PINN includes an input layer, a hidden layer, and an output layer, and the physical constraint equation is selected as the formula (1) above. The PINN is based on spatial coordinates. and time is the input and the output is the temperature value .

[0098] In one embodiment, the model parameters of the source model can be transferred to the candidate intermediate model for the target domain as the initial model parameters of the candidate intermediate model. The candidate intermediate model can be fine-tuned using the real temperature data set and the simulated temperature data set to obtain the intermediate model. It includes the source model loss and the target domain loss, namely:

[0099] (5)

[0100] In formula (5), is the loss function of transfer learning, is the loss of the source model in the source domain (i.e., the source loss term), is the loss of the intermediate model in the target domain (i.e., the target loss term), which includes data fitting loss and physical constraint loss. is the transfer learning weight coefficient.

[0101] Taking the intermediate model as PINN, when training PINN, the loss function considers both physical equation constraints and data fitting loss. As an example, the loss function of PINN Including data fitting loss (also known as the first data fitting loss term) and physical constraint loss (Also the second physical constraint loss term):

[0102] (5)

[0103] In formula (5), and is a weight coefficient used to balance the importance of data fitting and physical constraints.

[0104] Data fitting loss It includes a first sub-loss item determined based on the sensor temperature data and a second sub-loss item determined based on the simulated temperature data.

[0105] The first sub-loss term is calculated as:

[0106] (6)

[0107] In formula (6), is the number of sensor data samples; is the output of the PINN model Temperature prediction values; For the The actual temperature value measured by the sensor.

[0108] The second sub-loss term is calculated as:

[0109] (7)

[0110] In formula (7), is the number of simulated data samples; is the output of the PINN model Temperature prediction values; For the simulated data values.

[0111] Data fitting loss The calculation method is:

[0112] (8)

[0113] In formula (8), , which represents the sum of the number of simulation data samples and the number of sensor data samples.

[0114] Physical constraint loss The calculation method is:

[0115] (9)

[0116] In formula (9), is the number of physical constraint samples; For the The temperature values ​​predicted by PINN are used to calculate the physical constraint loss.

[0117] Figure 4A A schematic diagram of transfer learning according to an embodiment of the present invention is shown.

[0118] like Figure 4A As shown, a source server motherboard with similar components to the target server motherboard can be selected as the source domain. Training can be performed on the source domain to obtain a pre-trained source model. The model parameters of the source model can be transferred to a candidate intermediate model, serving as the initial model parameters for the candidate intermediate model. Transfer learning training can be performed on the candidate intermediate model based on a set of real temperature data and a set of simulated temperature data from the target domain. For example, a target loss function for transfer learning can be defined that includes the loss of the source model in the source domain and the loss of the intermediate model in the target domain. The candidate intermediate model can then be trained based on the transfer learning loss to obtain an intermediate model.

[0119] Figure 4B A schematic diagram of training a physical information neural network according to an embodiment of the present invention is shown.

[0120] like Figure 4B As shown in Figure 1, when training a PINN, both the physical equation constraint loss and the data fitting loss are considered. For example, the data fitting loss can be calculated based on a set of real and simulated temperature data. The physical constraint loss can be calculated based on the physical equation constraints defined by the PINN network. A PINN loss function can be constructed to calculate the total loss based on the data fitting loss and the physical equation constraint loss, allowing the PINN network to be trained based on this total loss.

[0121] According to an embodiment of the present invention, the loss function of the temperature field determination model includes a second data fitting loss term determined based on the generated temperature data set, as well as a first data fitting loss term and a second physical constraint loss term.

[0122] According to one embodiment of the present invention, the real temperature data set, the simulated temperature data set, and the generated temperature data set can be used as sample data and input into the intermediate model after transfer learning for training to obtain a temperature field determination model (hereinafter referred to as the target model). The model structure of the target model remains unchanged as PINN, and continues to use spatial coordinates. and time (or operating parameters) as input and output the predicted value of temperature field.

[0123] In one embodiment, the real temperature data set and the simulated temperature data set can be input into the intermediate model for training, and the first data fitting loss term can be calculated based on formula (8), and the second physical constraint loss term can be calculated based on formula (9). The generated temperature data set can be input into the intermediate model for training, and the generated data fitting loss term can be calculated. (This is also known as the second data fitting loss term.) After receiving the generated data as input, the target model calculates using its own network structure and parameters and outputs the predicted temperature.

[0124] Loss function of the temperature field determination model The calculation method is:

[0125] (10)

[0126] In formula (10), is the fitting loss of the target model to the generated data, is the loss of the intermediate model in the target domain (i.e., the target loss term).

[0127] The calculation formula is:

[0128] (11)

[0129] In formula (11), is the number of generated data samples used to calculate the fitting loss, that is, the total number of temperature field data samples generated by the Generative Adversarial Network (GAN) generator. For the target model The predicted temperature values ​​for the generated data samples. For the The true temperature value of the generated data sample (the "true" here refers to the temperature value generated by the generator, which is the object of the target model fitting during training) is the temperature field data generated by the GAN generator based on the input random noise.

[0130] According to an embodiment of the present invention, in the training of the target model, the losses of the target model for real data and generated data can be combined through the designed loss function, and the model's fitting ability and generalization performance for different data can be improved by alternately optimizing the target model and the generator.

[0131] In an optional embodiment, the root mean square error (RMSE) and the mean absolute error (MAE) may be used as evaluation indicators to evaluate the temperature field reconstruction effect of the trained temperature field determination model.

[0132] The formula for calculating the root mean square error is:

[0133] (12)

[0134] In formula (12), To verify the number of samples, is the actual temperature value.

[0135] The mean absolute error is calculated as:

[0136] (13)

[0137] Figure 5 A flow chart of a temperature field determination method according to an embodiment of the present invention is shown.

[0138] like Figure 5 As shown, the method 500 includes operations S510 to S520.

[0139] In operation S510 , historical temperature data of a target hardware module is acquired. The historical temperature data includes historical collection time, historical temperature values, and collection positions. The collection positions represent the placement position of the temperature sensor relative to the target hardware module.

[0140] In operation S520, the target time and historical temperature data are input into a temperature field determination model, which outputs a target temperature data set that characterizes the temperature field of the target hardware module at the target time. The temperature field determination model is trained using a temperature field determination model training method according to an embodiment of the present invention.

[0141] In one embodiment, the target hardware module may include a server motherboard component. The historical temperature data may include temperature data collected by multiple temperature sensors arranged on the server motherboard. For example, the historical temperature data may include historical temperature values. , historical collection time and collection location (such as the spatial coordinates of a temperature sensor).

[0142] For example, the historical temperature data [ , ] Input temperature field determination model and output reconstructed temperature field distribution .

[0143] The temperature field determination model of the embodiment of the present invention is obtained by training using the training method of the temperature field determination model according to the embodiment of the present invention. The model can be well applied to complex heat source systems with complex multi-physical field coupling and significant transient characteristics. It has high physical rationality, prediction accuracy, real-time performance, and dynamic heat source adaptability.

[0144] It can be understood that the temperature field determination model provided by the embodiments of the present invention addresses the core challenges of server motherboard component temperature field reconstruction, such as complex multi-physics field coupling, low computational efficiency, poor real-time performance, and strong data dependence. It can at least partially overcome the following technical problems:

[0145] (1) The high computational cost and insufficient real-time performance of the traditional finite element method (FEM): The multi-layer heterogeneous structure of the server motherboard components (such as PCB, copper traces, and heat sinks) requires fine mesh division, which results in a limited time step for the FEM transient solution and cannot meet the millisecond-level real-time monitoring requirements.

[0146] (2) Lack of physical consistency and data dependence of pure data-driven models: Traditional neural networks (such as CNN and LSTM) ignore the heat conduction equation, and the prediction results may violate the conservation of energy (such as the local temperature exceeds the melting point of the material), and require tens of thousands of sets of labeled data (infrared thermal imagers are expensive to collect).

[0147] (3) Training instability of the single PINN method under complex geometry and boundary conditions: The complex geometry of the server motherboard components, such as the curved heat sink and tiny vias, leads to low efficiency of PINN PDE residual point sampling, and the anisotropic thermal conductivity characteristics exacerbate the optimization difficulty.

[0148] (4) Prediction distortion of transient temperature field under dynamic load: Sudden changes in server chip power consumption (such as cloud computing task scheduling) cause rapid changes in the temperature field, and traditional methods are difficult to capture transient characteristics.

[0149] Figure 6 A structural block diagram of a training device for a temperature field determination model according to an embodiment of the present invention is shown.

[0150] like Figure 6 As shown, the apparatus 600 includes a first acquisition module 610 , a first training module 620 and a second training module 630 .

[0151] The first acquisition module 610 is used to obtain a sample temperature data set about the target hardware module, and the sample temperature data set includes a real temperature data set, a simulated temperature data set and a generated temperature data set; wherein, the real temperature data set includes a plurality of first sample temperature data collected by a plurality of temperature sensors arranged in the target hardware module, the simulated temperature data set includes a plurality of second sample temperature data obtained by finite element simulation based on the simulation physical model and simulation conditions of the target hardware module, and the generated temperature data set includes a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data and the third sample temperature data all include a sample time, a sample temperature value and a sample temperature position corresponding to the sample temperature value.

[0152] The first training module 620 is used to perform transfer learning training on the pre-trained source model based on the real temperature data set and the simulated temperature data set to obtain an intermediate model, wherein the source model includes a physical information neural network with target heat conduction equation constraints.

[0153] The second training module 630 is used to train the intermediate model based on the real temperature data set, the simulated temperature data set and the generated temperature data set to obtain a trained temperature field determination model.

[0154] According to an embodiment of the present invention, the simulation conditions include material property information, specification property information, heat source distribution information, ambient temperature information and workload status information of the target hardware module; the simulated temperature data set is determined by the following operations: based on the simulation physical model and simulation conditions, multiple second sample temperature data are solved according to the target heat conduction equation by the finite element method to obtain the simulated temperature data set.

[0155] According to an embodiment of the present invention, a target generative adversarial network includes a generator and a discriminator, and the target generative adversarial network is trained using the following operations: random sample noise is input into a candidate generator, and a sample generation temperature data set is output; a sample reference temperature data set or a sample generation temperature data set is input into a candidate discriminator, and a discrimination result is output; wherein, the sample reference temperature includes a real temperature data set and a simulated temperature data set, and the discrimination result represents the probability that the data input into the candidate discriminator is the sample reference temperature data set; the candidate generator and the candidate discriminator are trained independently and alternately until the end condition is met, thereby obtaining a generator and a discriminator; wherein, the loss function of the generator includes an adversarial loss term determined based on the discrimination result and a first physical constraint loss term determined based on the target heat conduction equation.

[0156] According to an embodiment of the present invention, the first training module 620 may include a migration submodule and a first training submodule.

[0157] The migration submodule is used to migrate the source model parameters of the source model to the candidate intermediate model as the model parameters of the candidate intermediate model.

[0158] The first training submodule is used to perform transfer learning training on the candidate intermediate model based on the real temperature data set and the simulated temperature data set to obtain the intermediate model.

[0159] Among them, the loss function of transfer learning training includes a source loss term and a target loss term; the target loss term includes a first data fitting loss term determined based on the real temperature data set and the simulated temperature data set, and a second physical constraint loss term determined based on the target heat conduction equation. The first data fitting loss term includes a first sub-loss term determined based on the real temperature data set and a second sub-loss term determined based on the simulated temperature data set.

[0160] According to an embodiment of the present invention, the loss function of the temperature field determination model includes a second data fitting loss term determined based on the generated temperature data set, as well as a first data fitting loss term and a second physical constraint loss term.

[0161] According to embodiments of the present invention, any multiple modules among the first acquisition module 610, the first training module 620, and the second training module 630 may be combined into a single module, or any one of these modules may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in a single module. According to embodiments of the present invention, at least one of the first acquisition module 610, the first training module 620, and the second training module 630 may be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or may be implemented in hardware or firmware through any other reasonable means of circuit integration or packaging, or may be implemented in any one of the three implementation methods of software, hardware, and firmware, or any suitable combination of any of these. Alternatively, at least one of the first acquisition module 610, the first training module 620, and the second training module 630 may be at least partially implemented as a computer program module that, when executed, performs the corresponding functionality.

[0162] Figure 7 A structural block diagram of a temperature field determination device according to an embodiment of the present invention is shown.

[0163] like Figure 7 As shown, the device 700 includes a second acquisition module 710 and a prediction module 720 .

[0164] The second acquisition module 710 is used to acquire historical temperature data of the target hardware module. The historical temperature data includes historical acquisition time, historical temperature value and acquisition position. The acquisition position represents the layout position of the temperature sensor relative to the target hardware module.

[0165] The prediction module 720 is used to input the target time and historical temperature data into the temperature field determination model, and output a target temperature data set, which is used to characterize the temperature field of the target hardware module at the target time.

[0166] The temperature field determination model is obtained by training using the training method for the temperature field determination model according to an embodiment of the present invention.

[0167] According to embodiments of the present invention, any multiple modules in the second acquisition module 710 and the prediction module 720 can be combined into a single module, or any one of them can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in a single module. According to embodiments of the present invention, at least one of the second acquisition module 710 and the prediction module 720 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented in hardware or firmware through any other reasonable means of circuit integration or packaging, or can be implemented in any one of the three implementation methods of software, hardware, and firmware, or any suitable combination of any of these. Alternatively, at least one of the second acquisition module 710 and the prediction module 720 can be at least partially implemented as a computer program module that, when executed, performs the corresponding functionality.

[0168] Figure 8 The block diagram schematically shows an electronic device suitable for implementing a training method for a temperature field determination model and a temperature field determination method according to an embodiment of the present invention.

[0169] like Figure 8As shown, an electronic device 800 according to an embodiment of the present invention includes a processor 801, which can perform various appropriate actions and processes based on programs stored in a read-only memory (ROM) 802 or programs loaded from a storage unit 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or related chipsets and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0170] Various programs and data required for the operation of the electronic device 800 are stored in the RAM 803. The processor 801, ROM 802, and RAM 803 are connected to each other via a bus 804. The processor 801 executes the programs in the ROM 802 and / or RAM 803 to perform various operations according to the method flow of the embodiment of the present invention. It should be noted that the programs may also be stored in one or more memories other than the ROM 802 and RAM 803. The processor 801 may also execute the programs stored in the one or more memories to perform various operations according to the method flow of the embodiment of the present invention.

[0171] According to an embodiment of the present invention, electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to bus 804. Electronic device 800 may also include one or more of the following components connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including devices such as a cathode ray tube (CRT), liquid crystal display (LCD), and speakers; a storage section 808 including a hard disk; and a communication section 809 including a network interface card such as a LAN card or modem. Communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. Removable media 811, such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed in drive 810 as needed, so that computer programs read from the removable media can be installed into storage section 808 as needed.

[0172] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments, or may exist independently and not incorporated into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0173] According to an embodiment of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, and may include, for example, but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, a computer-readable storage medium may include the ROM 802 and / or RAM 803 described above, and / or one or more memories other than ROM 802 and RAM 803.

[0174] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for executing the method shown in the flowchart. When the computer program product is executed in a computer system, the program code is configured to cause the computer system to implement the temperature field determination model training method and temperature field determination method provided in embodiments of the present invention.

[0175] The computer program executes the above functions defined in the system / device of the embodiment of the present invention when executed by the processor 801. According to the embodiment of the present invention, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0176] In one embodiment, the computer program may be stored on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may be transmitted and distributed in the form of a signal on a network medium, downloaded and installed via the communication portion 809, and / or installed from a removable medium 811. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to wireless, wired, or any suitable combination thereof.

[0177] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 809 and / or installed from a removable medium 811. When the computer program is executed by the processor 801, the above-described functions defined in the system of the embodiment of the present invention are performed. According to the embodiment of the present invention, the systems, devices, means, modules, units, etc. described above can be implemented by computer program modules.

[0178] According to an embodiment of the present invention, the program code for executing the computer program provided by the embodiment of the present invention can be written in any combination of one or more programming languages. Specifically, these computer programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C" or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect via the Internet).

[0179] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the above-mentioned module, program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0180] It will be understood by those skilled in the art that the features described in the various embodiments of the present invention may be combined and / or coupled in various ways, even if such combinations or couplings are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention may be combined and / or coupled in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or couplings fall within the scope of the present invention.

[0181] The above describes embodiments of the present invention. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. Although each embodiment has been described separately above, this does not mean that the measures in each embodiment cannot be advantageously used in combination. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A training method for a temperature field determination model, characterized in that: The method comprises: Acquire a sample temperature data set about a target hardware module, the sample temperature data set including a real temperature data set, a simulated temperature data set, and a generated temperature data set; wherein the real temperature data set includes a plurality of first sample temperature data acquired by a plurality of temperature sensors disposed on the target hardware module, the simulated temperature data set includes a plurality of second sample temperature data acquired by finite element simulation based on a simulated physical model and simulation conditions of the target hardware module, and the generated temperature data set includes a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data, and the third sample temperature data each include a sample time, a sample temperature value, and a sample temperature position corresponding to the sample temperature value; Based on the real temperature data set and the simulated temperature data set, performing transfer learning training on a pre-trained source model to obtain an intermediate model, wherein the source model includes a physical information neural network with target heat conduction equation constraints; and The intermediate model is trained based on the real temperature data set, the simulated temperature data set and the generated temperature data set to obtain a trained temperature field determination model.

2. The method according to claim 1, characterized in that The simulation conditions include material property information, specification property information, heat source distribution information, ambient temperature information, and workload status information of the target hardware module; the simulation temperature data set is determined by the following operations: Based on the simulation physical model and the simulation conditions, a plurality of second sample temperature data are solved according to the target heat conduction equation by using a finite element method to obtain the simulated temperature data set.

3. The method according to claim 1, characterized in that The target generative adversarial network includes a generator and a discriminator, and the target generative adversarial network is trained using the following operations: Input random sample noise into the candidate generator and output the sample to generate the temperature data set; Inputting a sample reference temperature data set or the sample generated temperature data set into a candidate discriminator and outputting a discrimination result; wherein the sample reference temperature includes the real temperature data set and the simulated temperature data set, and the discrimination result represents the probability that the data input to the candidate discriminator is the sample reference temperature data set; The candidate generator and the candidate discriminator are trained alternately and independently until an end condition is met, thereby obtaining the generator and the discriminator; wherein the loss function of the generator includes an adversarial loss term determined based on the discrimination result and a first physical constraint loss term determined based on the target heat conduction equation.

4. The method according to claim 1, wherein The step of performing transfer learning training on the pre-trained source model based on the real temperature data set and the simulated temperature data set to obtain an intermediate model includes: Migrating source model parameters of the source model to the candidate intermediate model as model parameters of the candidate intermediate model; Based on the real temperature data set and the simulated temperature data set, performing transfer learning training on the candidate intermediate model to obtain the intermediate model; Among them, the loss function of the transfer learning training includes a source loss term and a target loss term; the target loss term includes a first data fitting loss term determined based on the real temperature data set and the simulated temperature data set, and a second physical constraint loss term determined based on the target heat conduction equation, and the first data fitting loss term includes a first sub-loss term determined based on the real temperature data set and a second sub-loss term determined based on the simulated temperature data set.

5. The method according to claim 4, characterized in that The loss function of the temperature field determination model includes a second data fitting loss term determined based on the generated temperature data set, as well as the first data fitting loss term and the second physical constraint loss term.

6. A method for determining a temperature field, characterized in that: The method comprises: Acquire a set of historical temperature data of a target hardware module, the historical temperature data including historical collection time, historical temperature value, and collection location, the collection location representing the placement position of a temperature sensor relative to the target hardware module; Inputting the target time and the historical temperature data into a temperature field determination model, and outputting a target temperature data set, wherein the target temperature data set is used to characterize the temperature field of the target hardware module at the target time; The temperature field determination model is trained using the method according to any one of claims 1 to 5.

7. A training device for a temperature field determination model, characterized in that: The device comprises: a first acquisition module, configured to acquire a sample temperature data set about a target hardware module, the sample temperature data set comprising a real temperature data set, a simulated temperature data set, and a generated temperature data set; wherein the real temperature data set comprises a plurality of first sample temperature data acquired by a plurality of temperature sensors disposed on the target hardware module, the simulated temperature data set comprises a plurality of second sample temperature data acquired by finite element simulation based on a simulated physical model and simulation conditions of the target hardware module, and the generated temperature data set comprises a plurality of third sample temperature data generated by a target-based generative adversarial network; the first sample temperature data, the second sample temperature data, and the third sample temperature data each comprise a sample time, a sample temperature value, and a sample temperature position corresponding to the sample temperature value; a first training module, configured to perform transfer learning training on a pre-trained source model based on the real temperature data set and the simulated temperature data set to obtain an intermediate model, wherein the source model includes a physical information neural network with target heat conduction equation constraints; and The second training module is used to train the intermediate model based on the real temperature data set, the simulated temperature data set and the generated temperature data set to obtain a trained temperature field determination model.

8. A temperature field determination device, characterized in that: The device comprises: A second acquisition module is configured to acquire historical temperature data of a target hardware module, wherein the historical temperature data includes a historical acquisition time, a historical temperature value, and an acquisition location, wherein the acquisition location represents a location of a temperature sensor relative to the target hardware module; a prediction module, configured to input a target time and the historical temperature data into a temperature field determination model, and output a target temperature data set, wherein the target temperature data set is used to characterize the temperature field of the target hardware module at the target time; The temperature field determination model is trained using the method according to any one of claims 1 to 5.

9. An electronic device comprising: one or more processors; a memory for storing one or more computer programs, It is characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images