An artificial intelligence-based temperature control method, device, equipment and storage medium

Abstract

This application discloses an artificial intelligence-based temperature control method and apparatus. This application can acquire satellite temperature prediction data and real-time sensor network data; establish a first digital twin based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; adjust the first digital twin to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; generate temperature control commands using the second digital twin based on the temperature prediction data and real-time sensor network data; and perform temperature control processing on the satellite based on the temperature control commands to improve the accuracy of satellite temperature control.

Description

Technical Field

[0001] This application relates to the field of computer technology, and specifically to a temperature control method, device, electronic device, and storage medium based on artificial intelligence. Background Technology

[0002] Currently, in the field of spacecraft thermal control, commonly used thermal control strategies mainly include threshold-based passive response control and ground command control. Threshold-based passive response control primarily uses temperature sensors to monitor the temperature of critical components, triggering corresponding thermal control actions when the temperature exceeds a preset threshold. However, threshold-based passive response control suffers from problems such as response lag and large ground-satellite communication delays. Therefore, there is an urgent need in existing technologies for an autonomous decision-making method for thermal control systems that can autonomously sense, make real-time decisions, execute intelligently, and continuously learn in orbit, and support constellation-level thermal load collaborative scheduling. Summary of the Invention

[0003] This application proposes an artificial intelligence-based temperature control method and apparatus, which can improve the accuracy of thermal management of spacecraft / satellites.

[0004] This application provides an artificial intelligence-based temperature control method, including: Acquire satellite temperature prediction data and real-time sensor network data; A first digital twin is established based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; The first digital twin is adjusted to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; The second digital twin generates temperature control commands based on the temperature prediction data and real-time sensor network data; The satellite is subjected to temperature control processing based on the temperature control command.

[0005] Accordingly, embodiments of this application also provide a temperature control device based on artificial intelligence, including: The acquisition unit is used to acquire satellite temperature prediction data and real-time sensor network data; A building unit is used to build a first digital twin based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; An adjustment unit is used to adjust the first digital twin to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; The generation unit is used to generate temperature control commands based on the temperature prediction data and real-time sensor network data using the second digital twin; The control unit is used to perform temperature control processing on the satellite based on the temperature control command.

[0006] In some embodiments, the adjustment unit may include: The dimensionality reduction subunit is used to perform dimensionality reduction processing on the first digital twin to obtain the third digital twin; The modeling subunit is used to perform physical data fusion modeling processing on the third digital twin using a physical information neural network to obtain a fourth digital twin; The assimilation subunit is used to assimilate the data of the fourth digital twin using ensemble Kalman filtering to obtain the second digital twin.

[0007] In some embodiments, the dimensionality reduction subunit may include: The sampling module is used to perform snapshot sampling on the first digital twin to obtain snapshot data; A construction module is used to construct a low-dimensional space based on the snapshot data; The truncation module is used to perform energy truncation processing on the low-dimensional space to obtain the third digital twin.

[0008] In some embodiments, the generation unit may include: The comparison subunit is used by the second digital twin to compare the temperature prediction data and the real-time sensor network data to obtain a residual vector; The statistical subunit is used to perform statistical analysis on the residual vector to obtain statistical analysis data. The first generation subunit is used to generate a thermal control strategy based on the statistical analysis data; A verification subunit is used to perform safety verification on the thermal control strategy. The second generation subunit is used to generate temperature control commands based on the thermal control strategy when the safety verification of the thermal control strategy passes.

[0009] In some embodiments, the first generating subunit may include: The judgment module is used to determine whether the satellite has a temperature anomaly based on the statistical analysis data; The generation module is used to generate the thermal control strategy based on a preset reward function and a greedy strategy using a reinforcement learning algorithm when an abnormal temperature occurs.

[0010] In some embodiments, the temperature control device may further include: The first acquisition unit is used to acquire the neighbor table of the satellite, the neighbor table including the thermal state data of at least one neighbor satellite of the satellite; The first determining unit is used to determine the non-real-time inference task of the satellite when the satellite's thermal margin data is less than a preset margin threshold. The second determining unit is used to determine the target neighboring satellite based on the thermal state data of the at least one neighboring satellite; A migration unit is used to migrate the non-real-time inference tasks of the satellite to the target neighboring satellite.

[0011] In some embodiments, the temperature control device may further include: The second acquisition unit is used to acquire the neighbor table of the satellite, the neighbor table including the thermal state data of at least one neighbor satellite of the satellite; The balancing unit is used to perform global thermal load balancing processing on the satellite and the at least one neighboring satellite based on the thermal state data of the satellite and the thermal state data of the at least one neighboring satellite.

[0012] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative embodiments described above.

[0013] Accordingly, this application embodiment also provides a storage medium storing instructions, which, when executed by a processor, implement any of the artificial intelligence-based temperature control methods provided in this application embodiment.

[0014] This application embodiment can acquire satellite temperature prediction data and real-time sensor network data; establish a first digital twin based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; adjust the first digital twin to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; generate temperature control commands using the second digital twin based on the temperature prediction data and real-time sensor network data; and perform temperature control processing on the satellite based on the temperature control commands to improve the accuracy of satellite temperature control. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of a scenario for the temperature control method based on artificial intelligence provided in an embodiment of this application; Figure 2 This is a schematic flowchart of the temperature control method based on artificial intelligence provided in an embodiment of this application; Figure 3 This is another scenario illustration of the temperature control method based on artificial intelligence provided in the embodiments of this application; Figure 4 This is another scenario illustration of the temperature control method based on artificial intelligence provided in the embodiments of this application; Figure 5 This is another scenario illustration of the temperature control method based on artificial intelligence provided in the embodiments of this application; Figure 6 This is another scenario illustration of the temperature control method based on artificial intelligence provided in the embodiments of this application; Figure 7 This is another scenario illustration of the temperature control method based on artificial intelligence provided in the embodiments of this application; Figure 8 This is another scenario illustration of the temperature control method based on artificial intelligence provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the temperature control device based on artificial intelligence provided in the embodiments of this application; Figure 10 This is a schematic diagram of the structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. However, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] This application proposes an artificial intelligence-based temperature control method, which can be executed by an artificial intelligence-based temperature control device integrated into a computer device. The computer device may include at least one of a terminal and a server. That is, the artificial intelligence-based temperature control method proposed in this application can be executed by a terminal, a server, or jointly by a terminal and a server capable of communicating with each other.

[0019] The terminal may include, but is not limited to, spacecraft, smartphones, tablets, laptops, personal computers (PCs), smart home appliances, wearable electronic devices, VR / AR devices, in-vehicle terminals, intelligent voice interaction devices, etc.

[0020] A server can be an interconnecting server between multiple heterogeneous systems or a backend server. It can also be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, and big data and artificial intelligence platforms, etc.

[0021] It should be noted that the embodiments of this application can be applied to various scenarios, including but not limited to heat identification, artificial intelligence, and aerospace applications.

[0022] In one embodiment, such as Figure 1 The AI-based temperature control device can be integrated into computer equipment such as terminals or servers to implement the AI-based temperature control method proposed in this application. Specifically, server 11 or terminal 10 can acquire satellite temperature prediction data and real-time sensor network data; establish a first digital twin based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; adjust the first digital twin to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; and generate temperature control commands using the second digital twin based on the temperature prediction data and real-time sensor network data.

[0023] The following will provide a detailed description of each example. It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of the embodiments.

[0024] This application embodiment will be described from the perspective of an artificial intelligence-based temperature control device, which can be integrated into a computer device, such as a server or a terminal.

[0025] like Figure 2 The present invention provides a temperature control method based on artificial intelligence, the specific process of which includes: 101. Acquire satellite temperature prediction data and real-time sensor network data.

[0026] The temperature prediction data is generated by a lightweight thermal control prediction model deployed on an onboard AI chip. This model predicts future short-term temperature field and heat load changes with a delay of less than 1 second, based on the satellite's real-time computing load, historical heat dissipation data, and orbital parameters. The real-time sensor network data includes real-time data collected by temperature sensors, pressure sensors, and flow sensors distributed throughout the satellite.

[0027] In this embodiment, the satellite carries an onboard AI chip (such as the NVIDIA Jetson AGX Orin), on which a lightweight thermal control prediction model is deployed. This model predicts temperature field and heat load changes over the next 5-15 minutes with a delay of less than 1 second, based on the satellite's real-time computational load, historical heat dissipation data, and orbital parameters (including solar radiation angle and Earth's albedo). Simultaneously, real-time sensor network data is collected through temperature sensors (200), pressure sensors (50), and flow sensors (20) distributed throughout the satellite, with a sampling frequency of 10Hz.

[0028] In some embodiments, prior to acquiring satellite temperature prediction data, this application may further include: Obtain a first temperature prediction model, wherein the first temperature prediction model is a full-order model; The first temperature prediction model is subjected to knowledge distillation to obtain the second temperature prediction model. The second temperature prediction model is subjected to quantitative sensing training to obtain the third temperature prediction model; The third temperature prediction model is pruned in a structured manner to obtain a fourth temperature prediction model, wherein the order of the fourth temperature prediction model is lower than that of the first temperature prediction model. The fourth temperature prediction model is used to predict the satellite's load data, historical heat consumption data, and orbital parameters to obtain the satellite's temperature prediction data. Temperature control processing is performed on the satellite based on the satellite's temperature prediction data.

[0029] In this embodiment, the first temperature prediction model is a high-precision ground-based computational fluid dynamics (CFD) full-order model with 10 million grid cells. A single calculation takes approximately 2 hours, and the computational complexity is O(10^25). 9The model is on the order of magnitude of 1000. It can perform high-precision simulations of the satellite's three-dimensional transient temperature field and serves as a teacher model for subsequent knowledge distillation. The ground-based high-precision CFD full-order model can be understood as a complete and unsimplified CFD calculation model established during the ground phase for accurately simulating the satellite's thermal environment.

[0030] The knowledge distillation process uses the first temperature prediction model (full-order model) as the teacher model and the on-board lightweight model as the student model, transferring the knowledge from the teacher model to the student model, thus achieving a significant reduction in model size for the first time.

[0031] The quantization-aware training employs a symmetrical uniform quantization strategy, which quantizes the model weights from high-precision floating-point format to low-precision integer format, thereby compressing the model size and improving inference speed while maintaining prediction accuracy.

[0032] The structured pruning further reduces model complexity and inference latency while maintaining accuracy by removing unimportant channels and connections in the model.

[0033] The satellite's real-time load data (such as the type of computing task and computing power requirements), historical heat consumption data, and orbital parameters (such as solar radiation angle, Earth albedo, etc.) are input into the fourth temperature prediction model. The model outputs prediction data of future short-term temperature field and heat load changes with a delay of less than 1 second.

[0034] The predicted temperature data is used as a feedforward input, combined with real-time sensor feedback, to generate temperature control commands and send them to the onboard thermal control actuators (including liquid-cooled pumps, radiators, thermal switches, etc.) to achieve precise control of the satellite temperature.

[0035] Preferably, the step of performing knowledge distillation on the first temperature prediction model to obtain the second temperature prediction model specifically includes: Obtain a fifth temperature prediction model, wherein the fifth temperature prediction model is a student model of the first temperature prediction model; The first temperature prediction model is used to identify satellite-related training data to obtain the hard label prediction results output by the first temperature prediction model and the soft label probability distribution output by the target layer in the model structure of the first temperature prediction model. Based on a preset distillation compression ratio, the knowledge of the first temperature prediction model is transferred to the fifth temperature prediction model to obtain the sixth temperature prediction model; The sixth temperature prediction model is trained based on the distillation loss function to obtain the second temperature prediction model.

[0036] A distillation loss function is constructed based on the hard label prediction results and the soft label probability distribution.

[0037] The fifth temperature prediction model is a lightweight multilayer perceptron (MLP) on space, which contains three hidden layers with 128 neurons in each layer. Its number of parameters is much smaller than that of the teacher model, making it suitable for deployment on spaceborne AI chips.

[0038] Satellite-related training data (including historical payload data, heat dissipation data, orbital parameters, etc.) are input into the first temperature prediction model to obtain the hard-label prediction results (i.e., the determined temperature prediction values) output by the model. Simultaneously, the soft-label probability distribution output by the model's last softmax layer at a preset distillation temperature is also obtained. This soft-label probability distribution contains the teacher model's "understanding" of various temperature states and information on inter-category similarity, serving as a key carrier for transmitting "dark knowledge" in knowledge distillation.

[0039] The distillation loss function is L = α·L_hard + (1-α)·L_soft, where α = 0.7, L_hard is the cross-entropy loss (based on hard labels), L_soft is the KL divergence loss (based on soft labels), and the distillation temperature T = 4. Through this loss function, the student model simultaneously learns the teacher model's hard-label output (deterministic knowledge) and soft-label probability distribution (uncertain knowledge), thus inheriting the teacher model's predictive ability more completely.

[0040] The distillation compression ratio is greater than 10:1, preferably 15:1. Through knowledge distillation, the knowledge of the teacher model is efficiently transferred to the student model, and the model size is significantly reduced.

[0041] The sixth temperature prediction model is iteratively trained until the distillation loss function converges, resulting in the second temperature prediction model. After knowledge distillation, the second temperature prediction model has a mean absolute error (MAE) of less than 0.5K and a root mean square error (RMSE) of less than 1.0K on the LEO orbital thermal environment simulation dataset.

[0042] Preferably, the step of performing quantitative sensing training on the second temperature prediction model to obtain the third temperature prediction model specifically includes: Based on the symmetric uniform quantization strategy and preset quantization parameters, the model weights of the second temperature prediction model are subjected to format quantization to obtain the seventh temperature prediction model. The seventh temperature prediction model is trained, and during the training process, a pass-through estimator is used to perform gradient approximation on the seventh temperature prediction model to obtain the third temperature prediction model.

[0043] A symmetric uniform quantization strategy was adopted, setting the model weights from FP32 format to INT8 format, and the activation value quantization range to [-10, 10]. The quantization scaling factor was determined adaptively through backpropagation during training. After quantization, the model size was compressed from 32MB to 8MB, achieving a compression ratio of 75%.

[0044] During forward propagation, the quantization parameters are applied to the model weights and activation values ​​to perform INT8 quantization simulation, generating the quantized output. During backpropagation, since the gradient of the quantization operation (Round) is zero or non-existent at most points, a straight-through estimator is used to approximate the back gradient of the quantization function as 1, allowing the gradient to bypass the quantization nodes and be directly transmitted to the parameters of the previous network layer, thus maintaining the continuous flow of gradients. After iterative training, the model parameters gradually adapt to the low-precision quantization representation. On the NVIDIA Jetson AGX Orin onboard AI chip, the INT8 inference latency is reduced from 50ms in FP32 to 25ms, a 50% reduction, while the throughput is increased by 2.1 times.

[0045] Preferably, the step of performing structured pruning on the third temperature prediction model to obtain the fourth temperature prediction model specifically includes: The scaling factor of the structural layer in the third temperature prediction model is subjected to sparsity constraints to obtain the constrained scaling factor. The constrained scaling factor is compared with the preset pruning threshold to obtain a target scaling factor that is less than the preset pruning threshold; Remove the model channel corresponding to the target scaling factor from the third temperature prediction model to obtain the pruned model; fine-tune and train the pruned model to obtain the fourth temperature prediction model.

[0046] An L1 regularized channel pruning method is employed, using the scaling factor of the BN layer in the third temperature prediction model as a metric for channel importance. L1 regularization imposes a sparsity constraint on the scaling factor of the BN layer, causing the scaling factors of unimportant channels to approach zero during training.

[0047] Set a pruning threshold λ=1e-4, and determine channels with scaling factors lower than the threshold as redundant channels.

[0048] Redundant channels and their corresponding connections are removed from the third temperature prediction model, with a pruning rate of no less than 40% and an accuracy loss of no more than 2%. The sparsity of the pruned model is 40%, and fine-tuning is performed to restore accuracy. After pruning, TensorRT sparse inference is used for acceleration, resulting in an inference speed improvement of no less than 35%.

[0049] Through the three-step compression process of knowledge distillation, quantitative perception training, and structured pruning, the fourth temperature prediction model has a size of less than 10MB and a prediction latency of less than 1 second, meeting the requirements for real-time deployment on spacecraft.

[0050] In some embodiments, such as Figure 3 and Figure 4 As shown, this invention deploys a systematically compressed lightweight thermal control prediction model on a spaceborne AI chip. The model is first compressed using knowledge distillation: a high-precision ground-based CFD full-order model (10 million grids, 2 hours computation time) is used as the teacher model, and a lightweight multilayer perceptron on the spacecraft (containing 3 hidden layers, 128 neurons per layer) is used as the student model. The student model simultaneously learns the hard-label output of the teacher model and the soft-label probability distribution of the last softmax layer (distillation temperature T=4). The distillation loss function is L = 0.7·L_hard + 0.3·L_soft (where L_hard is the cross-entropy loss and L_soft is the KL divergence loss). The distillation compression ratio reaches 15:1. The student model's temperature prediction MAE on the LEO orbital thermal environment simulation dataset is 0.3K, and the RMSE is 0.5K. Building upon this foundation, a symmetric uniform quantization strategy was employed for quantization-aware training. Model weights were quantized from FP32 to INT8, and the activation value quantization range was set to [-10, 10]. After quantization, the model size was compressed from 32MB to 8MB (a 75% compression ratio). On the NVIDIA Jetson AGX Orin, the INT8 inference latency decreased from 50ms with FP32 to 25ms (a 50% reduction), resulting in a 2.1x increase in throughput. Further, an L1 regularized channel pruning method was adopted, using the BN layer scaling factor as a measure of channel importance. A pruning threshold of λ=1e-4 was set, achieving a pruning rate of 40% with only a 1.5% accuracy loss. The model sparsity after pruning was 40%, and TensorRT sparse inference acceleration improved the actual inference speed by 35%. After these three compression steps, the model size was less than 10MB, and the prediction latency was less than 1 second, meeting the deployment requirements for real-time on-orbit hot prediction.

[0051] In some embodiments, Table 1 presents a performance comparison of lightweight model compression techniques, as shown below:

[0052] Table 1 This application's embodiments employ a three-step systematic compression process—knowledge distillation, quantized perception training, and structured pruning—to compress a high-precision, full-order ground-based CFD model into a lightweight model smaller than 10MB, achieving a compression ratio of over 15:1. This enables a leapfrog deployment of the full-order model from ground to space. The inference latency of the lightweight model on the spaceborne AI chip is reduced from 50ms on the FP32 to 25ms, a 50% reduction, and the prediction latency is less than 1 second, providing ample feedforward time for active thermal control. While achieving significant compression, the model's temperature prediction MAE on the LEO orbital thermal environment simulation dataset is less than 0.5K, and RMSE is less than 1.0K, with quantization accuracy loss controlled within 2%, balancing lightweight design and high accuracy. Through INT8 quantization inference and sparse inference acceleration, the model throughput is increased by 2.1 times, and the inference speed is increased by over 35%, meeting the computational power requirements for real-time spaceborne inference.

[0053] 102. Establish a first digital twin based on temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin.

[0054] The first digital twin is a full-order digital twin model of satellite thermal-electrical-mechanical multiphysics coupling. Its state space contains all the degrees of freedom of the satellite's three-dimensional transient thermal network model (e.g., 1 million degrees of freedom grid cells), which can completely characterize the overall temperature field distribution and thermal state of the satellite.

[0055] 103. Adjust the first digital twin to obtain the second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin.

[0056] The adjustment process includes dimensionality reduction, physical data fusion modeling, and data assimilation. Through these processes, the high-dimensional full-order digital twin is transformed into a low-dimensional, high-precision, and real-time updated reduced-order digital twin.

[0057] Preferably, the adjustment process of the first digital twin to obtain the second digital twin specifically includes: Step S31: Perform dimensionality reduction processing on the first digital twin to obtain the third digital twin.

[0058] The step further includes: taking a snapshot of the first digital twin to obtain snapshot data; constructing a low-dimensional space based on the snapshot data; and performing energy truncation processing on the low-dimensional space to obtain the third digital twin.

[0059] Specifically, Proper Orthogonal Decomposition (POD) is used to reduce the order of the high-dimensional state space. By performing snapshot sampling on the full-order digital twin, the covariance matrix is ​​constructed and the eigenvalue problem is solved. Principal component basis functions are selected based on the cumulative energy truncation criterion (retaining 99.9% of the system energy), projecting the original high-dimensional space to a low-dimensional principal component space (typically reducing the dimension to 50-200). After the reduction, the computational complexity of a single forward solution is reduced from O(10^6) for the full-order model. 9 The magnitude was reduced to O(10) 6 The computational complexity is reduced by 99.9% to the order of magnitude of the previous version, and the time taken for a single update is less than 10ms, meeting the requirements for real-time updates.

[0060] After order reduction, the computational complexity of a single forward pass is reduced from O(10^6) for the full-order model. 9 The magnitude was reduced to O(10) 6 The computational complexity is reduced by 99.9% to the order of magnitude of the original model. On the onboard NVIDIA Jetson AGX Orin processor, the reduced-order model takes less than 10ms to update, meeting the 10Hz real-time update requirement, while the full-order model takes more than 1 hour to update, making it impossible to run in orbit in real time.

[0061] Step S32: Use a physical information neural network to perform physical data fusion modeling on the third digital twin to obtain a fourth digital twin.

[0062] Specifically, a Physics-Informed Neural Network (PINN) is used for physical-data fusion modeling. PINN employs a multi-layer fully connected network structure, and its loss function consists of a data-driven term and a physical constraint term, where the physical constraint term is the residual of the heat conduction equation (such as the Fourier heat conduction equation). By embedding physical equation constraints, PINN achieves a temperature field reconstruction error of less than 1% under sparse sensor conditions, ensuring the physical consistency of the digital twin.

[0063] In some embodiments, a Physical Information Neural Network (PINN) is used for physical-data fusion modeling. PINN employs a 4-layer fully connected network with 64 neurons per layer and an activation function of tanh. The loss function consists of a data-driven term MSE_data and a physical constraint term MSE_pde, i.e., L_PINN = MSE_data + λ·MSE_pde (λ=0.1), where MSE_pde is the residual of the Fourier heat conduction equation. By embedding physical equation constraints, PINN achieves a temperature field reconstruction error of less than 1% while relying on only 200 sparse temperature measurement points.

[0064] Step S33: Assimilate the data of the fourth digital twin using ensemble Kalman filtering to obtain the second digital twin.

[0065] Specifically, ensemble Kalman filtering (EnKF) is employed for real-time data assimilation. Ensemble prediction of the digital twin's state is performed with a preset ensemble size, fusing real-time data from temperature, pressure, and flow sensors distributed across the entire satellite into the digital twin at a preset frequency. EnKF estimates the state covariance matrix through ensemble prediction, avoiding explicit computation of high-dimensional covariance matrices, with a single assimilation cycle taking less than 5ms.

[0066] In some embodiments, such as Figure 3 and Figure 5 As shown, this invention establishes a reduced-order real-time thermo-electric-mechanical multiphysics coupled digital twin on a satellite. This digital twin first employs intrinsic orthogonal decomposition (POD) to reduce the order of the high-dimensional state space: based on the 1 million-DOF full-order state space of the satellite's three-dimensional transient thermal network model, a covariance matrix is ​​constructed through snapshot sampling, and the eigenvalue problem is solved. Based on the cumulative energy truncation criterion (retaining 99.9% of the system energy), the eigenvectors corresponding to the top 100 largest eigenvalues ​​are selected as principal component basis functions, projecting the original high-dimensional space to a 100-dimensional principal component space. After order reduction, the computational complexity of a single forward solution is reduced from O(10^6) for the full-order model. 9 The magnitude was reduced to O(10) 6The computational complexity is reduced by 99.9% to the order of magnitude of the original model. On the onboard NVIDIA Jetson AGX Orin processor, a single update takes less than 10ms, meeting the 10Hz real-time update requirement. In contrast, a single update of the full-order model takes more than 1 hour, making it impossible to run in orbit in real time. Based on this, a Physical Information Neural Network (PINN) is used for physical-data fusion modeling: PINN uses a 4-layer fully connected network (64 neurons per layer, with tanh activation function). The loss function consists of the data-driven term MSE_data and the physical constraint term MSE_pde, i.e., L_PINN = MSE_data + λ·MSE_pde (λ=0.1), where MSE_pde is the residual of the Fourier heat conduction equation. By embedding physical equation constraints, PINN can achieve a temperature field reconstruction error of less than 1% with only 200 sparse temperature measurement points. Furthermore, ensemble Kalman filtering (EnKF) is employed for real-time data assimilation: ensemble prediction is performed using 50 pairs of digital twin states. Real-time data from 200 temperature sensors, 50 pressure sensors, and 20 flow sensors distributed across the entire satellite are fused into the digital twin at a frequency of 10Hz. EnKF estimates the state covariance matrix through ensemble prediction, avoiding the enormous computational burden of explicitly calculating high-dimensional covariance matrices. A single assimilation operation takes less than 5ms. Through the combination of POD order reduction, PINN physical constraint modeling, and EnKF real-time data assimilation, the digital twin achieves high-precision, low-latency physical-data fusion modeling, providing real-time updated state input for subsequent autonomous decision-making.

[0067] By combining the above-mentioned POD reduction, PINN physical constraint modeling and EnKF real-time data assimilation, a second digital twin is obtained, namely a lightweight, high-fidelity, and real-time updated reduced-order digital twin.

[0068] 104. Use the second digital twin to generate temperature control commands based on temperature prediction data and real-time sensor network data.

[0069] Specifically, the second digital twin compares the temperature prediction data with the real-time sensor network data to obtain a residual vector; performs statistical analysis on the residual vector to obtain statistical analysis data; generates a thermal control strategy based on the statistical analysis data; performs security verification on the thermal control strategy; and generates a temperature control command based on the thermal control strategy when the security verification of the thermal control strategy passes.

[0070] Preferably, generating a thermal control strategy based on the statistical analysis data specifically includes: determining whether the satellite has a temperature anomaly based on the statistical analysis data; when a temperature anomaly exists, generating the thermal control strategy using a reinforcement learning algorithm based on a preset reward function and a greedy strategy.

[0071] Specifically, the digital twin predicts the sensor readings for the next moment based on the current state, compares the predicted values ​​with the actual sensor readings, and obtains the residual vector r = y_obs - y_pred. Statistical analysis is performed on the residual vector to calculate the mean μ and standard deviation σ of the residuals at each measurement point. When the residual at any measurement point satisfies |r_i| > 3σ, an anomaly is determined in the corresponding area. After fault detection is triggered, policy generation based on the DQN (Deep Q-Network) reinforcement learning algorithm is initiated: the DQN network adopts a structure of convolutional layers plus fully connected layers, using the principal components of the digital twin after order reduction as the state space, and the action space as the discretized action combination of the control commands of the adjustable actuator; the reward function comprehensively considers multiple objective constraints such as temperature deviation, power consumption, and actuator switching frequency; in the current state, DQN selects an action through an ε-greedy policy, the digital twin virtually executes the action and predicts the next state and reward, and the network is trained through an experience replay pool. After the strategy is generated, its security must be verified by a digital twin before the control strategy can be sent to the executor.

[0072] In some embodiments, the second digital twin predicts the sensor readings at the next moment based on the current state, compares the predicted values ​​with the actual sensor readings, and obtains a residual vector r = y_obs - y_pred. Statistical analysis is performed on the residual vector to calculate the mean μ and standard deviation σ of the residuals at each measuring point. When the residual at any measuring point satisfies |r_i| > 3σ, an anomaly is determined in the corresponding area. Further, fault types are identified through residual space pattern matching: pump performance degradation manifests as a systematic shift in the fluid loop flow residual; pipeline blockage manifests as an abnormal increase in the local temperature gradient; and sensor drift manifests as a continuous deviation of the residual at a single point while neighboring measuring points remain normal.

[0073] Upon fault detection, policy generation based on the DQN reinforcement learning algorithm is initiated. The DQN network employs a structure of 3 convolutional layers (3×3 kernels, 32-64-128 channels) plus 2 fully connected layers (256-128 neurons). The state space is based on the 100-dimensional principal components of the digital twin after reduction, and the action space is based on 500 action combinations discretized from the control instructions of 20 adjustable actuators. The reward function is designed as R = -w1·Σ(Ti-Tsafe)² - w2·Ppump - w3·Nswitch, where w1=0.5, w2=0.3, w3=0.2, Tsafe is the safe temperature threshold, and Nswitch is the number of actuator switching operations. In the current state s, DQN selects action a through an ε-greedy policy (ε=0.1). The digital twin virtually executes the action and predicts the next state s' and reward r'. The network is trained through an experience replay pool (capacity 10000). The target network is updated every 100 steps, with a learning rate of 0.001 and a discount factor γ=0.99.

[0074] After the strategy is generated, its safety is confirmed through virtual verification using a digital twin (simulating execution of 5 orbital cycles), and then the control strategy is converted into temperature control commands.

[0075] In some embodiments, such as Figure 3 and Figure 6As shown, this invention enables autonomous decision-making based on a real-time updated digital twin. This autonomous decision-making includes two branches: autonomous reconstruction decision-making and rolling optimization decision-making. The autonomous reconstruction decision-making is implemented through the following process: First, the digital twin predicts the sensor readings for the next moment based on the current state, compares the predicted values ​​with the actual sensor readings, and obtains a residual vector r = y_obs - y_pred. Statistical analysis is performed on the residual vector to calculate the mean μ and standard deviation σ of the residuals at each measuring point. When the residual at any measuring point satisfies |r_i| > 3σ, it is determined that there is an anomaly in the corresponding area. Further, fault types are identified through residual space pattern matching—pump performance degradation manifests as a systematic shift in the fluid loop flow residual; pipeline blockage manifests as an abnormal increase in the local temperature gradient; and sensor drift manifests as a continuous deviation of the single-point residual while neighboring measuring points remain normal. Upon fault detection, the system initiates a reconstruction strategy generation based on the DQN reinforcement learning algorithm. The DQN network employs a structure of 3 convolutional layers (3×3 kernels, 32-64-128 channels) plus 2 fully connected layers (256-128 neurons). The state space is based on the 100-dimensional principal components of the digital twin after reduction, and the action space is based on 500 action combinations discretized from the control instructions of 20 adjustable actuators. The reward function is designed as R = -w1·Σ(Ti-Tsafe)^2 - w2·Ppump - w3·Nswitch, where w1=0.5, w2=0.3, w3=0.2, Tsafe is the safe temperature threshold, and Nswitch is the number of actuator switching times; in the current state s, DQN selects action a through an ε-greedy policy (ε=0.1), the digital twin virtually executes the action and predicts the next state s' and reward r', and trains the network through an experience replay pool (capacity 10000), the target network is updated every 100 steps, the learning rate is set to 0.001, and the discount factor γ=0.99; after the policy is generated, the safety of the policy is confirmed by virtual verification by the digital twin (simulating execution of 5 track cycles) before the control policy can be issued to the actuator. The rolling optimization decision is based on AI load prediction and thermal control-task-energy collaborative optimization, and works in conjunction with autonomous reconfiguration decision to form a complete autonomous decision-making mechanism.

[0076] In some embodiments, such as Figure 3 and Figure 7As shown, the digital twin proposed in this application embodiment can also be optimized by rolling. Specifically, this invention employs Model Predictive Control (MPC) combined with Sequential Quadratic Programming (SQP) to achieve rolling optimization, comprehensively minimizing the multi-objective costs of temperature deviation, pump power consumption, and equivalent carbon emissions. The objective function of the MPC is designed as J = w1·Σ(Ti-Tref)² + w2·Ppump + w3·Emission, where w1=0.4, w2=0.3, and w3=0.3 are the weighting coefficients for temperature, power consumption, and carbon emissions, respectively, and Tref is the reference temperature (typically 60℃). During the optimization process, the following constraints must be satisfied simultaneously: temperature safety constraint Ti_min (40℃) ≤ Ti ≤ Ti_max (80℃), pump power consumption constraint Ppump ≤ Prated (rated power 50W), and actuator adjustment range constraints (louver opening θ∈[0%,100%], fluid distribution ratio q∈[0,1]). The solver employs a quasi-Newton method to approximate the Hessian matrix to accelerate convergence, with a convergence tolerance of 1e-6 and a maximum of 50 iterations. Actual measurements on the NVIDIA Jetson AGX Orin onboard processor show a single optimization solution time of 600-800ms, less than 1 second, meeting the time constraints of on-orbit real-time rolling optimization. The optimization cycle is set to 5-15 minutes, adapting to the changing thermal environment of low Earth orbit and the dynamic fluctuations of onboard AI workloads. Through the aforementioned MPC-SQP rolling optimization, the system achieves optimal synergy between thermal control efficiency and mission performance while ensuring temperature safety.

[0077] In some embodiments, the present invention executes control commands issued by the autonomous decision-making module through a software-defined reconfigurable thermal control actuator. Each actuator communicates with the autonomous decision-making module through a standardized hardware interface protocol. This interface protocol adopts a layered design: the physical layer supports dual redundant interfaces of CAN bus (2.0B, 1Mbps) and RS-485 (half-duplex, 115.2kbps), and the protocol layer adopts a custom lightweight instruction set TLC, whose instruction format is [frame header 2B | device ID 1B | instruction code 1B | parameter field 4B | checksum 1B | frame tail 1B], totaling 10 bytes, and the transmission time of a single frame is less than 1ms. The control instruction set includes SET_EMIT (sets radiator emissivity, parameter range 0.1-0.9, resolution 0.01), SET_FLOW (sets fluid flow rate, parameter range 0-100%, resolution 1%), SET_CONDUCT (sets thermal conductivity of thermal switch, parameter range 0.1-10 W / (m·K), resolution 0.1), GET_STATUS (reads actuator status), and RESET (actuator reset). The software-defined reconfigurable thermal actuators specifically include three types: Electrically addressable multilayer radiators employ electrochromic materials (WO3 / NiO double-layer structure), changing the material's optical properties by applying a 0-3V voltage to achieve continuously adjustable emissivity within the range of 0.1-0.9. The radiator is divided into 8 independently addressable regions (each region 10cm×10cm, thermally isolated between regions), with a response time of less than 1 second (measured 500-800ms) and an adjustment accuracy of ±0.02. Programmable fluid distribution networks utilize a miniature solenoid valve array (16 channels, 2mm diameter), adjusting the flow ratio of each branch through 100Hz PWM control (duty cycle 0-100%). The valve response time is less than 500ms (measured 200-400ms), with a flow adjustment accuracy of ±2% and a leakage rate of less than 0.1%. Intelligent thermal switch arrays are driven by shape memory alloys (SMA), controlling the SMA phase change through Joule heating to achieve a thermal conductivity of 0.1-10. Adjustable within the W / (m·K) range, with a switching response time of less than 2 seconds (measured at 1.2-1.8 seconds) and a cycle life exceeding 100,000 cycles. The three actuators described above achieve precise and flexible adjustment of radiative heat dissipation, fluid distribution, and heat conduction, respectively, with response times ranging from milliseconds to seconds, forming a complete software-defined thermal control actuator system.

[0078] 105. Perform temperature control processing on the satellite based on temperature control commands.

[0079] The temperature control command is sent to the onboard thermal control actuator, which includes, but is not limited to, adjustable thermal control devices such as liquid cooling pumps, radiators, and thermal switches. The precise control of the satellite temperature is achieved by adjusting the operating status of each actuator.

[0080] Tests showed that, using the method described in this embodiment, the system could run continuously for 500 orbital cycles in an LEO orbital environment, with the surface temperature difference of the cold plate remaining stable within ±0.8K, and the overall energy efficiency ratio improved by 28% compared to the traditional threshold control strategy.

[0081] Preferably, the method further includes a constellation-level cooperative scheduling step: Obtain the neighbor table of the satellite, which includes the thermal state data of at least one neighbor satellite; when the thermal margin data of the satellite is less than a preset margin threshold, determine the non-real-time inference task of the satellite; determine the target neighbor satellite based on the thermal state data of the at least one neighbor satellite; migrate the non-real-time inference task of the satellite to the target neighbor satellite.

[0082] And / or: Obtain the neighbor table of the satellite, which includes the thermal state data of at least one neighbor satellite; perform global thermal load balancing processing on the satellite and the at least one neighbor satellite based on the thermal state data of the satellite and the thermal state data of the at least one neighbor satellite.

[0083] Specifically, a thermal mutual assistance network is constructed within the constellation, employing a two-layer ring communication topology: satellites in the same orbital plane form an inner ring through inter-satellite links, while satellites in different orbital planes form an outer ring through inter-orbit links. Each satellite transmits thermal status data packets at a preset broadcast period. The thermal mutual assistance network achieves thermal energy transfer through two modes: (1) Task migration mode—when the thermal margin of a satellite is lower than a preset threshold, non-real-time inference tasks are migrated to cryogenic satellites; (2) Thermal load redistribution mode—the ADMM (Alternating Direction Method of Multipliers) distributed optimization algorithm is used to solve the global thermal load balancing problem. Each satellite uses the local actuator adjustment amount as the decision variable, constructs a local objective function and satisfies global consensus constraints, and achieves overall constellation thermal load balancing through distributed iterative solution.

[0084] In some embodiments, a thermal mutual assistance network is constructed within the constellation, employing a two-layer ring communication topology: satellites in the same orbital plane form an inner ring via inter-satellite links (ISL, Ka-band, 100Mbps bandwidth, communication latency less than 10ms); satellites in different orbital planes form an outer ring via inter-orbit links (IOL, 50Mbps bandwidth, communication latency less than 50ms). Each satellite maintains a neighbor table, recording the thermal status of two neighboring satellites in the same ring and two in the outer ring. Each satellite broadcasts a thermal status data packet (HSP) every 1-5 seconds, containing information such as satellite ID, timestamp, temperature field principal component coefficient vector (100-dimensional), available thermal margin, task queue load rate, and energy status.

[0085] Heat-cooperative networks achieve heat transfer through two modes: (a) Task Migration Mode: When a satellite's thermal margin falls below the 20% threshold, it sends a task migration request to neighboring satellites. The receiving satellite decides whether to accept the request based on its own thermal margin and task queue status. Task migration is implemented through the onboard AI load scheduler, migrating non-real-time inference tasks (such as image preprocessing) to cryogenic satellites to reduce local thermal load. The task migration data size is less than 100MB, and the transmission time on the ISL link is less than 10 seconds.

[0086] (II) Heat Load Redistribution Mode: The ADMM distributed optimization algorithm is used to solve the global heat load balancing problem. Each satellite uses the local actuator adjustment xi as the decision variable to construct a local objective function fi(xi)=ci·(Ti-Tref)²+ri·Pi, satisfying the global consensus constraint xi=z (global average heat load). The ADMM iterative formulas are: xi^(k+1)=argmin fi(xi)+(ρ / 2)||xi-z^(k)+ui^(k)||²; z^(k+1)=(1 / N)Σ(xi^(k+1)+ui^(k)); ui^(k+1)=ui^(k)+xi^(k+1)-z^(k+1). The penalty parameter ρ=0.5, the typical number of convergence iterations is 20-50, and the solution delay is less than 1 second.

[0087] Through the synergistic mechanism of mission migration and heat load redistribution, the risk of local overheating in the constellation is dispersed to the whole, the overall energy efficiency of the constellation is improved by 15%, and the peak heat load of a single star is reduced by 20%.

[0088] In some embodiments, such as Figure 3 and Figure 8As shown, this invention constructs a thermal cooperation network within a constellation and employs the ADMM distributed optimization algorithm to achieve overall constellation thermal load balancing and energy efficiency optimization. The communication topology of this thermal cooperation network adopts a two-layer ring structure: satellites in the same orbital plane form an inner ring via inter-satellite links (ISL, Ka-band, bandwidth 100Mbps, communication latency less than 10ms), while satellites in different orbital planes form an outer ring via inter-orbit links (IOL, bandwidth 50Mbps, communication latency less than 50ms). Each satellite maintains a neighbor table, recording the thermal status of two neighboring satellites in the same ring and two in the outer ring. Each satellite broadcasts a thermal status data packet (HSP) every 1-5 seconds. This data packet contains information such as satellite ID, timestamp, temperature field principal component coefficient vector (100-dimensional, 400B), available heat margin, task queue load rate, and energy status, with a single packet size of approximately 420B. The thermal mutual assistance network achieves thermal energy transfer through two modes: The first is a task migration mode, where when a satellite's thermal margin falls below the 20% threshold, it sends a task migration request to neighboring satellites, migrating non-real-time inference tasks (such as image preprocessing) to cryogenic satellites to reduce local thermal load. The task migration data size is less than 100MB, and the transmission time on a 100Mbps ISL link is less than 10 seconds. The second is a thermal load redistribution mode, which solves the global thermal load balancing problem using the ADMM distributed optimization algorithm. Each satellite uses its local actuator adjustment xi as the decision variable, constructing a local objective function fi(xi)=ci·(Ti-Tref)²+ri·Pi, while satisfying the global consensus constraint xi=z (global average thermal load). The ADMM iterative formula xi^(k+1)=argmin is then used to determine the thermal load. The functions fi(xi)+(ρ / 2)||xi-z^(k)+ui^(k)||², z^(k+1)=(1 / N)Σ(xi^(k+1)+ui^(k)), and ui^(k+1)=ui^(k)+xi^(k+1)-z^(k+1) are solved in a distributed manner with a penalty parameter ρ=0.5. The typical convergence iteration count is 20-50, and the solution latency is less than 1 second. Through the collaborative mechanism of task migration and heat load redistribution described above, the risk of local overheating in the constellation is dispersed globally, improving the overall energy efficiency of the constellation by 15%, reducing the peak heat load of a single star by 20%, and achieving a solution latency of less than 1 second.

[0089] In some embodiments, for satellites equipped with high-performance onboard AI chips (computing power greater than 275 TOPS), the dimensionality can be reduced to a 200-dimensional principal component space to retain more detailed temperature field information, further improving the temperature field reconstruction accuracy to within 0.5%. For satellites equipped with low-performance onboard AI chips (limited computing power), the dimensionality can be reduced to a 50-dimensional principal component space to further reduce the computational load, shortening the time for a single update to within 5ms.

[0090] In some embodiments, for satellites with a small number of actuators (less than 10), the DQN network can adopt a lighter structure—two convolutional layers plus one fully connected layer—to reduce computational overhead, and the policy generation time can be shortened to less than 40 seconds. For satellites with a large number of actuators (more than 30), the DQN network can increase its depth—four convolutional layers plus three fully connected layers—to improve its ability to represent complex action spaces, and the policy generation accuracy can be improved to over 98%.

[0091] In this embodiment, through the collaboration of a lightweight temperature prediction model and a digital twin, the temperature prediction latency is reduced from minutes in existing technologies to less than one second, providing sufficient feedforward time for proactive temperature control. Based on the autonomous decision-making mechanism of the digital twin, the strategy generation time under fault conditions is shortened from several hours to less than one minute, and the strategy switching time is less than 10 seconds, achieving on-orbit autonomous temperature control without waiting for ground commands. Through multi-objective rolling optimization, comprehensively considering the costs of temperature deviation, power consumption, and other dimensions, the overall energy efficiency is improved by more than 25%, and the surface temperature difference of the cold plate is controlled within ±0.8K. The overheating mutual assistance network and ADMM distributed optimization algorithm improve the overall energy efficiency of the constellation by 15%, reduce the peak heat load of a single satellite by 20%, and achieve a solution latency of less than one second, realizing constellation-level heat load balancing. In addition, this embodiment can be deployed in the onboard computers of various on-orbit satellites, space stations, and deep space probes to realize on-orbit autonomous perception, real-time decision-making, and intelligent execution of satellite thermal control systems. This invention is particularly applicable to the thermal management of low Earth orbit computing satellites and constellations, and can effectively improve the adaptive capability, control accuracy and on-orbit reliability of satellite thermal control systems, and has broad industrial application prospects.

[0092] To better implement the AI-based temperature control method provided in this application, one embodiment also provides an AI-based temperature control device, which can be integrated into a computer device. The meanings of the terms used are the same as in the AI-based temperature control method described above, and specific implementation details can be found in the description of the method embodiment.

[0093] In one embodiment, an artificial intelligence-based temperature control device is provided, which can be specifically integrated into a computer device, such as... Figure 9 As shown, the AI-based temperature control device includes: an acquisition unit 201, an establishment unit 202, an adjustment unit 203, a generation unit 204, and a control unit 205, as detailed below: Acquisition unit 201 is used to acquire satellite temperature prediction data and real-time sensor network data; Establishment unit 202 is used to establish a first digital twin based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; The adjustment unit 203 is used to adjust the first digital twin to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; Generation unit 204 is used to generate temperature control commands based on the temperature prediction data and real-time sensor network data using the second digital twin; Control unit 205 is used to perform temperature control processing on the satellite based on the temperature control command.

[0094] In some embodiments, the adjustment unit 203 may include: The dimensionality reduction subunit is used to perform dimensionality reduction processing on the first digital twin to obtain the third digital twin; The modeling subunit is used to perform physical data fusion modeling processing on the third digital twin using a physical information neural network to obtain a fourth digital twin; The assimilation subunit is used to assimilate the data of the fourth digital twin using ensemble Kalman filtering to obtain the second digital twin.

[0095] In some embodiments, the dimensionality reduction subunit may include: The sampling module is used to perform snapshot sampling on the first digital twin to obtain snapshot data; A construction module is used to construct a low-dimensional space based on the snapshot data; The truncation module is used to perform energy truncation processing on the low-dimensional space to obtain the third digital twin.

[0096] In some embodiments, the generation unit 204 may include: The comparison subunit is used by the second digital twin to compare the temperature prediction data and the real-time sensor network data to obtain a residual vector; The statistical subunit is used to perform statistical analysis on the residual vector to obtain statistical analysis data. The first generation subunit is used to generate a thermal control strategy based on the statistical analysis data; A verification subunit is used to perform safety verification on the thermal control strategy. The second generation subunit is used to generate temperature control commands based on the thermal control strategy when the safety verification of the thermal control strategy passes.

[0097] In some embodiments, the first generating subunit may include: The judgment module is used to determine whether the satellite has a temperature anomaly based on the statistical analysis data; The generation module is used to generate the thermal control strategy based on a preset reward function and a greedy strategy using a reinforcement learning algorithm when an abnormal temperature occurs.

[0098] In some embodiments, the temperature control device may further include: The first acquisition unit is used to acquire the neighbor table of the satellite, the neighbor table including the thermal state data of at least one neighbor satellite of the satellite; The first determining unit is used to determine the non-real-time inference task of the satellite when the satellite's thermal margin data is less than a preset margin threshold. The second determining unit is used to determine the target neighboring satellite based on the thermal state data of the at least one neighboring satellite; A migration unit is used to migrate the non-real-time inference tasks of the satellite to the target neighboring satellite.

[0099] In some embodiments, the temperature control device may further include: The second acquisition unit is used to acquire the neighbor table of the satellite, the neighbor table including the thermal state data of at least one neighbor satellite of the satellite; The balancing unit is used to perform global thermal load balancing processing on the satellite and the at least one neighboring satellite based on the thermal state data of the satellite and the thermal state data of the at least one neighboring satellite.

[0100] The aforementioned AI-based temperature control device can improve the accuracy of AI-based temperature control.

[0101] This application also provides a computer device, which may include a terminal or a server. For example, the computer device may serve as an artificial intelligence-based temperature control terminal, such as a mobile phone, tablet computer, etc.; or, for example, the computer device may serve as a server, such as an artificial intelligence-based temperature control server. Figure 10 As shown, it illustrates a structural schematic diagram of the terminal involved in an embodiment of this application. Specifically:

[0102] The computer device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 10 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0103] The processor 401 is the control center of the computer device, connecting various parts of the computer device through various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 402, and by calling data stored in the memory 402. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user page, and application programs, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 401.

[0104] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0105] The computer device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0106] The computer device may also include an input unit 404, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0107] Although not shown, the computer device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the computer device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions, as follows:

[0108] Acquire satellite temperature prediction data and real-time sensor network data; A first digital twin is established based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; The first digital twin is adjusted to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; The second digital twin generates temperature control commands based on the temperature prediction data and real-time sensor network data; The satellite is subjected to temperature control processing based on the temperature control command.

[0109] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0110] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the above embodiments.

[0111] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by a computer program, or by a computer program controlling related hardware. The computer program can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0112] Therefore, embodiments of this application also provide a storage medium storing a computer program that can be loaded by a processor to execute steps in any of the artificial intelligence-based temperature control methods provided in embodiments of this application. For example, the computer program can execute the following steps:

[0113] Acquire satellite temperature prediction data and real-time sensor network data; A first digital twin is established based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; The first digital twin is adjusted to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; The second digital twin generates temperature control commands based on the temperature prediction data and real-time sensor network data; The satellite is subjected to temperature control processing based on the temperature control command.

[0114] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0115] Since the computer program stored in the storage medium can execute the steps of any of the artificial intelligence-based temperature control methods provided in the embodiments of this application, the beneficial effects that any of the artificial intelligence-based temperature control methods provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0116] The above provides a detailed description of an artificial intelligence-based temperature control method, apparatus, computer device, and storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A temperature control method based on artificial intelligence, characterized in that, include: Acquire satellite temperature prediction data and real-time sensor network data; A first digital twin is established based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; The first digital twin is adjusted to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; The second digital twin generates temperature control commands based on the temperature prediction data and real-time sensor network data; The satellite is subjected to temperature control processing based on the temperature control command.

2. The method according to claim 1, characterized in that, The step of adjusting the first digital twin to obtain the second digital twin includes: The first digital twin is subjected to dimensionality reduction processing to obtain the third digital twin; A fourth digital twin is obtained by performing physical data fusion modeling on the third digital twin using a physical information neural network. The fourth digital twin is assimilated using ensemble Kalman filtering to obtain the second digital twin.

3. The method according to claim 2, characterized in that, The step of performing dimensionality reduction on the first digital twin to obtain the third digital twin includes: A snapshot of the first digital twin is performed to obtain snapshot data; A low-dimensional space is constructed based on the snapshot data; The energy truncation process is performed on the low-dimensional space to obtain the third digital twin.

4. The method according to claim 1, characterized in that, The step of generating temperature control commands using the second digital twin based on the temperature prediction data and real-time sensor network data includes: The second digital twin compares the temperature prediction data with the real-time sensor network data to obtain a residual vector; Statistical analysis is performed on the residual vector to obtain statistical analysis data; A thermal control strategy is generated based on the statistical analysis data; The thermal control strategy was verified for safety. When the safety verification of the thermal control strategy is passed, a temperature control command is generated based on the thermal control strategy.

5. The method according to claim 4, characterized in that, A thermal control strategy is generated based on the statistical analysis data, including: Based on the statistical analysis data, determine whether the satellite has any temperature anomalies; When an abnormal temperature occurs, the heat control strategy is generated using a reinforcement learning algorithm based on a preset reward function and a greedy strategy.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Obtain the neighbor table of the satellite, which includes the thermal state data of at least one of the satellite's neighboring satellites; When the satellite's thermal margin data is less than a preset margin threshold, the satellite's non-real-time inference task is determined. The target neighboring satellite is determined based on the thermal state data of the at least one neighboring satellite; The non-real-time inference task of the satellite is migrated to the target neighbor satellite.

7. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Obtain the neighbor table of the satellite, which includes the thermal state data of at least one of the satellite's neighboring satellites; Based on the thermal state data of the satellite and the thermal state data of at least one neighboring satellite, a global thermal load balancing process is performed on the satellite and the at least one neighboring satellite.

8. A temperature control device, characterized in that, include: The acquisition unit is used to acquire satellite temperature prediction data and real-time sensor network data; A building unit is used to build a first digital twin based on the temperature prediction data and real-time sensor network data, wherein the first digital twin is a full-order digital twin; An adjustment unit is used to adjust the first digital twin to obtain a second digital twin, wherein the order of the second digital twin is lower than that of the first digital twin; The generation unit is used to generate temperature control commands based on the temperature prediction data and real-time sensor network data using the second digital twin; The control unit is used to perform temperature control processing on the satellite based on the temperature control command.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores an application program, and the processor runs the application program within the memory to perform the operations in the artificial intelligence-based temperature control method according to any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores multiple instructions adapted for loading by a processor to execute the steps of the artificial intelligence-based temperature control method according to any one of claims 1 to 7.

Images