A deep learning-based gas film cooling temperature field digital twinning method and system

Abstract

The application discloses a kind of based on deep learning's gas film cooling temperature field digital twinborn method and system, it is related to heat protection technical field, the method includes the following steps: constructing gas film cooling experimental device, the experimental device includes sensor, measurement and data acquisition system, complete the measurement and acquisition of experimental data;Acquisition experimental data, use infrared camera to shoot the temperature image of incomplete orifice plate and save, experimental data are used to fine-tune pre-training model, generate migration model;Training pre-training model, using CFD result as pre-training data source, generate pre-training model;Assimilate numerical data, assimilate numerical data using migration learning method based on iterative neural operator, numerical data include data from different image size and image layer;Respectively carry out space, structure and working condition deduction, update physical-virtual synchronous model.The application has data assimilation and deduction function, can support the efficient and accurate design of gas film cooling.

Description

Technical Field

[0001] This invention relates to the field of thermal protection technology, and in particular to a digital twin method and system for air film cooling temperature field based on deep learning. Background Technology

[0002] The intelligentization of design, manufacturing, testing, and operation and maintenance is the development trend of thermal protection technology. For correlation / empirical models, the disadvantages are high dimensionality and low accuracy; while for experimental / CFD (Computational Fluid Dynamics) methods, the disadvantages are high cost and long time consumption. Therefore, it is essential to establish reliable high-dimensional and high-precision prediction models and realize rapid prediction of multiphysics fields for the design and optimization of aerospace thermal protection.

[0003] Gas turbine systems demand higher operating temperatures for higher efficiency. Currently, the turbine inlet temperature of gas turbine engines far exceeds the tolerable range of parallel high-temperature alloy materials. Therefore, developing high-temperature alloy / ceramic materials and applying active cooling technologies to protect the hot sections of gas turbine systems has become crucial. Current cooling technologies generally integrate various internal and external cooling methods. Among these methods, film cooling (FSU) is one of the most widely used external cooling methods. Coolant air is discharged through slots or discrete orifices, forming a thin film of coolant air on the outer surface of hot components, preventing thermal erosion of the turbine blades by high-temperature gases and achieving the purpose of cooling protection.

[0004] The cooling effect of the adiabatic wall is the main criterion for evaluating film cooling (FSV). FSV efficiency is influenced by various factors such as the air-to-water ratio, density ratio, mainstream turbulence intensity, orifice shape, and orifice location. To achieve higher cooling performance, accurate prediction of local temperature details and optimization of FSV parameters are necessary. However, due to the complex interaction between the coolant jet and the mainstream flow, it is difficult to understand the FSV flow field and its corresponding cooling efficiency distribution. Currently, the feasibility evaluation of FSV is still based on experiments, utilizing mature non-contact measurement technologies such as pressure-sensitive coatings, infrared thermal imaging, and liquid crystals. To alleviate the challenges mentioned above and obtain a high-fidelity and fast-response two-dimensional efficiency distribution, some methods proposed for FSV modeling have the following advantages and disadvantages:

[0005] 1) In predicting the cooling effect of a single-hole array, one of the most widely used correlations is Baldauf's one-dimensional correlation. Following this, Sellers first proposed a superposition calculation model for film cooling channels. Its drawbacks are that Baldauf's correlation cannot accurately predict the effectiveness of different types of film cooling; Sellers' superposition equations cannot account for local conditions or details of local geometry. Simple explicit equations are insufficient to simulate the complex surface temperature distribution of superimposed film cooling structures, even for lateral averages. Besides accuracy issues, these correlations also have low dimensionality, providing only one-dimensional results for regularly arranged film cooling holes. Using randomly distributed discrete holes, few correlations can predict the two-dimensional distribution of surface cooling efficiency. This is due to the extremely high complexity and nonlinearity of the flow field resulting from the interaction and superposition of randomly distributed jets from upstream to downstream. Especially when considering high-temperature and high-pressure conditions, the problem involves more nonlinear characteristics, such as changes in gas / air properties, making the cooling effect response even more complex.

[0006] 2) In the field of CFD, research on cross-flow jet phenomena over the past decade has revealed significant challenges in simulating near-wall turbulence in film-cooled structures. Anisotropic turbulent diffusion is considered a crucial mechanism that traditional two-equation turbulence methods (such as shear stress transport models) cannot simulate. Furthermore, Prandtl number calculations are inaccurate, and the anti-gradient scalar diffusion mechanism exhibits anomalies exceeding the capabilities of diffusion models. Research attempts to address these film-cooling prediction problems generally rely on prior knowledge of turbulence and analytical derivations, resulting in low integration with data. Pioneer research following this approach came from Li et al. To obtain more accurate Reynolds stress and turbulent scalar flux models, they proposed an algebraic anisotropic turbulence model and validated it through experiments with leading-edge lotus-head divergence and full-coverage divergence blades. Through comparison of four different turbulence models, they demonstrated good qualitative and quantitative consistency between the algebraic anisotropic model and experimental data. Subsequent research by Li et al. proposed an improved algebraic anisotropic eddy viscosity (AAEV) method that considers the effects of wall and mean flow field strain on the anisotropy ratio. The new AAEV k-ε model correctly predicts the spanwise expansion of the film and the decay of the secondary eddies.

[0007] 3) In machine learning, while turbulence modeling is an unavoidable problem in computational fluid dynamics, data-driven methods can circumvent this difficulty at the cost of data. Deep learning methods, which emerged in 2012, have shown great potential in regressing many physical fields in image or matrix formats and are naturally compatible with film cooling (FSD) prediction. Under the guarantee of universal approximation theorems, researchers have attempted to establish a mapping between geometry and FSD effects. Results show that deep learning models can identify abstract features in datasets with high accuracy and good generalization ability. The main advantage of these deep learning methods is their extremely fast FSD prediction speed, typically 105 times faster than CFD methods. However, because numerical data is used to train deep learning models, concerns remain regarding the accuracy of FSD predictions, as the regression model will not be more accurate than the data that drives it. Possible solutions lie in experimental data or LES data. Unfortunately, such high-fidelity data is still far from sufficient for training deep neural networks, both in terms of the quantity and distribution of FSD data.

[0008] 4) In data fusion methods for thermofluid problems, a potential technique to improve the modeling accuracy of film cooling is to integrate simulated data with experimental data, a process known as "data assimilation." These methods are typically based on existing partially analytical physics, such as the Navier-Stokes equations, and model several parameters of these equations using experimental data. Li et al. proposed a continuous adjoint data assimilation model that successfully reconstructed the global turbulent mean flow from a limited number of wall pressure measurements. This model showed good predictive performance for high Reynolds number separated flows, strong adverse pressure gradient flows, and complex flows. They introduced an anisotropic adjoint data assimilation scheme to supplement the limited spatial range and measurable turbulent mean flow measurement data. They achieved good results in reconstructing complex flow fields such as circular jet flow fields, flow fields over blunt plates, and flow fields over ribbed walls. Deng et al. used an integrated-Kalman filter-based data assimilation method to optimize RANS model constants and recover the global flow field from local measurement data. They validated the effectiveness of the data assimilation method on four different RANS models. All research models using the data assimilation method showed significant improvement, with the kk model performing best. While data assimilation methods can be a good tool for improving the accuracy of film cooling predictions, they still have two major drawbacks: a) experimental film cooling data are generally interface temperatures (effectiveness), while a large amount of simulated data is in the mainstream, leading to a data distribution mismatch; b) data assimilation methods still require solving the Navier-Stokes equations, which is too slow to provide designers with an immediate response. Solutions may still involve machine learning.

[0009] Transfer learning is a special category of machine learning and data assimilation methods that provides the ability to improve the modeling accuracy of a dataset by leveraging a pre-trained model obtained from a similar dataset. In most cases, the source dataset has a large amount of data with a good distribution, while the target dataset may have very little data. The pre-trained model learns an initial mapping on the source dataset and is then fine-tuned on the target dataset. If the two learning tasks are similar, the transferred model will successfully inherit some of the internal structure of the pre-trained model and perform accurately on the target dataset. Existing transfer learning methods include model-based transfer learning (fine-tuning), instance-based transfer learning, and feature-based transfer learning.

[0010] In recent research on transfer learning and physics, several publications on thermofluid problems have revealed the potential of transfer learning to achieve high accuracy on small datasets. Wang et al. proposed a transfer learning model based on Conditional Generative Adversarial Networks (CGANs) to predict the three-dimensional pressure distribution on turbine blade surfaces. The model transfers knowledge learned from large-scale low-fidelity datasets to small-scale high-fidelity datasets. Results show that transfer learning effectively improves generalization accuracy on small datasets. Obiols-Sales et al. introduced Super-Resolution Flow Network (SURFNet) to accelerate high-resolution turbulent CFD simulations. SURFNet is trained on low-resolution simulation data and then fine-tuned on high-resolution datasets. This model exhibits good resolution invariance and generalization ability at the cost of 1 / 15th the training data.

[0011] Therefore, those skilled in the art are dedicated to developing a deep learning-based digital twin method and system for air film cooling temperature fields. Summary of the Invention

[0012] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is that experimental film cooling data are generally interface temperatures, while a large amount of simulation data are in the mainstream flow, resulting in data distribution mismatch. The data assimilation method has a slow solution speed and cannot provide designers with an immediate response. To achieve the above objective, the present invention provides a deep learning-based digital twin method for film cooling temperature fields, characterized in that the method includes the following steps:

[0013] S101: Construct an air film cooling experimental device, the experimental device including sensors, measurement and data acquisition system, the experimental device completes the measurement and acquisition of experimental data;

[0014] S103: Collect experimental data, use an infrared camera to capture and save the temperature image of the incomplete orifice plate. The experimental data is used to fine-tune the pre-trained model and generate a transfer model.

[0015] S105: Train the pre-trained model by using CFD results as the pre-training data source to generate the pre-trained model;

[0016] S107: Assimilate the numerical data using a transfer learning method based on iterative neural operators, wherein the numerical data includes data from different image sizes and image layers;

[0017] S109: Perform spatial, structural, and operational condition simulations respectively, and update the physical-virtual synchronization model.

[0018] Further, in step S101, the air-film cooling experimental device includes a wind tunnel, a pressure vessel, a flow meter, a cryostat, an infrared thermal imager, and a measurement and data acquisition system; wherein, the wind tunnel has a rectangular cross-section, and a honeycomb structure is installed upstream of the measurement area of ​​the wind tunnel to provide uniform airflow; the pressure vessel provides driving force for the main flow and secondary flow; the flow meter is installed in the middle of the wind tunnel to measure the velocity of the main flow; the secondary flow is cooled to -5°C by the cryostat before pressure and temperature measurements are performed; the infrared thermal imager records the temperature distribution on the outer surface of the jet cooling plate; the measurement and data acquisition system uses thermocouples to measure the temperature of the main flow and secondary flow in real time and performs data acquisition; the mixture of the main flow and the secondary flow is finally discharged into the environment from the wind tunnel outlet.

[0019] Further, in step S103, the experimental data is obtained under predetermined experimental conditions, which include: the experiment is conducted at room temperature and pressure, the back pressure of the main flow and the secondary flow is 1 atm, the temperature of the main flow is 299.60 K, and the temperature of the secondary flow is 273.15 K; the experimental sample material is an additive manufacturing material, the additive manufacturing material is an unsaturated polyester resin, and the thermal conductivity of the material is 0.1 W / m / K.

[0020] Further, in step S105, training the pre-trained model includes the following steps:

[0021] S1051: Use CFD results as a pre-training data source;

[0022] S1052: Selecting suitable hyperparameters using five-fold cross-validation;

[0023] S1053: The learning rate during training is set to 1e-3;

[0024] S1054: Update the weights of the pre-trained model using the Adam optimizer;

[0025] S1055: Convert the numerical data samples into images for machine learning to obtain the pre-trained model with generalization accuracy.

[0026] Furthermore, in step S1051, a gas film geometry is obtained by simulating a plate through which a randomly distributed perforation hole passes under isothermal conditions, and CFD simulation is performed using the gas film geometry to generate CFD numerical data.

[0027] Furthermore, the computational domain of the CFD simulation only includes the fluid domain, the computational unit is an unstructured polyhedron, and a prism mesh is generated in the boundary layer region; the gas film geometry is generated under preset constraints, the constraints being that randomly distributed seepage holes simulate passing through a flat plate under isothermal conditions, the length of the seepage hole distribution area is 40mm, the diameter of the seepage hole is 1mm, the number of seepage holes is 10-15, and the seepage holes are distributed according to predetermined conditions; wherein, the predetermined conditions include: the holes do not overlap, the center distance of the holes is not less than twice the diameter of the holes; the hole cavity configuration is diverse; the holes on the outer surface are distributed within a rectangular area of ​​40mm×20mm.

[0028] Further, in step S107, the partial differential equation of the iterative neural operator is:

[0029]

[0030] The difference is expressed in the following form:

[0031]

[0032] , where u is the physical quantity to be solved, F is a nonlinear function of u and its derivative, Ω is the computational domain, and Γ is the boundary.

[0033] Further, in step S109, the pre-trained model is transferred to the experimental data using a model-based transfer learning method, the pre-trained model is fine-tuned to generate the transfer model, the learning rate of the fine-tuning process is set to 1e-4, and all convolutional layers in the pre-trained model are set to trainable during the fine-tuning process.

[0034] Furthermore, the inference uses predetermined parameters for convolutional kernel size, number of hidden layers, and number of hidden state layers to train the iterative neural operator network model, and updates the model weights using the Adam optimizer.

[0035] The spatial extrapolation uses incomplete air film cooling efficiency images to train the iterative neural operator network, enabling the iterative neural operator network to learn the common physical laws of space, extrapolate the physical image of the incomplete part, output the complete air film cooling efficiency image, and recover the image of the entire temperature field.

[0036] The structural deduction uses test samples with a number of holes and a hole arrangement completely different from the training set to train the iterative neural operator network. This enables the iterative neural operator network to deduce geometric layers with different hole distributions and positions and predict the gas film superposition phenomenon, and output a complete gas film cooling efficiency image to recover the image of the entire temperature field.

[0037] The operating condition simulation uses test samples with operating conditions completely different from the training set to train the iterative neural operator network. The iterative neural operator network learns the influence of operating conditions on the film cooling efficiency and deduces the physical picture under different operating conditions, outputs a complete film cooling efficiency picture, and recovers the picture of the entire temperature field. The operating conditions include: different air blowing ratios, different turbulence intensities, and different density ratios.

[0038] On the other hand, the present invention also provides a deep learning-based digital twin system for air-film cooling temperature fields, characterized in that the system is constructed using a deep learning-based digital twin method for air-film cooling temperature fields, and the system includes a model training module, a sensor module, a data acquisition module, a data processing module, a physical-virtual synchronization module, and an inference module; wherein,

[0039] The model training module uses machine learning to pre-train the model to obtain a pre-trained model, and then fine-tunes the pre-trained model through transfer learning to generate a transfer model.

[0040] The sensor module includes a sensor interface, through which the data acquisition module completes the acquisition of air film cooling test data.

[0041] The data acquisition module acquires CFD simulation data and air film cooling experimental data. After being processed by the data processing module, the data is used as training data and test data for the model training module.

[0042] The data processing module processes the data acquired by the data acquisition module. The processing includes converting the CFD simulation values ​​into images, assimilating the numerical data, and performing regression processing on the CFD simulation data and the experimental data.

[0043] The physical-virtual synchronization module uses a model-based transfer learning method to transfer the pre-trained model to the experimental data, fine-tunes the pre-trained model, and generates the transfer model, thereby achieving synchronization between the physical model and the virtual model.

[0044] The inference module uses the model trained by the model training module to infer space, structure, and working conditions, and updates the physical-virtual synchronous model.

[0045] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The digital twin method for air film cooling designed in this invention has data assimilation and extrapolation functions, which can support the efficient and precise design of air film cooling.

[0047] 2. The iterative neural operator model designed in this invention can accurately capture the nonlinear characteristics of fluid accumulation and predict the local fluid cooling effect of random pore configurations with high quality. Compared with direct machine learning methods, transfer learning significantly reduces the requirements for the number of training samples (reduced by 2-3 times) and training cycles (reduced by 5-6 times) while achieving the same accuracy. With the same data cost and training cost, the error of transfer learning is reduced by 63.7%, 41.8%, and 46.2%, respectively.

[0048] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a preferred embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the construction framework of a preferred embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram illustrating the implementation steps of a preferred embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram illustrating the steps of assimilating numerical data using a preferred embodiment of the transfer learning method based on iterative neural operators of the present invention.

[0053] Figure 5 This is a schematic diagram of the transfer learning steps in a preferred embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of hardware and software synchronization according to a preferred embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of a preferred embodiment of the gas film cooling experimental apparatus of the present invention;

[0056] Figure 8 This is a structural deduction diagram of a preferred embodiment of the present invention;

[0057] Figure 9 This is a working condition simulation diagram of a preferred embodiment of the present invention.

[0058] Among them: 1-compressor, 2-valve, 3-flow meter, 4-honeycomb, 5-infrared thermal imager, 6-infrared transparent glass, 7-airflow channel, 8-air film cooling plate, 9-wind tunnel wall, 10-air supply chamber, 11-cooling air inlet, 12-host computer, 13-data acquisition unit, 14-low temperature thermostat, 15-flow controller, 16-pressure regulator. Detailed Implementation

[0059] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0060] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0061] like Figure 1 As shown in the figure, an embodiment of the present invention provides a digital twin method for air film cooling temperature field based on deep learning, which includes the following steps:

[0062] S101: Construct an air film cooling experimental device to complete the measurement and acquisition of experimental data.

[0063] The air-film cooling experimental setup includes a wind tunnel, a pressure vessel, a flow meter 3, a cryostat 14, an infrared thermal imager 5, and a measurement and data acquisition system, such as... Figure 7As shown in the diagram, the wind tunnel has a rectangular cross-section. A honeycomb structure 4 is installed upstream of the measurement area of ​​the wind tunnel to provide uniform airflow. A pressure vessel provides driving force for the main flow and secondary flow. A flow meter 3 is installed in the middle of the wind tunnel to measure the velocity of the main flow. The secondary flow is cooled to -5°C by a cryostat 14 before pressure and temperature measurements are performed. An infrared thermal imager 5 records the temperature distribution on the outer surface of the jet cooling plate. The measurement and data acquisition system uses thermocouples to measure the temperature of the main flow and secondary flow in real time and collects the data. The mixture of the main flow and secondary flow is finally discharged into the environment from the wind tunnel outlet.

[0064] S103: Collect experimental data, use an infrared camera to capture and save the temperature image of the incomplete orifice plate. The experimental data is used to fine-tune the pre-trained model and generate a transfer model.

[0065] The above experimental data were obtained under predetermined experimental conditions, which included: the experiment was conducted at room temperature and pressure, the back pressure of the main flow and the secondary flow was 1 atm, the temperature of the main flow was 299.60 K, and the temperature of the secondary flow was 273.15 K; the experimental sample material was an additive manufacturing material, which was an unsaturated polyester resin with a thermal conductivity of 0.1 W / m / K.

[0066] S105: Train the pre-trained model by using CFD results as the pre-training data source.

[0067] The training of a pre-trained model includes the following steps:

[0068] S1051: Using CFD results as a pre-training data source: A gas film geometry is obtained by simulating a randomly distributed perforated plate under isothermal conditions. This gas film geometry is then used for CFD simulation to generate CFD numerical data, which serves as the pre-training data source. The computational domain of the CFD simulation only includes the fluid domain, the computational unit is an unstructured polyhedron, and a prismatic mesh is generated in the boundary layer region. The gas film geometry is generated under the following constraints: a randomly distributed perforated plate is simulated under isothermal conditions, with a perforation distribution area length of 40 mm, a perforation diameter of 1 mm, and a number of 10–15 perforations. These perforations are distributed according to the following conditions: the perforations do not overlap, the center-to-center distance between perforations is not less than twice the perforation diameter, the perforation configuration is diverse, and the perforations on the outer surface are distributed within a 40 mm × 20 mm rectangular area.

[0069] S1052: Selecting suitable hyperparameters using five-fold cross-validation;

[0070] S1053: The learning rate during training is set to 1e-3;

[0071] S1054: Update the weights of the pre-trained model using the Adam optimizer;

[0072] S1055: Convert the numerical data samples into images for machine learning to obtain the pre-trained model with generalization accuracy.

[0073] S107: Assimilate the numerical data by using a transfer learning method based on iterative neural operators, wherein the numerical data includes data from different image sizes and image layers.

[0074] Transfer learning is a special category of machine learning and data assimilation methods that provides the ability to improve the modeling accuracy of a dataset by leveraging a pre-trained model obtained from a similar dataset. In most cases, the source dataset has a large amount of data with a good distribution, while the target dataset may have very little data. The pre-trained model learns an initial mapping on the source dataset and is then fine-tuned on the target dataset. If the two learning tasks are similar, the transferred model will successfully inherit part of the internal structure of the pre-trained model and perform accurately on the target dataset. Existing transfer learning methods include model-based transfer learning (fine-tuning), instance-based transfer learning, and feature-based transfer learning. Iterative neural operator models can accurately capture the nonlinear features of effusion and predict the local effusion cooling effect of random pore configurations with high quality. Compared with direct machine learning methods, transfer learning significantly reduces the requirements for the number of training samples (reduced by 2-3 times) and training cycles (reduced by 5-6 times) to achieve the same accuracy. Simultaneously, with the same data cost and training cost, transfer learning reduces errors by 63.7%, 41.8%, and 46.2%, respectively.

[0075] The partial differential equation of the iterative neural operator used in this embodiment of the invention is as follows:

[0076]

[0077] The difference is expressed in the following form:

[0078]

[0079] , where u is the physical quantity to be solved, F is a nonlinear function of u and its derivative, Ω is the computational domain, and Γ is the boundary.

[0080] S109: Perform spatial, structural, and operational condition simulations respectively, and update the physical-virtual synchronization model.

[0081] A model-based transfer learning method is used to transfer a pre-trained model to experimental data. The pre-trained model is then fine-tuned to generate a transfer model. The learning rate for fine-tuning is set to 1e-4, and all convolutional layers in the pre-trained model are set to trainable during the fine-tuning process. During extrapolation, the iterative neural operator network model is trained using predetermined parameters for convolutional kernel size, number of hidden layers, and number of hidden state layers. The model weights are updated using the Adam optimizer.

[0082] Spatial extrapolation involves training an iterative neural operator network using incomplete images of the air film cooling efficiency. This allows the network to learn common physical laws in space and extrapolate the physical images of the incomplete parts, outputting the complete image of the air film cooling efficiency and recovering the image of the entire temperature field.

[0083] Structural deduction: The iterative neural operator network is trained using test samples with a number of holes and a hole arrangement that are completely different from the training set. This enables the iterative neural operator network to deduce the geometric layers with different hole distributions and positions and predict the gas film superposition phenomenon, and output a complete gas film cooling efficiency image to recover the image of the entire temperature field.

[0084] The operating condition simulation uses test samples with operating conditions completely different from the training set to train the iterative neural operator network. The iterative neural operator network learns the impact of operating conditions on the film cooling efficiency and deduces the physical picture under different operating conditions, outputting a complete film cooling efficiency picture and recovering the picture of the entire temperature field. Different operating conditions include: different air blowing ratios, different turbulence intensities, and different density ratios.

[0085] This invention presents a digital twin method for film cooling, encompassing data assimilation and extrapolation functions, supporting efficient and precise design of film cooling systems. This invention is applicable to film cooling problems with spatially and physically incomplete data, obtaining accurate global physical fields with minimal experimental information. A pre-trained model is obtained by performing machine learning on numerical data from CFD results. A transfer learning method based on iterative neural operators, driven by physical laws, is used to assimilate the numerical data, regressing it onto image data obtained from CFD simulations and experiments. The pre-trained model is then transferred to the experimental dataset using a model-based transfer learning method (fine-tuning). The digital twin system constructed using this method, with only 1% of the data known, extrapolates complete data that closely matches the experimental data, demonstrating significant effectiveness in assimilating incomplete data.

[0086] like Figure 2As shown in the figure, this invention provides a deep learning-based digital twin system for air-film cooling temperature fields. The system is constructed using a deep learning-based digital twin method for air-film cooling temperature fields. The system includes a model training module, a sensor module, a data acquisition module, a data processing module, a physical-virtual synchronization module, and a deduction module.

[0087] The model training module uses machine learning to pre-train the model to obtain a pre-trained model, and then fine-tunes the pre-trained model through transfer learning to generate a transfer model.

[0088] The sensor module includes a sensor interface, which, in conjunction with the data acquisition module, completes the acquisition of air film cooling test data.

[0089] The data acquisition module collects CFD simulation data and film cooling experimental data. After being processed by the data processing module, the above data is used as training data and test data for the model training module.

[0090] The data processing module processes the data acquired by the data acquisition module, including converting CFD simulation values ​​into images, assimilating numerical data, and performing regression processing on CFD simulation data and experimental data.

[0091] The physical-virtual synchronization module uses a model-based transfer learning method to transfer the pre-trained model to experimental data, fine-tune the pre-trained model, generate a transfer model, and achieve synchronization between the physical model and the virtual model.

[0092] The deduction module uses the model trained by the application model training module to deduce spatial, structural, and operational conditions, and update the physical-virtual synchronous model.

[0093] The hardware and software integrated deep learning-based digital twin system for film cooling temperature fields provided in this invention embodiment has multiple functions including spatial deduction, structural deduction, and operational condition deduction of film cooling physical information. Utilizing the model trained by this system, it can perform deductions of space, structure, and different operational conditions. This system uses incomplete images to supervise the network, from which the network learns common physical laws in space, thereby deducing the physical image of the incomplete parts and reconstructing the entire temperature field image. This system can also perform structural deduction. For samples in the test set with completely different numbers and arrangements of holes compared to the training set, the network can reconstruct the distribution of cooling efficiency well, demonstrating good generalization ability. When continuously moving a single film cooling hole along a certain direction, the network can effectively predict film cooling superposition phenomena, exhibiting good continuous response capability. Figure 8This is a structural simulation diagram of a preferred embodiment of the present invention. This system can perform operational condition simulations. For new operational conditions in the test set (such as different airflow ratios, turbulence intensities, density ratios, etc.), the network can accurately predict the film cooling efficiency. Figure 9 This is a working condition simulation diagram of a preferred embodiment of the present invention.

[0094] The present invention will now be described in detail with reference to preferred embodiments.

[0095] Example 1

[0096] like Figures 3-6 As shown in the preferred embodiment of the present invention, a digital twin method for air film cooling temperature field based on deep learning is provided, comprising the following steps:

[0097] Step 1: Construct a film cooling experimental setup, mainly including: a wind tunnel, a pressure vessel, a flow meter, a cryostat, an infrared thermal imager, and a measurement and data acquisition system, such as... Figure 7 As shown. In each experiment, a pressure vessel containing eight rods of compressed air enters the test rig, providing driving force for the main and secondary flows. The main air at ambient temperature passes through a series of valves 2 and a vortex flowmeter 3 before entering the test section. A honeycomb 4 is installed upstream of the measurement area to provide uniform airflow. The wind tunnel has a rectangular cross-section, 30 mm high and 30 mm wide.

[0098] Step 2: Set up the experimental setup and begin the experiment. The velocity of the main stream was further measured using a Pitot tube installed in the middle of the wind tunnel, verifying the measurement results of flow meter 3. Before measuring the pressure and temperature of the secondary flow, the secondary flow was cooled to -5°C using a cryostat 14, and compressor 1 was turned on, raising the supply pressure to 0.8 MPa. An infrared thermal imager 5 was installed and connected to an infrared image acquisition host computer (e.g., [missing information]). Figure 6 Computer 1), connected to the infrared image acquisition host computer (e.g., computer 1), uses a data transmission cable. Figure 6 The computer in the middle) and the deep learning host computer (such as Figure 6(Computer 2 in the middle). After installing the experimental orifice plate, the main and secondary flows were turned on, the flow rates were set, and the experiment began. A digital flow controller 15 was used to adjust the secondary air mass flow rate, with an adjustable range of 20–50 slm. Ni-Cr / Ni-Al k-type thermocouples were used to measure the temperature of the two airflows in real time and to collect data. A 3D-printed liquid accumulation test section was installed on the bottom of the wind tunnel and positioned and sealed using a 3D-printed static pressure box connected to the wind tunnel. The distance between the honeycomb 4 outlet and the upstream edge of the liquid accumulation cooling plate was 500 mm. The temperature distribution on the outer surface of the jet cooling plate was recorded using a FLIR A615 high-resolution infrared thermal imager 5. Black paint was sprayed on the surface of the experimental sample to improve the emissivity of the specimen. Before the experiment, the infrared thermal imager was calibrated using a heating plate with a temperature controller to control the temperature measurement error of the infrared thermal imager 5 within ±0.5%. The mixture of secondary air and main air was finally discharged into the environment from the wind tunnel outlet.

[0099] The general setup of the experiment is as follows: (1) The experiment was conducted at room temperature and pressure, with a back pressure of 1 atm for the main and secondary flows, and temperatures of approximately 299.60 K and 273.15 K for the main and secondary flows, respectively. (2) The material of the experimental sample was unsaturated polyester resin (8200) used in additive manufacturing, with a thermal conductivity of approximately 0.1 W·m⁻¹·K⁻¹. Due to the extremely low thermal conductivity of the experimental sample, the experimental conditions in this study can be approximated as adiabatic. Figure 7 This is a schematic diagram of the air film cooling experimental device.

[0100] Step 3: Data Acquisition. After the temperature field stabilizes, use an infrared camera to capture temperature images of the incomplete orifice plate and save them as a CSV file. Use data transfer software to transfer the CSV file to the deep learning computer, read the CSV file, and load the pre-trained model. For example... Figure 6 As shown, the infrared thermal imager 5 acquires infrared images in real time and transmits them to the infrared image acquisition host computer (e.g., [data cable]) via a data cable. Figure 6 The computer in the middle) and the deep learning host computer (such as Figure 6 Computer 2 in the middle).

[0101] Step 4: Construct a pre-trained model using CFD results as the pre-training data source. A plate with randomly distributed perforations was simulated under isothermal conditions. The rectangular cross-section of the main channel is Y = 20 mm × H = 50 mm. Coolant first flows into the static pressure tank and then into the main channel through overflow holes at a 30-degree angle. The length L of the perforation distribution area is 40 mm. The channel is extended by 40 mm upstream and downstream. The diameter D of the perforations is 1 mm. The geometric parameters are within the normal range for general turbine cooling designs. The number of perforations varies between 10 and 15. The perforation distribution follows several criteria and constraints: (i) perforations do not overlap, and the center-to-center distance is not less than twice the diameter; (ii) the perforation configuration has as much diversity as possible to ensure model generation; (iii) the perforations on the outer surface are within a rectangular area of ​​L = 40 mm × Y = 20 mm. Under these constraints, 54 film geometry models were generated for CFD simulation and machine learning.

[0102] The computational domain contains only the fluid domain. The elements are unstructured polyhedra. Prismatic meshes are generated in the boundary layer region. The first prism layer has a thickness of 0.0008 mm within the pores and 0.002 mm elsewhere, ensuring that the minimum dimensionless distance y+ of the fluid simulation mesh is less than 1 in most regions of the fluid domain. The thickness growth ratio is 1.2. To obtain a suitable mesh density, mesh independence was verified in the thermal fluid simulation. The test shell has 10 pores. Cell number sequences of 1.56 million, 2.19 million, 2.92 million, and 4.5 million were detected, respectively. The corresponding mesh substrate sizes are 0.84 mm, 0.67 mm, 0.53 mm, and 0.42 mm, respectively. The numerical dataset samples obtained are converted into images for machine learning to obtain a pre-trained model, such as... Figure 4 As shown.

[0103] Step 5: Assimilate Numerical Data. Since this model must be compatible with different data sizes (image size and image layers), we use a transfer learning method based on iterative neural operators driven by physical laws to assimilate the numerical data for regression on image data obtained from CFD simulations and experiments, such as... Figure 4 As shown. Film cooling is essentially a convective heat transfer problem, which can be comprehensively described by a set of convective-diffusion partial differential equations and their boundary conditions. The physical quantities obtained from solving these equations generally follow the laws determined by the PDE system:

[0104]

[0105] The difference is expressed in the following form:

[0106]

[0107] Where u is the physical quantity to be solved (i.e., the liquid cooling effect in this study), F is a nonlinear function of u and its derivative, Ω is the computational domain, and Γ is the boundary.

[0108] First, five-fold cross-validation is used to select suitable hyperparameters. A pre-trained model is then trained under these parameters to achieve good generalization accuracy. Finally, a model-based transfer learning method (fine-tuning) is used to transfer the pre-trained model to the experimental dataset. Figure 5 As shown, the learning rate for the pre-trained model and the non-transfer model is set to 1e-3 during training, while the learning rate for the fine-tuning process is set to 1e-4. During fine-tuning, all convolutional layers in the model are set to trainable.

[0109] Step 6: Perform spatial, structural, and operational condition simulations. The convolutional kernel size used in the simulations is 15*15, with 12 hidden layers and 7 hidden state layers. The iterative operator network model was compiled using the open-source deep learning framework TensorFlow 2.0. The neural network model was trained on an NVIDIA Quadro RTX 4000 GPU processor, with the Adam optimizer used to update the model weights during training.

[0110] In the spatial extrapolation section, the iterative operator network is trained using the acquired incomplete images of the film cooling efficiency. By using these incomplete images to supervise the network, it learns common physical laws in space and uses this to extrapolate the physical image of the incomplete parts. The network outputs a complete image of the film cooling efficiency, reconstructing the entire temperature field.

[0111] In the structural deduction section, the principle is that the samples in the training set exhibit variations in the number and position of holes. Therefore, the iterative operator network learns underlying physical laws during training, such as the low downstream temperature of the film cooling vents and the superposition of film cooling vents. Thus, by modifying the hole distribution and position in the geometric layer input to the network, the network can quickly predict the film cooling efficiency of the orifice plate. Even for samples in the test set with completely different hole numbers and arrangements compared to the training set, simply inputting their geometric layers into the iterative operator network allows the network to reconstruct the cooling efficiency distribution effectively, demonstrating good generalization ability. When continuously moving a single film cooling vent in a certain direction, the network can effectively predict film cooling superposition phenomena, exhibiting good continuous response capability. The steps for structural deduction are: after training the model using the training set, modify the geometric layer in the network input, load the trained weights, run the code, and the network predicts the cooling efficiency under different film cooling vent distributions. Figure 8 This is a schematic diagram illustrating the effect of structural deduction.

[0112] In the operating condition extrapolation section, the samples in the training set also include variations in operating conditions (such as different airflow ratios, turbulence intensities, and density ratios). The iterative operator network learns the impact of these operating conditions on film cooling efficiency during training. Therefore, by modifying the values ​​of the airflow ratio, turbulence intensities, and density ratio layers in the network input layer, the film cooling efficiency under these operating conditions can be quickly predicted. For operating conditions in the test set not included in the training set, the network can predict the film cooling efficiency well, demonstrating good generalization ability. The steps for operating condition extrapolation are: after training the model using the training set, modify the boundary condition layer in the network input, load the trained weights, run the code, and the network predicts the cooling efficiency under different operating conditions. Figure 9 This is a schematic diagram illustrating the effect of performing a working condition simulation.

[0113] Example 2

[0114] like Figure 2 As shown in the preferred embodiment of the present invention, a deep learning-based digital twin system for air-film cooling temperature fields is provided. This system encompasses the integration of hardware and software systems, data assimilation, and inference functions, supporting efficient and precise design of air-film cooling systems. It can be widely applied in design, manufacturing, operation, maintenance, and testing scenarios related to air-film cooling. The system is constructed using a deep learning-based digital twin method for air-film cooling temperature fields and includes a model training module, a sensor module, a data acquisition module, a data processing module, a physical-virtual synchronization module, and an inference module.

[0115] The model training module constructs a simulation model suitable for the temperature field of film cooling, uses image deep learning to quickly predict physical field information, uses machine learning to pre-train the model to obtain a pre-trained model, and uses transfer learning to fine-tune the pre-trained model to generate a transfer model.

[0116] The sensor module, which includes various sensors, works with the data acquisition module through the sensor interface to collect data from the film cooling test.

[0117] The data acquisition module collects CFD simulation data and air film cooling experimental data. The data is then processed by the data processing module to achieve compatibility of multi-source heterogeneous data. The processed data is used as training data and test data for the model training module.

[0118] The data processing module processes the data acquired by the data acquisition module, including converting CFD simulation values ​​into images, assimilating numerical data, and performing regression processing on CFD simulation data and experimental data.

[0119] The physical-virtual synchronization module uses a model-based transfer learning method to transfer the pre-trained model to experimental data, fine-tune the pre-trained model, generate a transfer model, and realize the synchronization of the physical model and the virtual model, thus completing the synchronization of software and hardware.

[0120] The deduction module uses the model trained by the application model training module to deduce spatial, structural, and operational conditions, and update the physical-virtual synchronous model.

[0121] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A deep learning-based digital twin method for temperature field generation in film cooling, characterized in that, The method includes the following steps: S101: Construct an air film cooling experimental device, the experimental device including sensors, measurement and data acquisition system, the experimental device completes the measurement and acquisition of experimental data; S103: Collect experimental data, use an infrared camera to capture and save the temperature image of the incomplete orifice plate. The experimental data is used to fine-tune the pre-trained model and generate a transfer model. S105: Train the pre-trained model by using CFD results as the pre-training data source to generate the pre-trained model; S107: Assimilate the numerical data using a transfer learning method based on iterative neural operators, wherein the numerical data includes data from different image sizes and image layers; S109: Perform spatial, structural, and operational condition simulations respectively, and update the physical-virtual synchronization model; in, In step S105, training the pre-trained model includes the following steps: S1051: Use CFD results as a pre-training data source; S1052: Selecting hyperparameters using five-fold cross-validation; S1053: The learning rate during training is set to 1e-3; S1054: Update the weights of the pre-trained model using the Adam optimizer; S1055: Convert numerical data samples into images for machine learning to obtain the pre-trained model with generalization accuracy; In step S1051, the gas film geometry is obtained by simulating a plate through which a randomly distributed perforation hole passes under isothermal conditions, and CFD simulation is performed using the gas film geometry to generate CFD numerical data. The computational domain of the CFD simulation only includes the fluid domain, the computational unit is an unstructured polyhedron, and a prismatic mesh is generated in the boundary layer region. The gas film geometry is generated under preset constraints, namely, simulating the passage of a flat plate by randomly distributed seepage holes under isothermal conditions. The length of the seepage hole distribution area is 40 mm, the diameter of the seepage hole is 1 mm, and the number of seepage holes is 10 to 15. The seepage holes are distributed according to predetermined conditions, including: the holes do not overlap, the center-to-center distance of the holes is not less than twice the diameter of the holes, the hole configuration is diverse, and the holes on the outer surface are distributed within a rectangular area of ​​40 mm × 20 mm. In step S109, the inference uses predetermined parameters such as convolutional kernel size, number of hidden layers, and number of hidden state layers to train the iterative neural operator network model, and updates the model weights using the Adam optimizer. The spatial extrapolation uses incomplete air film cooling efficiency images to train the iterative neural operator network model, enabling the iterative neural operator network model to learn the common physical laws of space, extrapolate the physical image of the incomplete part, output the complete air film cooling efficiency image, and recover the image of the entire temperature field. The structural deduction uses test samples with a number of holes and a hole arrangement completely different from the training set to train the iterative neural operator network model, so that the iterative neural operator network model can deduce the geometric layers with different hole distributions and positions and predict the gas film superposition phenomenon, and output a complete gas film cooling efficiency image to recover the image of the entire temperature field. The operating condition simulation uses test samples with operating conditions completely different from the training set to train the iterative neural operator network model. The iterative neural operator network model learns the influence of operating conditions on the film cooling efficiency and deduces the physical picture under different operating conditions, outputs a complete film cooling efficiency picture, and recovers the picture of the entire temperature field. The operating conditions include: different blowing ratios, different turbulence intensities, and different density ratios.

2. The deep learning-based digital twin method for air film cooling temperature field as described in claim 1, characterized in that, In step S101, the air-film cooling experimental device includes a wind tunnel, a pressure vessel, a flow meter, a cryostat, an infrared thermal imager, and a measurement and data acquisition system. The wind tunnel has a rectangular cross-section, and a honeycomb structure is installed upstream of the measurement area to provide uniform airflow. The pressure vessel provides driving force for the main flow and secondary flow. The flow meter is installed in the middle of the wind tunnel to measure the velocity of the main flow. Before pressure and temperature measurements are performed, the secondary flow is cooled to -5°C by the cryostat. The infrared thermal imager records the temperature distribution on the outer surface of the jet cooling plate. The measurement and data acquisition system uses thermocouples to measure the temperature of the main flow and secondary flow in real time and acquires the data. The mixture of the main flow and the secondary flow is finally discharged into the environment from the wind tunnel outlet.

3. The deep learning-based digital twin method for air film cooling temperature field as described in claim 2, characterized in that, In step S103, the experimental data is obtained under predetermined experimental conditions, which include: the experiment is conducted at room temperature and pressure, the back pressure of the main flow and the secondary flow is 1 atm, the temperature of the main flow is 299.60 K, and the temperature of the secondary flow is 273.15 K; the experimental sample material is an additive manufacturing material, the additive manufacturing material is an unsaturated polyester resin, and the thermal conductivity of the material is 0.1 W / m / K.

4. The deep learning-based digital twin method for air film cooling temperature field as described in claim 1, characterized in that, In step S107, the partial differential equation of the iterative neural operator is: ， The difference is expressed in the form of: ， Where u is the physical quantity to be solved, F is a nonlinear function of u and its derivative, Ω is the computational domain, and Γ is the boundary value. For time, The first spatial derivative of the physical field. The second spatial derivative of the physical field. The physical field at the current time. for The physical field after a certain time, for film cooling, is any one of film cooling effect, temperature, velocity, and pressure.

5. The deep learning-based digital twin method for air film cooling temperature field as described in claim 1, characterized in that, In step S107, the pre-trained model is transferred to the experimental data using a model-based transfer learning method. The pre-trained model is then fine-tuned to generate the transfer model. The learning rate for the fine-tuning process is set to 1e-4, and all convolutional layers in the pre-trained model are set to trainable during the fine-tuning process.

6. A deep learning-based digital twin system for air-film cooling temperature fields, characterized in that, The system is constructed using the digital twin method as described in any one of claims 1-5, and includes a model training module, a sensor module, a data acquisition module, a data processing module, a physical-virtual synchronization module, and a deduction module; wherein, The model training module uses machine learning to pre-train the model to obtain a pre-trained model, and then fine-tunes the pre-trained model through transfer learning to generate a transfer model. The sensor module includes a sensor interface, through which the data acquisition module completes the acquisition of air film cooling test data. The data acquisition module acquires CFD simulation data and film cooling experimental data. The film cooling experimental data is processed by the data processing module and used as training data and test data for the model training module. The data processing module processes the data acquired by the data acquisition module. The processing includes converting the CFD simulation values ​​into images, assimilating the numerical data, and performing regression processing on the CFD simulation data and the experimental data. The physical-virtual synchronization module uses a model-based transfer learning method to transfer the pre-trained model to the experimental data, fine-tunes the pre-trained model, and generates the transfer model, thereby achieving synchronization between the physical model and the virtual model. The inference module uses the model trained by the model training module to infer space, structure, and working conditions, and updates the physical-virtual synchronous model.

Images