CN116467960A - Deep learning-based digital twin method and system for air film cooling temperature field - Google Patents

Abstract

The invention discloses a deep learning-based digital twin method and a deep learning-based digital twin system for a film cooling temperature field, which relate to the technical field of thermal protection, and comprise the following steps: constructing an air film cooling experimental device, wherein the experimental device comprises a sensor, a measuring and data acquisition system, and measuring and acquiring experimental data are completed; collecting experimental data, shooting and storing incomplete pore plate temperature images by using an infrared camera, wherein the experimental data are used for fine tuning a pre-training model to generate a migration model; training a pre-training model, and generating the pre-training model by adopting a CFD result as a pre-training data source; assimilating numerical data, the numerical data comprising data from different image sizes and image layers, using an iterative neural operator based transfer learning method; and respectively carrying out space, structure and working condition deduction, and updating the physical-virtual synchronous model. The invention has the functions of data assimilation and deduction, and can support the efficient and accurate design of air film cooling.

Description

Deep learning-based digital twin method and system for air film cooling temperature field

Technical Field

The invention relates to the technical field of thermal protection, in particular to a deep learning-based digital twin method and system for a film cooling temperature field.

Background

The intellectualization of design, manufacture, detection and operation and maintenance is a development trend of heat protection technology. For the correlation/experience model, the defects are high dimension, low precision; however, for the experimental/CFD (computational fluid dynamics) method, the disadvantage is high cost and long time consumption; therefore, a reliable high-dimensional and high-precision prediction model is established, the rapid prediction of multiple physical fields is realized, and the design and optimization of aerospace thermal protection are necessary.

Gas turbine systems require higher operating temperatures for higher efficiencies. The turbine inlet temperature of the current gas turbine engines is far beyond the tolerable range of the parallel superalloy materials. Therefore, it becomes critical to develop superalloy/ceramic materials and to apply active cooling techniques to protect the hot section of the gas turbine system. The cooling technology currently practiced generally integrates various internal and external cooling methods. Among these methods, film cooling is one of the most widely used external cooling methods. The coolant air is discharged through the grooves or the discrete holes, and a coolant air film is formed on the outer surface of the hot component, so that the hot erosion of the turbine blade by the high-temperature gas is avoided, and the purpose of cooling protection is achieved.

The adiabatic wall cooling effect is the main criterion for evaluating film cooling. Film cooling efficiency is affected by a number of factors such as the ratio of blowing, the density ratio, the turbulence intensity of the main stream, the hole shape and the hole position. In order to achieve higher cooling performance, accurate prediction of local temperature details is required, and film cooling parameters are optimized. However, due to the complex interactions between the coolant jet and the main flow, it is difficult to understand the film cooling flow field and its corresponding cooling efficiency profile. The current feasibility assessment of film cooling is still based on experiments, with the aid of established non-contact measurement techniques such as pressure sensitive coatings, infrared thermal imaging and liquid crystals. In order to alleviate the above-mentioned challenges, to obtain a two-dimensional effectiveness distribution of high fidelity and fast response, some of the methods currently proposed in film cooling modeling have the following advantages and disadvantages:

1) One of the most widely used correlations in predicting the cooling effect of a single row of holes is the one-dimensional correlation of Baldauf, followed by the one-dimensional correlation, and the earliest proposed superposition calculation model for film cooling slots by Sellers. The disadvantage is that the Baldauf correlation cannot accurately predict the effectiveness of different types of film cooling; the superposition equations of the Sellers cannot take into account the details of the local conditions or local geometry. Simple explicit equations are not sufficient to model complex surface temperature distributions, even lateral averages, of the superimposed film cooling structure. In addition to the accuracy problem, these correlations are also of low dimensionality and can only provide one-dimensional results for regularly arranged film cooling holes. With random distribution of discrete holes, there is little correlation that can predict the two-dimensional cooling efficiency distribution of a surface. This is due to the extremely high complexity and nonlinearity of the randomly distributed flow fields that interact and overlap from upstream to downstream. Particularly when considering high temperature and high pressure conditions, more nonlinear characteristics such as changes in gas/air physical properties should be involved in the problem, and the response of the cooling effect becomes more complicated.

2) In terms of CFD, research into cross-flow jet phenomena has been conducted in recent decades to demonstrate that there are great difficulties in simulating near-wall turbulence in film cooling structures. Anisotropic turbulent diffusion is considered to be an important mechanism which cannot be simulated by the traditional two-party Cheng Tuanliu method (such as a shear stress transport model). Meanwhile, inaccuracy exists in the calculation of the Plandter number, and the inverse gradient scalar diffusion mechanism is abnormal and exceeds the capability of a diffusion model. Research attempts to solve these film cooling prediction problems are generally based on a priori knowledge of turbulence and analytical deductions, with a low degree of integration with the data. The prior studies following this technical route were from Li et al. To obtain a more accurate model of reynolds stress and turbulent scalar flux, they propose an algebraic anisotropic turbulent model and verify it by leading edge zone showerhead divergence and full coverage divergent vane experiments. By comparing four different turbulence models, they demonstrate that algebraic anisotropy models have better consistency in nature and quantification with experimental data. Subsequent studies by Li et al have proposed a modified algebraic anisotropic vortex viscosity method (AAEV) which takes into account the influence of wall and mean flow field strain on the anisotropy ratio. The new AAEVk-epsilon model correctly predicts the spanwise expansion of the membrane and the attenuation of the secondary vortex.

3) In terms of machine learning, while turbulence modeling is an unavoidable issue of computational fluid dynamics, data driven methods can circumvent this difficulty at the expense of data. The deep learning approach presented in 2012 has shown great potential in regressing many physical fields with images or matrix formats and is naturally compatible with film cooling prediction. Under the guarantee of the general approximation theorem, researchers try to establish a mapping relation between the geometric shape and the air film cooling effect. The results show that the deep learning model can identify abstract features in the dataset with high accuracy and good generalization ability. The main advantage of these deep learning methods is that film cooling prediction speeds are extremely fast, typically 105 times faster than CFD methods. However, also because numerical data is used to train the deep learning model, there is still concern about the accuracy of film cooling because the regression model does not have higher accuracy than the data driving it. Possible methods lie in experimental data or LES data. Unfortunately, this higher fidelity data is far from adequate for training deep neural networks, both for the amount and distribution of film cooling data.

4) In the data fusion method of the hot fluid problem, one potential technology for improving the modeling precision of the air film cooling is to integrate simulation data with experimental data, which is called data assimilation. Such methods are typically based on existing partially resolved physics, such as the Navier-Stokes equations, and model several parameters in these equations with experimental data. Li et al propose a continuous companion data assimilation model that successfully reconstructs global turbulent flow from a limited number of wall pressure measurements. The model has good prediction effect on high-Reynolds number separation flow, strong reverse pressure gradient flow and complex flow. They introduced an anisotropic concomitant data assimilation scheme supplementing the limited spatial range and measurable turbulent average flow measurement data. They achieved good results for complex flow field reconstruction of circular jet flow fields, blunt plate flow fields and rib wall flow fields. Deng et al used an integrated-Kalman filter based data assimilation method to optimize the RANS model constants to recover the global flow field from the local measurement data. They validated the effect of the data assimilation method for four different RANS models, respectively. All study models using the data assimilation method performed significantly better with the k-k model. Although the data assimilation method can be used as a good tool for improving the air film prediction precision, two major disadvantages still exist, namely a) experimental air film cooling data are generally interface temperature (effectiveness), and a large amount of simulation data are in main flow, so that data distribution is not matched; b) The data assimilation method still needs to solve the Navier-Stokes equation, which is too slow to provide immediate response to the designer. The solution may still be relevant to machine learning.

Transfer learning is a special class of machine learning and data assimilation methods that provides the ability to improve the accuracy of modeling of a data set by means of pre-trained models obtained from similar data sets. In most cases, the source data set has a large data volume and a good data distribution, while the target data set may have very little data. The pre-trained model learns an initial mapping on the source dataset and then fine-tunes on the target dataset. If the two learning tasks are similar, the transferred model will successfully complete part of the internal structure of the pre-trained model and execute accurately in the target dataset. Existing migration learning methods include model-based migration learning (fine tuning), instance-based migration learning, and feature-based migration learning.

In recent studies on transfer learning and physics, several publications on hot fluid problems have revealed the potential of transfer learning to achieve high accuracy on small data sets. Wang et al propose a condition-based generation of a transition learning model against a network (CGAN) for predicting the three-dimensional pressure distribution of the turbine blade surface. Knowledge learned from the large-scale low-fidelity dataset by the model is transferred to the small-scale high-fidelity dataset. The results show that transfer learning effectively improves generalization accuracy on small-scale data sets. The Obiols-Sales et al introduced a super-resolution streaming network (SURFNet) that accelerated high-resolution turbulent CFD simulation. SURFNet is trained in low resolution analog data and then fine tuned through a high resolution dataset. The model exhibits good resolution invariance and generalization ability at the cost of 1/15 of the training data.

Accordingly, those skilled in the art have focused their efforts on developing a deep learning based film cooling temperature field digital twin method and system.

Disclosure of Invention

In view of the above-mentioned drawbacks of the prior art, the present invention provides a method for digitally twinning an experimental air film cooling temperature field based on deep learning, which is characterized in that the method includes the following steps:

s101: constructing an air film cooling experimental device, wherein the experimental device comprises a sensor, a measuring system and a data acquisition system, and the experimental device is used for completing measurement and acquisition of experimental data;

s103: collecting experimental data, shooting and storing incomplete pore plate temperature images by using an infrared camera, wherein the experimental data are used for fine tuning a pre-training model to generate a migration model;

s105: training the pre-training model, and generating the pre-training model by adopting a CFD result as a pre-training data source;

s107: assimilating numerical data, the numerical data comprising data from different image sizes and image layers, using an iterative neural operator based transfer learning method;

S109: and respectively carrying out space, structure and working condition deduction, and updating the physical-virtual synchronous model.

Further, in the step S101, the film cooling experimental device includes a wind tunnel, a pressure vessel, a flowmeter, a cryostat, a thermal infrared imager, and a measurement and data acquisition system; the wind tunnel is rectangular in cross section, and a honeycomb is arranged at the upstream of a measuring area of the wind tunnel so as to provide uniform airflow; the pressure vessel providing a driving force for the primary and secondary streams; the flowmeter is arranged in the middle of the wind tunnel and measures the speed of the main flow; the secondary stream is cooled by the cryostat to-5 ℃ prior to taking pressure and temperature measurements; the infrared thermal imager records the temperature distribution of the outer surface of the jet cooling plate, the measuring and data acquisition system adopts a thermocouple to measure the temperature of the main flow and the secondary flow in real time and acquire data, and the mixture of the main flow and the secondary flow is finally discharged into the environment from the wind tunnel outlet.

Further, in the step S103, the experimental data is obtained under predetermined experimental conditions including: the experiment is carried out at normal temperature and pressure, the back pressure of the main flow and the secondary flow is 1atm, the temperature of the main flow is 299.60K, and the temperature of the secondary flow is 273.15K; the experimental sample material is an additive manufacturing material, the additive manufacturing material is unsaturated polyester resin, and the thermal conductivity coefficient of the material is 0.1W/m/K.

Further, in the step S105, the training pre-training model includes the steps of:

s1051: using CFD results as a source of pre-training data;

s1052: selecting proper super parameters by utilizing five-fold cross validation;

s1053: the learning rate in the training process is set to be 1e-3;

s1054: updating the weight of the pre-training model by adopting an Adam optimizer;

s1055: and converting the numerical data sample into an image for machine learning to obtain the pre-training model with generalization precision.

Further, in the step S1051, a gas film geometry is obtained by simulating a flat plate through which randomly distributed effusion holes pass under isothermal conditions, and CFD numerical data is generated by performing CFD simulation using the gas film geometry.

Further, the computational domain of the CFD simulation only comprises a fluid domain, the computational unit is an unstructured polyhedron, and the boundary layer region generates a prismatic grid; the gas film geometric figure is generated under a preset constraint condition, wherein the constraint condition is that exudation holes which are randomly distributed under an isothermal condition pass through a flat plate in a simulation mode, the length of a distribution area of the exudation holes is 40mm, the diameter of the exudation holes is 1mm, the number of the exudation holes is 10-15, and the exudation holes are distributed according to a preset condition; wherein the predetermined condition includes: the holes are not overlapped, and the center distance of the holes is not less than 2 times of the diameter of the holes; the hole configuration of the holes is diversified; the holes of the outer surface are distributed in a rectangular area of 40mm x 20 mm.

Further, in the step S107, the partial differential equation of the iterative nerve operator is:

，

the difference form is expressed as:

，

where u is the physical quantity of the desired solution, F is a nonlinear function of u and its derivatives, Ω is the computational domain, and Γ is the boundary.

Further, in the step S109, the pre-training model is migrated to the experimental data by a model-based migration learning method, the pre-training model is fine-tuned to generate the migration model, the learning rate of the fine-tuning process is set to 1e-4, and all convolution layers in the pre-training model in the fine-tuning process are set to be trainable.

Further, the deduction uses a predetermined convolution kernel size, number of hidden layers and hidden state diagram layer number parameters to train the iterative neural operator network model, and model weights are updated through an Adam optimizer, wherein,

the space deduction is carried out, the iterative neural operator network is trained by using the incomplete air film cooling efficiency image, so that the iterative neural operator network learns the physical rule of space commonality, the physical image of the incomplete part is deduced, the complete air film cooling efficiency image is output, and the image of the whole temperature field is restored;

The structure deduction is carried out, the iterative neural operator network is trained by using test samples with completely different hole numbers and hole arrangement modes from a training set, so that the iterative neural operator network can be used for deducing geometrical layers with different hole distribution and positions and predicting the phenomenon of air film superposition, outputting a complete air film cooling efficiency image and recovering the image of the whole temperature field;

the working condition deduction is carried out, a test sample with completely different working condition conditions and a training set is used for training the iterative neural operator network, so that the iterative neural operator network learns the influence of the working condition on the air film cooling efficiency, physical images under different working condition conditions are deduced, a complete air film cooling efficiency image is output, and the image of the whole temperature field is restored; the working conditions comprise: different blowing ratios, different turbulence intensities and different density ratios.

On the other hand, the invention also provides a deep learning-based air film cooling temperature field digital twin system, which is characterized in that the system is constructed by adopting a deep learning-based air film cooling temperature field digital twin method, and the system comprises 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 is used for pre-training the model by utilizing machine learning to obtain a pre-training model, and performing fine adjustment on the pre-training model through transfer learning to generate a transfer model;

the sensor module comprises a sensor interface, and the acquisition of the air film cooling test data is completed by matching with the data acquisition module through the sensor interface;

the data acquisition module acquires CFD simulation data and the air film cooling experimental data, and the data are processed by the data processing module and then serve as training data and test data of the model training module;

the data processing module is used for processing the data acquired by the data acquisition module, wherein the processing comprises the steps of converting the CFD simulation numerical value into an image, assimilating the numerical value data and carrying out regression processing on the CFD simulation data and the experimental data;

the physical-virtual synchronization module migrates the pre-training model to the experimental data through a model-based migration learning method, fine-tunes the pre-training model, generates the migration model, and realizes synchronization of a physical model and a virtual model;

and the deduction module is used for deducting space, structure and working condition by applying the model trained by the model training module and updating the physical-virtual synchronous model.

In the preferred embodiment of the invention, compared with the prior art, the invention has the following beneficial effects:

1. the digital twin method suitable for air film cooling has the functions of data assimilation and deduction, and can support the efficient and accurate design of air film cooling;

2. the iterative nerve operator model designed by the invention can accurately capture the nonlinear characteristics of effusion and can predict the local effusion cooling effect of random hole configuration with high quality. Compared with a direct machine learning method, under the condition of reaching the same precision, the transfer learning obviously reduces the requirements on the number of training samples (reduced by 2-3 times) and the training period (reduced by 5-6 times), and under the condition of the same data cost and training cost, the errors of the transfer learning are respectively reduced by 63.7%, 41.8% and 46.2%.

The conception, specific structure, and technical effects of the present invention will be further described with reference to the accompanying drawings to fully understand the objects, features, and effects of the present invention.

Drawings

FIG. 1 is a flow chart of a preferred embodiment of the present invention;

FIG. 2 is a schematic diagram of a build framework in accordance with a preferred embodiment of the present invention;

FIG. 3 is a schematic diagram showing the steps for implementing a preferred embodiment of the present invention;

FIG. 4 is a schematic diagram showing steps for assimilating numerical data by an iterative neural operator-based migration learning method according to a preferred embodiment of the present invention;

FIG. 5 is a schematic diagram showing a transfer learning procedure according to a preferred embodiment of the present invention;

FIG. 6 is a schematic diagram of the synchronization of software and hardware in a preferred embodiment of the invention;

FIG. 7 is a schematic diagram of an exemplary film cooling experimental setup according to the present invention;

FIG. 8 is a structural deduction effect diagram of a preferred embodiment of the present invention;

FIG. 9 is a graph showing the effect of working condition deduction according to a preferred embodiment of the present invention.

Wherein: 1-compressor, 2-valve, 3-flowmeter, 4-honeycomb, 5-infrared thermal imaging system, 6-infrared transparent glass, 7-air flow 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-cryostat, 15-flow controller, 16-pressure regulator.

Detailed Description

The following description of the preferred embodiments of the present invention refers to the accompanying drawings, which make the technical contents thereof more clear and easy to understand. The present invention may be embodied in many different forms of embodiments and the scope of the present invention is not limited to only the embodiments described herein.

In the drawings, like structural elements are referred to by like reference numerals and components having similar structure or function are referred to by like reference numerals. The dimensions and thickness of each component shown in the drawings are arbitrarily shown, and the present invention is not limited to the dimensions and thickness of each component. The thickness of the components is exaggerated in some places in the drawings for clarity of illustration.

As shown in fig. 1, the embodiment of the invention provides a film cooling temperature field digital twin method based on deep learning, which comprises the following steps:

s101: and constructing a gas film cooling experimental device, wherein the experimental device is used for completing measurement and acquisition of experimental data.

The air film cooling experimental device comprises a wind tunnel, a pressure container, a flowmeter 3, a cryostat 14, a thermal infrared imager 5 and a measuring and data acquisition system, as shown in fig. 7. Wherein the wind tunnel has a rectangular cross section, and a honeycomb 4 is arranged at the upstream of the measuring area of the wind tunnel to provide uniform airflow; the pressure vessel provides driving force for the primary and secondary streams; the flowmeter 3 is arranged in the middle of the wind tunnel and measures the speed of the main flow; the secondary stream is cooled to-5 ℃ by cryostat 14 prior to taking pressure and temperature measurements; the infrared thermal imager 5 records the temperature distribution of the outer surface of the jet cooling plate, the measuring and data acquisition system adopts a thermocouple to measure the temperature of the main flow and the secondary flow in real time, data acquisition is carried out, and the mixture of the main flow and the secondary flow is finally discharged into the environment from a wind tunnel outlet.

S103: and acquiring experimental data, shooting and storing incomplete pore plate temperature images by using an infrared camera, wherein the experimental data is used for fine tuning a pre-training model to generate a migration model.

The above experimental data are obtained at predetermined experimental conditions including: the experiment is carried out at normal temperature and pressure, the back pressure of the main flow and the secondary flow is 1atm, the temperature of the main flow is 299.60K, and the temperature of the secondary flow is 273.15K; the experimental sample material was an additive manufacturing material that was an unsaturated polyester resin with a thermal conductivity of 0.1W/m/K.

S105: training a pre-training model, and generating the pre-training model by adopting a CFD result as a pre-training data source.

When training the pre-training model, the method comprises the following steps:

s1051: CFD results were used as a source of pre-training data: and (3) obtaining a gas film geometric figure by simulating a flat plate penetrated by a random distribution of exudation holes under an isothermal condition, and performing CFD simulation by using the gas film geometric figure to generate CFD numerical data, wherein the CFD numerical data is used as a pre-training data source. The computational domain of the CFD simulation only comprises a fluid domain, the computational unit is an unstructured polyhedron, and the boundary layer region generates a prismatic grid; the film geometry is generated under the following constraints: the exudation holes randomly distributed under isothermal conditions were simulated to pass through the plate, the distribution area length of the exudation holes was 40mm, the diameter of the exudation holes was 1mm, and the number of the exudation holes was 10-15, and the above exudation holes were distributed according to the following conditions: the holes are not overlapped, and the center distance of the holes is not less than 2 times of the diameter of the holes; the hole configuration of the holes is diversified; the holes of the outer surface are distributed in a rectangular area of 40mm x 20 mm.

S1052: selecting proper super parameters by utilizing five-fold cross validation;

s1053: the learning rate in the training process is set to be 1e-3;

s1054: updating the weight of the pre-training model by adopting an Adam optimizer;

s1055: and converting the numerical data sample into an image for machine learning to obtain the pre-training model with generalization precision.

S107: assimilating numerical data, including data from different image sizes and image layers, using an iterative neural operator based transfer learning method.

Transfer learning is a special class of machine learning and data assimilation methods that provides the ability to improve the accuracy of modeling of a data set by means of pre-trained models obtained from similar data sets. In most cases, the source data set has a large data volume and a good data distribution, while the target data set may have very little data. The pre-trained model learns an initial mapping on the source dataset and then fine-tunes on the target dataset. If the two learning tasks are similar, the transferred model will successfully complete part of the internal structure of the pre-trained model and execute accurately in the target dataset. Existing migration learning methods include model-based migration learning (fine tuning), instance-based migration learning, and feature-based migration learning. The iterative nerve operator model can accurately capture the nonlinear characteristics of effusion and predict the local effusion cooling effect of random hole configuration with high quality. Transfer learning significantly reduces the number of training samples (2-3 times less) and training period (5-6 times less) requirements to achieve the same accuracy as compared to the direct machine learning method. Meanwhile, under the condition that the data cost and the training cost are the same, the errors of transfer learning are reduced by 63.7%, 41.8% and 46.2%, respectively.

The partial differential equation of the iterative nerve operator used in the embodiment of the invention is as follows:

，

the difference form is expressed as:

，

where u is the physical quantity of the desired solution, F is a nonlinear function of u and its derivatives, Ω is the computational domain, and Γ is the boundary.

S109: and respectively carrying out space, structure and working condition deduction, and updating the physical-virtual synchronous model.

And migrating the pre-trained model to experimental data through a model-based migration learning method, finely adjusting the pre-trained model to generate a migration model, setting the learning rate of the fine adjustment process to be 1e-4, and setting all convolution layers in the pre-trained model in the fine adjustment process to be trainable. Training the iterative neural operator network model by using a preset convolution kernel size, the number of hidden layers and hidden state diagram layer number parameters when deduction is carried out, updating model weights by an Adam optimizer, wherein,

space deduction, training an iterative neural operator network by using the incomplete air film cooling efficiency image, so that the iterative neural operator network learns the physical rule of space commonality, deducting the physical image of the incomplete part, outputting the complete air film cooling efficiency image, and recovering the image of the whole temperature field;

The structure deduction is carried out, and a test sample with completely different hole numbers and hole arrangement modes from a training set is used for training an iterative neural operator network, so that the iterative neural operator network can be used for deducing the superposition phenomenon of geometric layers and predicted air films with different hole distributions and positions, outputting a complete air film cooling efficiency image and recovering the image of the whole temperature field;

working condition deduction, training an iterative neural operator network by using a test sample with completely different working condition and training set, so that the iterative neural operator network learns the influence of the working condition on the air film cooling efficiency, deducts physical images under different working condition, outputs a complete air film cooling efficiency image and restores the image of the whole temperature field; different working conditions include: different blowing ratios, different turbulence intensities and different density ratios.

The digital twin method suitable for air film cooling, which is designed by the embodiment of the invention, covers the functions of data assimilation and deduction, and can support the efficient and accurate design of air film cooling. The method is suitable for the problem of air film cooling of spatial incomplete data and physical incomplete data, and an accurate global physical field is obtained through little actual measurement information. Obtaining a pre-training model by machine learning numerical data of the CFD result; numerical data is assimilated using an iterative neural operator-based transfer learning method driven by laws of physics for regression of image data obtained from CFD simulations and experiments. The pre-trained model is migrated to the experimental dataset by a model-based migration learning method (fine tuning). The digital twin system constructed by the method basically coincides with the data obtained by experiments under the condition that 1% of data is known, and has remarkable effect on assimilation ability of incomplete data.

As shown in fig. 2, the embodiment of the invention provides a deep learning-based digital twin system for a gas film cooling temperature field, which is constructed by adopting a deep learning-based digital twin method for the gas film cooling temperature field, and comprises 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 is used for pre-training the model by utilizing machine learning to obtain a pre-training model, and performing fine adjustment on the pre-training model through transfer learning to generate a transfer model;

the sensor module comprises a sensor interface, and the acquisition of the air film cooling test data is completed through the sensor interface and the data acquisition module;

the data acquisition module acquires CFD simulation data and air film cooling experimental data, and the data are processed by the data processing module and then serve as training data and test data of the model training module;

the data processing module is used for processing the data acquired by the data acquisition module, and comprises the steps of converting CFD simulation values into images, assimilating the data, and carrying out regression processing on the CFD simulation data and experimental data;

the physical-virtual synchronization module is used for migrating the pre-training model to experimental data through a model-based migration learning method, fine-tuning the pre-training model, generating a migration model and realizing synchronization of the physical model and the virtual model;

And the deduction module is used for deducting the space, the structure and the working condition by applying the model trained by the model training module and updating the physical-virtual synchronous model.

The software and hardware integrated deep learning-based air film cooling temperature field digital twin system provided by the embodiment of the invention has multiple functions of air film cooling physical information space deduction, structure deduction and working condition deduction, and the deduction of space, structure and different working conditions is realized by utilizing a model trained by the system. The system monitors the network by using the incomplete image, and the network learns the physical rule of space commonality from the network, so as to deduce the physical image of the incomplete part and restore the image of the whole temperature field. The system can also realize structure deduction, and for samples with completely different numbers and arrangement modes of holes in the test set and training sets, the network can reconstruct the distribution of cooling efficiency well, and the system has good generalization capability. When a certain air film hole is continuously moved along a certain direction, the network can effectively predict the air film superposition phenomenon, and has good continuous response capability, and fig. 8 is a structure deduction effect diagram of a preferred embodiment of the invention. The system can realize working condition deduction, and for some new working conditions (such as different blowing ratios, turbulence intensity, density ratios and the like) in the test set, the network can well predict the cooling efficiency of the air outlet film, and fig. 9 is a working condition deduction effect diagram of a preferred embodiment of the invention.

The present invention will be described in detail with reference to preferred embodiments thereof.

Example 1

As shown in fig. 3 to 6, the method for digitally twinning a film cooling temperature field based on deep learning provided by the preferred embodiment of the invention comprises the following steps:

step 1: a set of air film cooling experimental device is constructed, which mainly comprises a wind tunnel, a pressure vessel, a flowmeter 3, a cryostat 14, a thermal infrared imager 5 and a measuring and data acquisition system, as shown in figure 7. In each experiment, a pressure vessel containing 8 rods of compressed air was entered into the test bed to provide driving force for the primary and secondary flows. The main stream air at ambient temperature is passed through a series of valves 2 and a vortex flowmeter 3 before entering the test section. Upstream of the measuring area a honeycomb 4 is installed to provide a uniform air flow. The wind tunnel section is rectangular, and has a height of 30 mm and a width of 30 mm.

Step 2: and (5) preparing an experimental device and starting an experiment. The velocity of the main flow was further measured by a pitot tube installed in the middle of the wind tunnel, and the measurement result of the flow meter 3 was verified. Before measuring the pressure and temperature of the secondary stream, the secondary stream was cooled to-5℃by means of a cryostat 14, the compressor 1 was started and the gas supply pressure was increased to 0.8MPa. The thermal infrared imager 5 is installed and connected to an infrared image acquisition upper computer (such as a computer 1 in fig. 6), and the infrared image acquisition upper computer (such as the computer 1 in fig. 6) and the deep learning upper computer (such as the computer 2 in fig. 6) are connected by using a data transmission line. After the experimental hole plate is installed, the main flow and the secondary flow are started, the flow is set, and the experiment is started. The mass flow of the secondary air is regulated by adopting a digital flow controller 15, and the adjustable range is 20-50 slm. And (3) measuring the temperatures of the two airflows in real time by adopting a Ni-Cr/Ni-Al k type thermocouple, and acquiring data. And installing a 3d printing effusion test section on the bottom surface of the wind tunnel, and positioning and sealing by using a 3d printing static pressure box connected with the wind tunnel. The distance between the outlet of the honeycomb 4 and the upstream edge of the liquid-collecting cooling plate was 500mm. The jet cooling plate outer surface temperature profile was recorded using FLIR a615 high resolution thermal infrared imager 5. And spraying blackbody paint on the surface of the experimental sample to improve the emissivity of the test piece. Before the experiment, the thermal infrared imager is calibrated by adopting a heating plate with a temperature controller, so that the temperature measurement error of the thermal infrared imager 5 is controlled within +/-0.5 percent. The mixture of secondary and primary wind is eventually discharged from the wind tunnel outlet into the environment.

The experimental part is generally set up as follows (1) the experiment is carried out at normal temperature and pressure, the back pressure of the main stream and the secondary stream is 1atm, and the temperatures of the main stream and the secondary stream are about 299.60K and 273.15K respectively. (2) The material of the experimental sample is an additive manufacturing material unsaturated polyester resin (8200), and the heat conductivity coefficient is about 0.1 W.m < -1 > K < -1 >. Because the thermal conductivity of the experimental sample is extremely low, the experimental conditions of the study can be approximately regarded as thermal insulation, and fig. 7 is a schematic diagram of the air film cooling experimental device.

Step 3: data is collected. After the temperature field is stable, an infrared camera is used for shooting incomplete pore plate temperature images, and the incomplete pore plate temperature images are stored as CSV files. And transmitting the CSV file to a deep learning computer by using data transmission software, reading the CSV file, and loading the pre-training model. As shown in fig. 6, the thermal infrared imager 5 acquires infrared images in real time, and transmits the images to an infrared image acquisition upper computer (such as a computer 1 in fig. 6) and a deep learning upper computer (such as a computer 2 in fig. 6) through data lines.

Step 4: and constructing a pre-training model, and adopting a CFD result as a pre-training data source. A plate with randomly distributed weeping holes was simulated under isothermal conditions. The rectangular cross section of the main flow channel is y=20 mm×h=50 mm. The coolant flows into the static pressure tank first and then into the main flow channel through the overflow holes at an angle of 30 degrees. The length L of the hole distribution area was 40mm. The channel extends for a length of 40mm upstream and downstream respectively. The diameter of the hole D is 1mm. The geometric parameters are within the normal range of a typical turbine cooling design. The number of holes varies from 10 to 15. The distribution of holes follows several criteria and constraints (i) holes do not overlap, hole center distance is not less than 2 times the diameter; (ii) The hole configuration has as much diversity as possible, and the generation of the model is ensured; (iii) The holes of the outer surface are in a rectangular area of l=40 mm×y=20 mm. Under the constraints described above, 54 gas film geometries for CFD simulation and machine learning were generated.

The computational domain contains only the fluid domain. The cells are unstructured polyhedra. The boundary layer region creates a prismatic grid. The thickness of the first prismatic layer in the holes was 0.0008mm and at other locations 0.002mm, ensuring that the minimum dimensionless distance y+ of the fluid simulation grid was less than 1 in most areas of the fluid domains. The thickness growth ratio was 1.2. To obtain the proper grid density, grid independence verification was performed on the hot fluid simulation. The housing tested had 10 holes. 156 ten thousand, 219 ten thousand, 292 ten thousand and 450 ten thousand cell number sequences were detected, respectively. The corresponding mesh substrate sizes were 0.84mm, 0.67mm, 0.53mm and 0.42mm, respectively. The numerical dataset samples thus obtained are converted into images for machine learning to obtain a pre-trained model, as shown in fig. 4.

Step 5: assimilate numerical data. Since this model must be compatible with different data sizes (image size and image layer), we use an iterative neural operator based migration learning method driven by the laws of physics to assimilate numerical data for regression of image data obtained from CFD simulations and experiments, as shown in fig. 4. Film cooling is essentially a convective heat transfer problem and can be described in terms of a set of convective diffusion partial differential equations and their boundary conditions. The physical quantities solved for these equations generally follow the law determined by the PDE system:

，

The difference form is expressed as:

where u is the physical quantity of the desired solution (i.e., the cooling effect of the product liquid in this study), F is a nonlinear function of u and its derivatives, Ω is the computational domain, and Γ is the boundary.

The proper super parameters are selected through five-fold cross validation, a pre-training model is trained under the parameters, the model has better generalization precision, and the pre-training model is migrated to an experimental data set through a model-based migration learning method (fine tuning), as shown in fig. 5. The learning rate of the pre-training model and the non-transfer model in the training process is set to 1e-3, and the learning rate of the fine-tuning process is set to 1e-4. During the fine tuning process, all convolution layers in the model are set to be trainable.

Step 6: and respectively carrying out space, structure and working condition deduction. The convolution kernel size used is deduced: 15 x 15, number of hidden layers: 12 hidden state diagram layer number: 7. the iterative operator network model is compiled by the open source deep learning framework TensorFlow 2.0. Training of the neural network model is carried out on a NVIDIA Quadro RTX 4000GPU processor, and model weights are updated by adopting an Adam optimizer in the training process.

In the space deduction part, the iteration operator network is trained by using the collected images with incomplete air film cooling efficiency, the network is supervised by using the incomplete images, and the network learns the physical rule of space commonality from the network, so that the physical image of the incomplete part is deduced. And outputting a complete air film cooling efficiency image by the network, and recovering the image of the whole temperature field.

In the structure deduction part, the principle is that the sample in the training set has the change of the hole number and the hole position, so that the iteration operator network learns potential physical rules, such as low temperature of the downstream of the air film holes and the superposition rule of the air film holes, during training, and the network can rapidly predict the air film cooling efficiency of the pore plate by modifying the hole distribution and the position in the geometric layer input by the network. Even for samples with completely different numbers and arrangement modes of holes in the test set and training sets, the geometric layers are only required to be input into the iterative operator network, and the network can reconstruct the distribution of cooling efficiency well, so that the test set has good generalization capability. When a certain air film hole is continuously moved along a certain direction, the network can effectively predict the air film superposition phenomenon and has good continuous response capability. The step of structure deduction is to modify the geometric layer in the network input after using the training set training model, load the trained weight, run codes and predict the cooling efficiency under different air film hole distribution. Fig. 8 is an effect diagram of performing structure deduction.

In the working condition deduction part, samples in the training set also contain changes of working conditions (such as different blowing ratios, turbulence intensity, density ratios and the like), and the iterative operator network learns the influence of the working conditions on the air film cooling efficiency in the training process, so that the air film cooling efficiency under the working conditions can be rapidly predicted by modifying the values of the blowing ratios, the turbulence intensity and the density ratio layers in the network input layers. For some working conditions which are not contained in the training set in the test set, the network can well predict the cooling efficiency of the air outlet film and has good generalization capability. The working condition deduction is carried out by modifying a boundary condition layer in network input after a training set training model is used, loading trained weights, running codes and predicting the cooling efficiency under different working conditions by the network. Fig. 9 is a schematic diagram of the effect of operating mode deduction.

Example 2

As shown in FIG. 2, the deep learning-based digital twin system for the air film cooling temperature field provided by the preferred embodiment of the invention covers the functions of combination, data assimilation and deduction of a software system and a hardware system, can support the efficient and accurate design of air film cooling, can be widely applied to the related design, manufacture, transportation and detection scenes of air film cooling, and is constructed by adopting a deep learning-based digital twin method for the air film cooling temperature field, and comprises 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 is used for realizing the rapid prediction of physical field information through image deep learning by constructing a simulation body suitable for an air film cooling temperature field, pre-training a model through machine learning to obtain a pre-training model, and fine-tuning the pre-training model through transfer learning to generate a transfer model;

the sensor module comprises various sensors, and is matched with the data acquisition module to complete acquisition of air film cooling test data through a sensor interface;

the data acquisition module acquires CFD simulation data and air film cooling experimental data, the data are processed by the data processing module, compatibility of multi-source heterogeneous data is realized, and the processed data are used as training data and test data of the model training module;

The data processing module is used for processing the data acquired by the data acquisition module, and comprises the steps of converting CFD simulation values into images, assimilating the data, and carrying out regression processing on the CFD simulation data and experimental data;

the physical-virtual synchronization module is used for migrating the pre-training model to experimental data through a model-based migration learning method, fine-tuning the pre-training model to generate a migration model, synchronizing the physical model and the virtual model, and completing software and hardware synchronization;

and the deduction module is used for deducting the space, the structure and the working condition by applying the model trained by the model training module and updating the physical-virtual synchronous model.

The foregoing describes in detail preferred embodiments of the present invention. It should be understood that numerous modifications and variations can be made in accordance with the concepts of the invention without requiring creative effort by one of ordinary skill in the art. Therefore, all technical solutions which can be obtained by logic analysis, reasoning or limited experiments based on the prior art by the person skilled in the art according to the inventive concept shall be within the scope of protection defined by the claims.

Claims (10)

1. The method for digitally twinning the air film cooling temperature field based on deep learning is characterized by comprising the following steps of:

S101: constructing an air film cooling experimental device, wherein the experimental device comprises a sensor, a measuring system and a data acquisition system, and the experimental device is used for completing measurement and acquisition of experimental data;

s103: collecting experimental data, shooting and storing incomplete pore plate temperature images by using an infrared camera, wherein the experimental data are used for fine tuning a pre-training model to generate a migration model;

s105: training the pre-training model, and generating the pre-training model by adopting a CFD result as a pre-training data source;

s107: assimilating numerical data, the numerical data comprising data from different image sizes and image layers, using an iterative neural operator based transfer learning method;

s109: and respectively carrying out space, structure and working condition deduction, and updating the physical-virtual synchronous model.

2. The method of claim 1, wherein in step S101, the film cooling experimental device comprises a wind tunnel, a pressure vessel, a flow meter, a cryostat, a thermal infrared imager, a measurement and data acquisition system; the wind tunnel is rectangular in cross section, and a honeycomb is arranged at the upstream of a measuring area of the wind tunnel so as to provide uniform airflow; the pressure vessel providing a driving force for the primary and secondary streams; the flowmeter is arranged in the middle of the wind tunnel and measures the speed of the main flow; the secondary stream is cooled by the cryostat to-5 ℃ prior to taking pressure and temperature measurements; the infrared thermal imager records the temperature distribution of the outer surface of the jet cooling plate, the measuring and data acquisition system adopts a thermocouple to measure the temperature of the main flow and the secondary flow in real time and acquire data, and the mixture of the main flow and the secondary flow is finally discharged into the environment from the wind tunnel outlet.

3. The method according to claim 1, wherein in said step S103, said experimental data is obtained at predetermined experimental conditions comprising: the experiment is carried out at normal temperature and pressure, the back pressure of the main flow and the secondary flow is 1atm, the temperature of the main flow is 299.60K, and the temperature of the secondary flow is 273.15K; the experimental sample material is an additive manufacturing material, the additive manufacturing material is unsaturated polyester resin, and the thermal conductivity coefficient of the material is 0.1W/m/K.

4. The method according to claim 1, wherein in said step S105, said training pre-training model comprises the steps of:

s1051: using CFD results as a source of pre-training data;

s1052: selecting proper super parameters by utilizing five-fold cross validation;

s1053: the learning rate in the training process is set to be 1e-3;

s1054: updating the weight of the pre-training model by adopting an Adam optimizer;

s1055: and converting the numerical data sample into an image for machine learning to obtain the pre-training model with generalization precision.

5. The method of claim 4, wherein in step S1051, the gas film geometry is obtained by simulating a plate through which randomly distributed weeping holes pass under isothermal conditions, and CFD simulation is performed using the gas film geometry to generate CFD numerical data.

6. The method of claim 5, wherein the computational domain of the CFD simulation comprises only a fluid domain, the computational cells are unstructured polyhedra, and the boundary layer region generates a prismatic grid; the gas film geometric figure is generated under a preset constraint condition, wherein the constraint condition is that exudation holes which are randomly distributed under an isothermal condition pass through a flat plate in a simulation mode, the length of a distribution area of the exudation holes is 40mm, the diameter of the exudation holes is 1mm, the number of the exudation holes is 10-15, and the exudation holes are distributed according to a preset condition; wherein the predetermined condition includes: the holes are not overlapped, and the center distance of the holes is not less than 2 times of the diameter of the holes; the hole configuration of the holes is diversified; the holes of the outer surface are distributed in a rectangular area of 40mm x 20 mm.

7. The method of claim 1, wherein in the step S107, the partial differential equation of the iterative nerve operator is:

，

the difference form is expressed as:

，

where u is the physical quantity of the desired solution, F is a nonlinear function of u and its derivatives, Ω is the computational domain, and Γ is the boundary.

8. The method according to claim 1, wherein in the step S109, the pre-training model is migrated to the experimental data by a model-based migration learning method, the pre-training model is fine-tuned to generate the migration model, the learning rate of the fine-tuning process is set to 1e-4, and all convolution layers in the pre-training model in the fine-tuning process are set to be trainable.

9. The method of claim 8, wherein the deriving trains the iterative neural operator network model using a predetermined convolution kernel size, number of hidden layers, and hidden state diagram layer number parameters, wherein model weights are updated by an Adam optimizer,

the space deduction is carried out, the iterative neural operator network is trained by using the incomplete air film cooling efficiency image, so that the iterative neural operator network learns the physical rule of space commonality, the physical image of the incomplete part is deduced, the complete air film cooling efficiency image is output, and the image of the whole temperature field is restored;

the structure deduction is carried out, the iterative neural operator network is trained by using test samples with completely different hole numbers and hole arrangement modes from a training set, so that the iterative neural operator network can be used for deducing geometrical layers with different hole distribution and positions and predicting the phenomenon of air film superposition, outputting a complete air film cooling efficiency image and recovering the image of the whole temperature field;

the working condition deduction is carried out, a test sample with completely different working condition conditions and a training set is used for training the iterative neural operator network, so that the iterative neural operator network learns the influence of the working condition on the air film cooling efficiency, physical images under different working condition conditions are deduced, a complete air film cooling efficiency image is output, and the image of the whole temperature field is restored; the working conditions comprise: different blowing ratios, different turbulence intensities and different density ratios.

10. A deep learning-based digital twin system for a film cooling temperature field, which is characterized in that the system is constructed by adopting the digital twin method according to any one of claims 1-9, and comprises 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 is used for pre-training the model by utilizing machine learning to obtain a pre-training model, and performing fine adjustment on the pre-training model through transfer learning to generate a transfer model;

the sensor module comprises a sensor interface, and the acquisition of the air film cooling test data is completed by matching with the data acquisition module through the sensor interface;

the data acquisition module acquires CFD simulation data and the air film cooling experimental data, and the data are processed by the data processing module and then serve as training data and test data of the model training module;

the data processing module is used for processing the data acquired by the data acquisition module, wherein the processing comprises the steps of converting the CFD simulation numerical value into an image, assimilating the numerical value data and carrying out regression processing on the CFD simulation data and the experimental data;

The physical-virtual synchronization module migrates the pre-training model to the experimental data through a model-based migration learning method, fine-tunes the pre-training model, generates the migration model, and realizes synchronization of a physical model and a virtual model;

and the deduction module is used for deducting space, structure and working condition by applying the model trained by the model training module and updating the physical-virtual synchronous model.