Digital twinborn experiment-based high-fidelity online test method for complex turbulence kinetic parameters of numerical pool

Abstract

The invention establishes a digital twinborn experiment-based numerical pool complex turbulence dynamics parameter high-fidelity online test method, and solves the problems of low fidelity and poor timeliness of a traditional numerical pool experiment complex turbulence field online test. The method comprises the following steps: firstly, establishing a multi-source data acquisition system to perform real-time acquisition and transmission on monitoring data of various types of sensors in a physical pool; secondly, constructing a tensor-based neural network to establish a universal and robust turbulence model; thirdly, constructing a multi-scale residual convolutional neural network to extract anisotropic components of the Reynolds stress tensor and nonlinear evolution characteristics of the Reynolds stress tensor in space and time dimensions, and further realizing stable coupling iteration of a turbulence model and a Navier-Stokes equation; and finally, constructing a physical pool digital twinborn body by organically fusing monitoring data, a mechanism model and a machine learning method. According to the method, the problems of low fidelity and poor timeliness of the digital twinborn experiment caused by coupling iteration instability of the turbulence model and the Navier-Stokes equation can be solved.

Description

Technical Field

[0001] The present invention relates to numerical water tank online testing, machine learning and virtual-reality interactive digital twins in the field of experimental fluid mechanics, and in particular to a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water tank based on digital twin experiments. Background Art

[0002] With the development of digitalization and intelligence, numerical water pools will face a variety of virtual experimental scenarios. How to accurately simulate the turbulent fields of all possible operating scenarios of thousands or even millions in a short period of time is the key to avoiding local unreasonable designs. In addition, how to deeply integrate artificial intelligence algorithms and numerical simulations to find the optimal design scheme and even the automatic control scheme for numerical water pools is also a difficult problem faced by the field of experimental fluid mechanics. This type of digital technology has important practical significance for the refinement and low-carbonization of the design, construction, operation and maintenance of numerical water pool systems. At the same time, it also puts forward higher requirements on the ease of use of numerical experimental methods, as well as simulation accuracy and calculation speed.

[0003] Numerical water pools often have high Reynolds number turbulence characteristics, showing strong nonlinear and multi-scale dynamic characteristics. At present, the understanding of the evolution mechanism of turbulence under high Reynolds numbers is still unclear. Researchers mainly simulate turbulent fields by coupling various classic turbulence models with the Navier-Stokes equations and using fine grid division. However, complex turbulent dynamics experiments in numerical water pools under high Reynolds numbers are sensitive to factors such as grid division and the selection of turbulence models, requiring users to have high theoretical knowledge, poor usability in practical applications, and low reliability of flow field simulation. This type of traditional numerical water pool experimental method has entered a relatively bottleneck development stage. After decades of optimization, there is little room for step-by-step improvement. With the development of informatization, more and more high-precision, high-temporal and high-space resolution flow field experimental data are easily available. How to effectively use these big data, extract key flow information from them, and further promote the development of experimental fluid mechanics technology has been a research hotspot in the field of numerical water pools in recent years. Thanks to the rapid development of high-performance scientific computing, data mining and machine learning, digital twin experimental methods are gradually becoming a new research paradigm in experimental fluid mechanics. Through online simulation, identification inversion and diagnostic analysis at all time and space scales, digital twin experimental technology provides basic means and tools for online testing of complex turbulent dynamic parameters of numerical water tanks, which has great research significance for accelerating the digital transformation of numerical water tanks.

[0004] The digital twin experimental method can make full use of advanced fluid mechanics experimental testing technology, high-performance scientific computing, big data analysis and a new generation of artificial intelligence methods to perform high-fidelity virtual / real interaction on complex turbulent fields with high Reynolds numbers in numerical water pools. Compared with traditional turbulence modeling methods, this method greatly enhances the versatility and robustness of turbulence models through multi-source data fusion, and has achieved great success in a series of scenarios such as the exploration of the mechanism of turbulence evolution and the direct and inverse problems of complex flow simulation. Although the digital twin experimental method has great application potential, recent studies have found that even if the high-precision turbulence model constructed by this method controls the Reynolds stress tensor modeling error below 0.5%, the final simulated velocity field error can be as high as 35%. Due to the strong nonlinearity of the high Reynolds number turbulence dynamics process, when the turbulence model is coupled with the Navier-Stokes equations for iterative calculation, the small disturbances of the turbulence model are easily accumulated and amplified in the velocity field, causing large deviations in the velocity field. The deviation of the velocity field will further feed back as the input deviation of the turbulence model. This feedback makes it difficult for the solution process to converge, and ultimately leads to poor or even divergent simulation accuracy of the turbulence field. Up to now, the instability problem of the turbulence model constructed by the digital twin experimental method and the coupling calculation of the Navier-Stokes equation has not been reasonably solved, which to a certain extent restricts the improvement of the simulation fidelity of the complex turbulence field of the numerical water tank. In addition, unlike the traditional offline numerical simulation method, a key link of the digital twin experimental technology is to realize the online dynamic interaction between the physical entity and the digital twin. The cumbersome iterative process when coupling the turbulence model with the Navier-Stokes equation will also lead to a long calculation time, which is difficult to meet the timeliness requirements of the digital twin experiment. Summary of the invention

[0005] The present invention aims to overcome the problems of low fidelity and poor timeliness of digital twin experiments caused by iterative instability of coupling turbulence models with Navier-Stokes equations, and to provide a high-fidelity digital twin experimental method for online testing of complex turbulent fields in numerical water tanks. This method constructs a multi-source fused digital twin by organically integrating monitoring data, mechanism models, and machine learning, thereby achieving high-fidelity mirroring and online testing of complex turbulent processes in physical water tanks in digital space.

[0006] In view of this, the present invention provides a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment, which is characterized by comprising the following steps:

[0007] S1: Establish a multi-source data acquisition system to collect and transmit a wide variety of sensor monitoring data of physical pools in real time;

[0008] S2: construct a tensor-based neural network to establish an eddy viscosity model that describes the nonlinear relationship between the Reynolds stress tensor and the strain rate tensor;

[0009] S3: Generate a high-precision training sample set of a typical numerical pool turbulence field and train a tensor-based neural network;

[0010] S4: Modify the classical linear gradient diffusion hypothesis and preprocess the Reynolds stress tensor; decompose it into linear correlation terms and nonlinear terms by projecting it on the strain rate tensor, and then obtain the spatiotemporal evolution characteristics of the Reynolds stress tensor in different flow field characteristic regions;

[0011] S5: Summarize the non-uniform and anisotropic characteristics of the spatial distribution of the pre-processed Reynolds stress tensor, and construct a multi-scale residual convolutional neural network in the spatial dimension to extract the equivalent turbulent viscosity coefficient according to the local spatiotemporal discrete step length of the numerical pool flow field;

[0012] S6: Construct a Reynolds stress tensor form that can converge stably during the iteration process, and then establish a fast iterative algorithm for coupling the turbulence model with the Navier-Stokes equations;

[0013] S7: A digital twin of the physical water pool is constructed by organically integrating monitoring data, mechanism models, and machine learning methods, and complex turbulent processes in the physical water pool are mirrored and tested online with high fidelity in the digital space.

[0014] Furthermore, in step S1, a multi-source data acquisition system is established to collect and transmit various sensor monitoring data of the physical pool in real time, which specifically includes the following steps:

[0015] Step 1-1: Select a typical physical water pool with rich vortex structures to design and layout the measurement points;

[0016] In step 1-2, a multi-source sensor data acquisition system is developed based on the Labview software platform using signal conversion technology, real-time communication technology, multi-thread management technology and C / S programming technology to collect, analyze, display, store, convert and send a wide variety of sensor monitoring data from physical water pools in real time.

[0017] Furthermore, in step S2, the nonlinear eddy viscosity model is constructed based on the tensor-based neural network, which specifically includes the following steps:

[0018] Step 2-1: Based on the classical turbulence modeling theory, we deeply integrate the analysis tools such as characteristic invariant analysis, scaling rate, and dimensional analysis to select effective turbulence modeling input feature quantities; by streamlining the input feature variables and the dimension and complexity of the neural network, we can improve the compactness and generalization ability of the turbulence model;

[0019] Step 2-2: Construct the deviatoric stress tensor based on Reynolds stress The deviatoric stress tensor is dimensionless as bij =a ij / k. As in formula (1), the constitutive equation of the dimensionless deviatoric stress tensor is constructed:

[0020]

[0021] Among them, u′ is the velocity pulsation, k is the turbulent kinetic energy, and are the normalized mean strain rate tensor and rotation tensor, respectively.

[0022] Step 2-3, through the symmetric tensor and the antisymmetric tensor Constructing a complete tensor function basis for the deviatoric stress tensor as follows:

[0023]

[0024] Among them, I is the identity matrix, and Tr represents the trace of the matrix.

[0025] In step 2-4, the constitutive equation of the dimensionless deviatoric stress tensor is fitted as a linear combination of the above ten tensor bases, and the specific form is as follows:

[0026]

[0027] Among them, the fitting coefficient expression g(I1, I2, I3, I4, I5) depends on the following five invariants:

[0028]

[0029] In step 2-5, for the nonlinear dynamics of passive scalar transport in the turbulent field, a general turbulent diffusion tensor is constructed based on the anisotropic turbulent diffusion coefficient tensor for modeling, as shown in formula (3):

[0030]

[0031] Where θ represents the passive scalar field, and the dimensionless form of the passive scalar gradient is Here ε is the turbulent energy dissipation rate; g(I1, I2, I3, I4, I5; J0, J1, ..., J 10 ) is the fitting coefficient expression, J0~J 10 represents the invariant associated with the passive scalar; D ij is the anisotropic turbulent diffusion coefficient tensor.

[0032] Step 2-6, determine the invariant form of the passive scalar: in, Represents a set of tensor bases, which are in the following form:

[0033]

[0034] Steps 2-7: construct the anisotropic turbulent diffusion coefficient tensor D based on the following ten tensor function bases ij :

[0035]

[0036] in, represents the normalized Reynolds stress tensor.

[0037] Furthermore, in step S3, generating a high-precision training sample set of a typical numerical water pool turbulence field and training a tensor-based neural network specifically includes the following steps:

[0038] In step 3-1, a typical numerical pool turbulence field with rich vortex structures is selected as a training sample set (including flow around a cylinder, square cavity flow, and back-step flow, etc.), and a high-precision turbulence field database is established through direct numerical simulation methods;

[0039] In step 3-2, a tensor-based neural network is trained to obtain an interpretable and universal turbulence model, and the reliability and feasibility of the turbulence model are corrected through a posteriori analysis.

[0040] Furthermore, in step S4, the projection of the Reynolds stress tensor on the strain rate tensor is preprocessed, which specifically includes the following steps:

[0041] Step 4-1, correct the classical linear gradient diffusion hypothesis, establish a nonlinear eddy viscosity model that describes the relationship between the Reynolds stress tensor and the strain rate tensor, project the Reynolds stress tensor on the strain rate tensor, and obtain the effective turbulent viscosity coefficient of each component of the Reynolds stress tensor, as shown in formula (4):

[0042]

[0043] in, represents the effective turbulent viscosity coefficient; is the average strain rate tensor.

[0044] In step 4-2, the Reynolds stress tensor is preprocessed and decomposed into a quasi-linear part and a nonlinear part related to the strain rate tensor, as shown in formula (5):

[0045]

[0046] in, represents the nonlinear part of the Reynolds stress tensor.

[0047] Furthermore, in step S5, according to the local spatiotemporal discrete steps of the numerical pool flow field, a multi-scale residual convolutional neural network is constructed in the spatial dimension to extract an effective turbulent viscosity coefficient, which specifically includes the following steps:

[0048] Step 5-1: At any discrete node position in the numerical pool, determine the size of the turbulence field dependency domain based on the local disturbance propagation speed and the evolution time step;

[0049] Step 5-2, search for neighbor points in the turbulence field dependency domains of different sizes to determine the position of the spatial template point, and then reconstruct the turbulence field quantity of the spatial template point according to the neighbor point information in the influence domains of different sizes;

[0050] In step 5-3, a multi-scale residual convolutional neural network is constructed using convolutional layers, deconvolutional layers and residual connections to extract the equivalent turbulent viscosity coefficient at the above-mentioned spatial template points.

[0051] Furthermore, in step S6, a Reynolds stress tensor form that can converge stably is constructed to establish a fast iterative algorithm for coupling the turbulence model with the Navier-Stokes equations, which specifically includes the following steps:

[0052] In step 6-1, the linear part of the Reynolds stress tensor obtained in step 4-2 is directly implanted into the viscous diffusion term of the Navier-Stokes equation in the form of an equivalent shear stress tensor. The nonlinear part that does not satisfy the gradient diffusion assumption is implanted into the Navier-Stokes equation in the form of a force source term, as shown in equation (6):

[0053]

[0054] Where ρ represents the fluid density, is the average pressure, ν is the laminar viscosity coefficient, F i is the physical force source term.

[0055] In step 6-2, the convolution kernel evolution operator is used to calculate the shear stress diffusion term and the force source term in equation (6) respectively. The effective viscosity diffusion coefficient obtained in S5 is embedded in the convolution kernel evolution operator and coupled with the Navier-Stokes equations for iteration until convergence, and finally the Reynolds stress tensor form that can be stably calculated with the Navier-Stokes equations is obtained.

[0056] In step 6-3, for the nonlinear transport dynamics of passive scalar in the turbulent field, the turbulent diffusion tensor is projected in the corresponding velocity direction based on the field synergy enhancement theory as shown in equation (7), and then the dynamic evolution of the passive scalar transport process is realized based on the unified convolution kernel evolution operator:

[0057]

[0058] in, represents the equivalent turbulent diffusion coefficient.

[0059] In step 6-4, the robustness and reliability of the fast iteration algorithm of the turbulence model coupled Navier-Stokes equations are corrected through a posteriori analysis.

[0060] Furthermore, in step S7, a digital twin of the physical water pool is constructed based on multi-source sensor data fusion technology, and the complex turbulence process in the physical water pool is mirrored with high fidelity in the digital space. The specific steps are as follows:

[0061] Based on the above steps S1 to S6, a digital twin that organically integrates monitoring data, mechanism models and machine learning methods is constructed to mirror the complex turbulence process in the physical water pool with high fidelity in the digital space, and the statistics of turbulent dynamics of each order in the numerical water pool are tested online.

[0062] Compared with the prior art, the high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment described in the present invention has the following advantages:

[0063] (1) Make full use of advanced fluid mechanics experimental testing technology, high-performance scientific computing, big data analysis and the new generation of artificial intelligence methods to build digital twin experimental technology to change the existing experimental fluid mechanics research paradigm and accelerate the digital transformation of numerical water tanks; through real-time simulation, identification inversion and diagnostic analysis at all time and space scales, it provides basic means and tools for online testing of complex turbulent dynamic parameters of numerical water tanks.

[0064] (2) We fully exploit the potential of machine learning methods to extract strong nonlinear pattern characteristics of complex turbulent dynamics and construct a tensor-based neural network to establish a universal and general turbulence model, which greatly improves its transferability and scalability to different numerical water tank cases.

[0065] (3) The preprocessing and nonlinear feature extraction methods of the Reynolds stress tensor can greatly alleviate the problem of iterative instability when the turbulence model is coupled with the Navier-Stokes equations, improve the fidelity and timeliness of the numerical water tank digital twin experiment to a certain extent, and realize high-fidelity virtual / real interaction and online dynamic interaction between the physical entity and the digital twin. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is the digital twin experimental process of the marine environment numerical water tank in the embodiment of the method of the present invention.

[0067] Figure 2 The present invention provides a high-fidelity online testing method for complex turbulence dynamics parameters in a numerical water pool based on a digital twin experiment.

[0068] Figure 3 This is the data collection, transmission and interaction process of the numerical water tank digital twin experiment process in the embodiment of the method of the present invention.

[0069] Figure 4 The method of the present invention is a stable simulation process of a turbulence model coupled with the Navier-Stokes equations assisted by machine learning.

[0070] Figure 5 This is the online machine learning and posterior analysis process of the complex turbulence process in the numerical water pool of the method of the present invention. DETAILED DESCRIPTION

[0071] The following is a detailed description of an embodiment of the present invention in conjunction with the accompanying drawings. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0072] This embodiment is described based on a scenario of a ship sailing in a marine environment. Figure 1 The present invention provides a high-fidelity online testing method for complex turbulence dynamics parameters in a numerical water tank based on a digital twin experiment. Figure 2 As shown, the following steps are included;

[0073] S1: Establish a multi-source data acquisition system to collect and transmit a wide variety of sensor monitoring data in the marine environment numerical pool in real time. The brief steps are as follows:

[0074] Step 1-1: Determine the key measured physical quantities and optimize the design and layout of sensor measurement points;

[0075] Steps 1-2: Based on the Labview software platform, a multi-source sensor data acquisition system is developed using signal conversion technology, real-time communication technology, multi-threading management technology and C / S programming technology to collect, analyze, display, store, convert and send a wide variety of sensor monitoring data (including wind speed, water level, load stress, and ship speed and acceleration) of the marine environment numerical water tank in real time.

[0076] S2: Construct a tensor-based neural network to establish an eddy viscosity model that describes the nonlinear relationship between the Reynolds stress tensor and the strain rate tensor, including the following steps:

[0077] Step 2-1: Based on the classical turbulence modeling theory, we deeply integrate the analysis tools such as characteristic invariant analysis, scaling rate, and dimensional analysis to select effective turbulence modeling input feature quantities; by streamlining the input feature variables and the dimension and complexity of the neural network, we can improve the compactness and generalization ability of the turbulence model;

[0078] Step 2-2: Construct the deviatoric stress tensor based on Reynolds stress The deviatoric stress tensor is dimensionless as b ij =a ij / k. As in formula (1), the constitutive equation of the dimensionless deviatoric stress tensor is constructed:

[0079]

[0080] Among them, u′ is the velocity pulsation, k is the turbulent kinetic energy, and are the normalized mean strain rate tensor and rotation tensor, respectively.

[0081] Step 2-3, through the symmetric tensor and the antisymmetric tensor Constructing a complete tensor function basis for the deviatoric stress tensor as follows:

[0082]

[0083] Among them, I is the identity matrix, and Tr represents the trace of the matrix.

[0084] In step 2-4, the constitutive equation of the dimensionless deviatoric stress tensor is fitted as a linear combination of the above ten tensor bases, and the specific form is as follows:

[0085]

[0086] Among them, the fitting coefficient expression g(I1, I2, I3, I4, I5) depends on the following five invariants:

[0087]

[0088] In step 2-5, for the nonlinear dynamics of passive scalar transport in the turbulent field, a general turbulent diffusion tensor is constructed based on the anisotropic turbulent diffusion coefficient tensor for modeling, as shown in formula (3):

[0089]

[0090] Where θ represents the passive scalar field, and the dimensionless form of the passive scalar gradient is Here ε is the turbulent kinetic energy dissipation rate; g(I1, I2, I3, I4, I5; J0, J1, ..., J 10 ) is the fitting coefficient expression, J0~J 10 represents the invariant associated with the passive scalar; D ij is the anisotropic turbulent diffusion coefficient tensor.

[0091] Step 2-6, determine the invariant form of the passive scalar: in, Represents a set of tensor bases, which are in the following form:

[0092]

[0093]

[0094] Steps 2-7: Construct the anisotropic turbulent diffusion coefficient tensor D based on the following ten tensor function bases: ij :

[0095]

[0096] in, represents the normalized Reynolds stress tensor.

[0097] S3: Combine different marine environment parameters to generate a high-precision numerical pool turbulence field sample data set and train a tensor-based neural network. The main steps are as follows:

[0098] Step 3-1: Select classic flow scenarios with rich vortex structures (including flow around a cylinder, square cavity flow, and back-step flow), and establish a high-precision turbulence field database through direct numerical simulation methods;

[0099] In step 3-2, a tensor-based neural network is trained to obtain an interpretable and universal turbulence model, and the reliability and feasibility of the turbulence model are corrected through a posteriori analysis.

[0100] S4: Modify the classical linear gradient diffusion hypothesis and preprocess the Reynolds stress tensor; decompose it into linear correlation terms and nonlinear terms by projecting it on the strain rate tensor, and then obtain the spatiotemporal evolution characteristics of the Reynolds stress tensor in different flow field characteristic regions. The specific steps are as follows:

[0101] Step 4-1, correct the classical linear gradient diffusion hypothesis, establish a nonlinear eddy viscosity model that describes the relationship between the Reynolds stress tensor and the strain rate tensor, project the Reynolds stress tensor on the strain rate tensor, and obtain the effective turbulent viscosity coefficient of each component of the Reynolds stress tensor, as shown in formula (4):

[0102]

[0103] in, represents the effective turbulent viscosity coefficient; is the average strain rate tensor.

[0104] In step 4-2, the Reynolds stress tensor is preprocessed and decomposed into a quasi-linear part and a nonlinear part related to the strain rate tensor, as shown in equation (5):

[0105]

[0106] in, represents the nonlinear part of the Reynolds stress tensor.

[0107] S5: Summarize the non-uniformity and anisotropy of the spatial distribution of the pre-processed Reynolds stress tensor. According to the local space-time discrete step length of the numerical pool flow field in the offshore environment, a multi-scale residual convolutional neural network is constructed in the spatial dimension to extract the equivalent turbulent viscosity coefficient. The steps are as follows:

[0108] Step 5-1: At any discrete node position in the marine environment numerical water pool, determine the size of the turbulence field dependency domain according to the local disturbance propagation speed and evolution time step;

[0109] Step 5-2, search for neighbor points in the turbulence field dependency domains of different sizes to determine the position of the spatial template point, and then reconstruct the turbulence field quantity of the spatial template point according to the neighbor point information in the influence domains of different sizes;

[0110] In step 5-3, a multi-scale residual convolutional neural network is constructed using convolutional layers, deconvolutional layers and residual connections to extract the equivalent turbulent viscosity coefficient at the above-mentioned spatial template points.

[0111] S6: Construct a Reynolds stress tensor form that can converge stably during the iteration process, and then establish a fast iterative algorithm for coupling the turbulence model with the Navier-Stokes equations. The detailed steps are as follows:

[0112] In step 6-1, the linear part of the Reynolds stress tensor obtained in step 4-2 is directly implanted into the viscous diffusion term of the Navier-Stokes equation in the form of an equivalent shear stress tensor. The nonlinear part that does not satisfy the gradient diffusion assumption is implanted into the Navier-Stokes equation in the form of a force source term, as shown in equation (6):

[0113]

[0114] Where ρ represents the fluid density, is the average pressure, v is the laminar viscosity coefficient, F i is the physical force source term.

[0115] In step 6-2, the convolution kernel evolution operator is used to calculate the shear stress diffusion term and the force source term in equation (6) respectively. The effective viscosity diffusion coefficient obtained in S5 is embedded in the convolution kernel evolution operator and coupled with the Navier-Stokes equations for iteration until convergence, and finally the Reynolds stress tensor form that can be stably calculated with the Navier-Stokes equations is obtained.

[0116] In step 6-3, for the nonlinear transport dynamics of passive scalar in the turbulent field, the turbulent diffusion tensor is projected in the corresponding velocity direction based on the field synergy enhancement theory as shown in equation (7), and then the dynamic evolution of the passive scalar transport process is realized based on the unified convolution kernel evolution operator:

[0117]

[0118] in, represents the equivalent turbulent diffusion coefficient.

[0119] In step 6-4, the robustness and reliability of the fast iteration algorithm of the turbulence model coupled Navier-Stokes equations are corrected through a posteriori analysis.

[0120] S7: Organically integrate monitoring data, mechanism models and machine learning methods to construct a digital twin of the offshore water tank, and perform high-fidelity mirroring and online testing of the complex turbulence process in the offshore water tank in digital space. The specific steps are as follows:

[0121] Based on the above steps S1 to S6, a digital twin of the offshore water tank is constructed that organically integrates monitoring data, mechanism models and machine learning methods. The complex turbulence process in the offshore water tank is mirrored with high fidelity in the digital space, and the statistics of turbulence dynamics of various orders in the numerical water tank are tested online.

[0122] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A high-fidelity online testing method for complex turbulence dynamics parameters in a numerical water pool based on digital twin experiments, characterized in that: The following steps are involved: S1: Establish a multi-source data acquisition system to collect and transmit a wide variety of sensor monitoring data of physical pools in real time; S2: construct a tensor-based neural network to establish an eddy viscosity model that describes the nonlinear relationship between the Reynolds stress tensor and the strain rate tensor; S3: Generate a high-precision training sample set of a typical numerical pool turbulence field and train a tensor-based neural network; S4: Modify the classical linear gradient diffusion hypothesis and preprocess the Reynolds stress tensor; decompose it into linear correlation terms and nonlinear terms by projecting it on the strain rate tensor, and then obtain the spatiotemporal evolution characteristics of the Reynolds stress tensor in different flow field characteristic regions; S5: Summarize the non-uniform and anisotropic characteristics of the spatial distribution of the pre-processed Reynolds stress tensor, and construct a multi-scale residual convolutional neural network in the spatial dimension to extract the equivalent turbulent viscosity coefficient according to the local spatiotemporal discrete step length of the numerical pool flow field; S6: Construct a Reynolds stress tensor form that can converge stably during the iteration process, and then establish a fast iterative algorithm for coupling the turbulence model with the Navier-Stokes equations; S7: A digital twin of the physical water pool is constructed by organically integrating monitoring data, mechanism models, and machine learning methods, and complex turbulent processes in the physical water pool are mirrored and tested online with high fidelity in the digital space.

2. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment is characterized in that: In step S1, a multi-source data acquisition system is established to collect and transmit various sensor monitoring data of the physical pool in real time, which specifically includes the following steps: Step 1-1: Select a typical physical water pool with rich vortex structures to design and layout the measurement points; In step 1-2, a multi-source sensor data acquisition system is developed based on the Labview software platform using signal conversion technology, real-time communication technology, multi-thread management technology and C / S programming technology to collect, analyze, display, store, convert and send a wide variety of sensor monitoring data from physical water pools in real time.

3. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment is characterized in that: In the step S2, a nonlinear eddy viscosity model is constructed based on a tensor-based neural network, which specifically includes the following steps: Step 2-1: Based on the classical turbulence modeling theory, we deeply integrate the analysis tools such as characteristic invariant analysis, scaling rate, and dimensional analysis to select effective turbulence modeling input feature quantities; by streamlining the input feature variables and the dimension and complexity of the neural network, we can improve the compactness and generalization ability of the turbulence model; Step 2-2: Construct the deviatoric stress tensor based on Reynolds stress The deviatoric stress tensor is dimensionless as b ij =a ij / k. As in formula (1), the constitutive equation of the dimensionless deviatoric stress tensor is constructed: Among them, u′ is the velocity pulsation, k is the turbulent kinetic energy, and are the normalized mean strain rate tensor and rotation tensor, respectively. Step 2-3, through the symmetric tensor and the antisymmetric tensor Constructing a complete tensor function basis τ for the deviatoric stress tensor (n) ,as follows: Among them, I is the identity matrix, and Tr represents the trace of the matrix. In step 2-4, the constitutive equation of the dimensionless deviatoric stress tensor is fitted as a linear combination of the above ten tensor bases, and the specific form is as follows: Among them, the fitting coefficient expression g(I1, I2, I3, I4, I5) depends on the following five invariants: In step 2-5, for the nonlinear dynamics of passive scalar transport in the turbulent field, a general turbulent diffusion tensor is constructed based on the anisotropic turbulent diffusion coefficient tensor for modeling, as shown in formula (3): Where θ represents the passive scalar field, and the dimensionless form of the passive scalar gradient is Here ε is the turbulent energy dissipation rate; g(I1, I2, I3, I4, I5; J0, J1, ..., J 10 ) is the fitting coefficient expression, J0~J 10 represents the invariant associated with the passive scalar; D ij is the anisotropic turbulent diffusion coefficient tensor. Step 2-6, determine the invariant form of the passive scalar: (1≤n≤10); where, Represents a set of tensor bases, which are in the following form: Steps 2-7: construct the anisotropic turbulent diffusion coefficient tensor D based on the following ten tensor function bases ij : in, represents the normalized Reynolds stress tensor.

4. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment is characterized in that: The step S3 generates a high-precision training sample set of a typical numerical water pool turbulence field and trains a tensor-based neural network, which specifically includes the following steps: In step 3-1, a typical numerical pool turbulence field with rich vortex structures is selected as a training sample set (including flow around a cylinder, square cavity flow, and back-step flow, etc.), and a high-precision turbulence field database is established through direct numerical simulation methods; In step 3-2, a tensor-based neural network is trained to obtain an interpretable and universal turbulence model, and the reliability and feasibility of the turbulence model are corrected through a posteriori analysis.

5. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment is characterized in that: In step S4, the projection of the Reynolds stress tensor on the strain rate tensor is preprocessed, which specifically includes the following steps: Step 4-1, correct the classical linear gradient diffusion hypothesis, establish a nonlinear eddy viscosity model that describes the relationship between the Reynolds stress tensor and the strain rate tensor, project the Reynolds stress tensor on the strain rate tensor, and obtain the effective turbulent viscosity coefficient of each component of the Reynolds stress tensor, as shown in formula (4): in, represents the effective turbulent viscosity coefficient; is the average strain rate tensor. In step 4-2, the Reynolds stress tensor is preprocessed and decomposed into a quasi-linear part and a nonlinear part related to the strain rate tensor, as shown in formula (5): in, represents the nonlinear part of the Reynolds stress tensor.

6. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water tank based on a digital twin experiment is characterized in that: In step S5, according to the local space-time discrete step length of the numerical pool flow field, a multi-scale residual convolutional neural network is constructed in the spatial dimension to extract the effective turbulent viscosity coefficient, which specifically includes the following steps: Step 5-1: At any discrete node position in the numerical pool, determine the size of the turbulence field dependency domain based on the local disturbance propagation speed and the evolution time step; Step 5-2, search for neighbor points in the turbulence field dependency domains of different sizes to determine the position of the spatial template point, and then reconstruct the turbulence field quantity of the spatial template point according to the neighbor point information in the influence domains of different sizes; In step 5-3, a multi-scale residual convolutional neural network is constructed using convolutional layers, deconvolutional layers and residual connections to extract the equivalent turbulent viscosity coefficient at the above-mentioned spatial template points.

7. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water pool based on a digital twin experiment is characterized in that: In step S6, a Reynolds stress tensor form that can converge stably is constructed to establish a fast iterative algorithm for coupling the turbulence model with the Navier-Stokes equations, which specifically includes the following steps: In step 6-1, the linear part of the Reynolds stress tensor obtained in step 4-2 is directly implanted into the viscous diffusion term of the Navier-Stokes equation in the form of an equivalent shear stress tensor. The nonlinear part that does not satisfy the gradient diffusion assumption is implanted into the Navier-Stokes equation in the form of a force source term, as shown in equation (6): Where ρ represents the fluid density, is the average pressure, v is the laminar viscosity coefficient, F i is the physical force source term. In step 6-2, the convolution kernel evolution operator is used to calculate the shear stress diffusion term and the force source term in equation (6) respectively. The effective viscosity diffusion coefficient obtained in S5 is embedded in the convolution kernel evolution operator and coupled with the Navier-Stokes equations for iteration until convergence, and finally the Reynolds stress tensor form that can be stably calculated with the Navier-Stokes equations is obtained. In step 6-3, for the nonlinear transport dynamics of passive scalar in the turbulent field, the turbulent diffusion tensor is projected in the corresponding velocity direction based on the field synergy enhancement theory as shown in equation (7), and then the dynamic evolution of the passive scalar transport process is realized based on the unified convolution kernel evolution operator: in, represents the equivalent turbulent diffusion coefficient. In step 6-4, the robustness and reliability of the fast iteration algorithm of the turbulence model coupled Navier-Stokes equations are corrected through a posteriori analysis.

8. According to claim 1, a high-fidelity online testing method for complex turbulence dynamics parameters of a numerical water tank based on a digital twin experiment is characterized in that: In step S7, a digital twin of the physical water pool is constructed based on multi-source sensor data fusion technology, and the complex turbulence process in the physical water pool is mirrored with high fidelity in the digital space. The specific steps are as follows: Based on the above steps S1 to S6, a digital twin that organically integrates monitoring data, mechanism models and machine learning methods is constructed to mirror the complex turbulence process in the physical water pool with high fidelity in the digital space, and the statistics of turbulent dynamics of each order in the numerical water pool are tested online.

Images