Flow field reconstruction method and system based on fusion of sparse sensor data and physical knowledge

By using the encoder and decoder decoupling architecture and fine-tuning of the physical loss function of the Navi-Stokes equation in the flow field reconstruction, the problem of flow field reconstruction under sparse sensing data is solved, and the flow field is realized High precision and flexibility refactoring reduces labor costs and complexity.

CN118657087BActive Publication Date: 2024-11-08HUNAN UNIV
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202411117276.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-08
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The prior art is difficult to realize the full-domain refined reconstruction of flow fields under sparse sensing data conditions, especially in the case of sensor position movement and complex flow fields, and the lack of physical prior knowledge leads to insufficient overfitting and generalization capabilities of the model. .

Method used

Pre-training is performed using an encoder and decoder decoder architecture, combining the Navi-Stokes equation and the Reynolds number embedded physical loss function for fine-tuning, extracting flow field features through mask autoencoder and position encoding, and using depth The neural network performs transfer learning to ensure that the model complies with physical laws.

Benefits of technology

High-precision reconstruction of flow field under extremely sparse sensor conditions is realized, which reduces the complexity and labor cost of traditional methods, improves the generalization ability and flexibility of the model, and adapts to flow field reconstruction tasks in different scenarios and requirements. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118657087B_ABST
    Figure CN118657087B_ABST
Patent Text Reader

Abstract

The present invention applies deep learning algorithms to the field of fluid mechanics, and discloses a flow field reconstruction method and system based on the fusion of sparse sensor data and physical knowledge, so as to realize global refined flow field reconstruction under the condition of a small number of sensor data points. The method includes: obtaining a pre-trained model for the flow field under different scenarios and different Reynolds numbers; fine-tuning the pre-trained model for downstream tasks based on the target application scenario. During the fine-tuning process, the Navier-Stokes equations and the Reynolds number corresponding to the flow field are embedded in the physical loss function by numerical differentiation in the form of approximate corresponding partial derivatives, and then the physical loss function and the pre-trained loss function are multiplied by the corresponding weights to obtain the total loss function of the flow field reconstruction model; finally, the real-time data collected by each sparsely deployed sensor in the target application scenario is combined with the position of the sensor and input into the flow field reconstruction model to obtain the reconstructed global refined flow field.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the application of deep learning in the field of fluid mechanics, and in particular to a flow field reconstruction method and system based on the fusion of sparse sensor data and physical knowledge. Background Art

[0002] Flow field refers to the area in fluid mechanics that describes the distribution of physical quantities such as velocity, pressure, and temperature of a fluid in space. The flow field can be the motion state of gas, liquid, or other fluids. Flow field reconstruction refers to the process of estimating and reconstructing the unknown part of the flow field through an algorithm or method using experimental data or numerical simulation results. This process is of great significance for scientific research and engineering applications.

[0003] The importance of flow field reconstruction is reflected in many aspects. First, understanding the global distribution of the flow field is crucial to understanding the fluid motion behavior. By reconstructing the flow field, we can obtain key and more detailed flow information, which helps to deeply understand the physical mechanism of the flow. Secondly, flow field reconstruction has a wide range of applications in the engineering field. For example, in the fields of aerospace, automotive engineering, energy systems, etc., accurate reconstruction of the flow field can help optimize the design and improve the performance and efficiency of the system.

[0004] Reconstructing the flow field through sensors has many benefits. Sensors can capture local information in the flow field, such as velocity, pressure, etc. By integrating and processing these local data, we can reconstruct the global characteristics of the flow field. This method has high temporal and spatial resolution and real-time performance, and can provide effective support for real-time monitoring and control of fluid systems. In addition, sensor reconstruction of the flow field can also improve the efficiency and accuracy of data acquisition, reduce experimental costs, and thus promote the development of fluid mechanics research and engineering applications.

[0005] However, sparse data reconstruction aims to restore or reconstruct the global flow field based on limited sensor data from a space that we cannot fully observe. Sensors usually have extremely sparse coverage, and reconstructing complex flow field evolution using sensor data is a huge challenge.

[0006] CN116468602A discloses a turbulence super-resolution method based on the Transformer (a sequence model based on an attention mechanism) deep learning architecture. In the process of reconstructing a low-resolution flow field into a super-resolution flow field, it only performs training once for one Reynolds number, and the monitoring data required for the low-resolution flow field it is based on is still relatively large, and the applicable scenarios are limited.

[0007] Based on linear theoretical methods, such as Galerkin transform and linear stochastic estimation, it is often difficult to reconstruct the global field with a limited number of sensor data. However, neural networks have become an effective alternative due to their powerful nonlinear modeling capabilities. Experimental measurement or numerical simulation data usually have the characteristics of unstructured grids or non-uniform sensor arrangements, which are not compatible with traditional convolutional neural networks (CNN). Although multilayer perceptrons (MLP) can process unstructured grids, they cannot handle complex situations due to the limitations of their network structure, so their applications are limited. Graph Convolutional Nueral Network (GCN) can process unstructured data and perform convolution operations, but its scale is limited, and the current most advanced applications are limited to processing about 10 5 In addition, the neural network structures mentioned above cannot cope with the situation where the sensor position moves, which limits the scope of application of neural networks in actual on-site reconstruction. Neural networks usually learn models by fitting a large amount of data, but lack the guidance of physical prior knowledge, which may cause the learned models to be physically unreasonable or unexplainable. The lack of physical prior knowledge as a constraint can easily cause the neural network to overfit the data, thereby reducing the generalization ability on new data. Summary of the invention

[0008] The purpose of the present invention is to disclose a flow field reconstruction method and system based on the fusion of sparse sensor data and physical knowledge, so as to realize the global refined flow field reconstruction with a small number of sensor data points.

[0009] To achieve the above purpose, the flow field reconstruction method based on the fusion of sparse sensor data and physical knowledge disclosed in the present invention includes:

[0010] Step S1, constructing a first data set, wherein the first data set includes flow fields under different scenes and different Reynolds numbers;

[0011] Step S2, pre-training is performed according to the data set to obtain an initial model; the initial model adopts a decoupled architecture of the encoder and the decoder; wherein, during the pre-training process, the encoder maps the remaining blocks of the flow field after mask processing together with the position encoding corresponding to each remaining block to a potential representation, and the decoder reconstructs the flow field according to the potential representation, and the corresponding loss function is a data loss function, and the goal of the data loss function is to minimize the error between the original flow field snapshot and the reconstructed flow field;

[0012] Step S3, fine-tuning the initial model for downstream tasks according to the Navier-Stokes equations and the second data set of the target application scenario to obtain a flow field reconstruction model; wherein, in the fine-tuning process, the Navier-Stokes equations and the Reynolds number corresponding to the flow field are embedded in the physical loss function, and then the physical loss function and the data loss function are multiplied by corresponding weights respectively to obtain the total loss function of the flow field reconstruction model; and in the process of embedding the Navier-Stokes equations and the Reynolds number into the physical loss function, numerical differentiation is performed to approximate the corresponding partial derivatives;

[0013] Step S4: input the real-time data collected by the sensors sparsely deployed in the target application scenario into the flow field reconstruction model in combination with the relative position relationship between the position sensors to obtain a reconstructed real-time flow field.

[0014] Preferably, the physical loss function for:

[0015] ;

[0016] in, , , denote the axial velocity, radial velocity, and pressure predicted by the model, respectively. Reynolds number; subscript Express The partial derivative of Express The partial derivative of Express The second-order partial derivative of Express The second-order partial derivative of .

[0017] Preferably, in step S1, the flow field snapshot is obtained in the following manner:

[0018] In the numerical simulation software, the Navier-Stokes equations are used as the control equations, and simulation calculations are performed by setting the boundary conditions, initial conditions and required control parameters of the flow field. During the simulation, the flow field is sampled at set time intervals to obtain snapshots of the flow field corresponding to different time points.

[0019] Preferably, a non-gradient convolution filter is used to implement the processing of approximating the corresponding partial derivative by numerical differentiation. During the processing, the time conversion factor and position conversion factor The values ​​are:

[0020] ;

[0021] ;

[0022] in, and represent the time step and space step respectively.

[0023] To achieve the above objectives, the present invention also discloses a flow field reconstruction system based on the fusion of sparse sensor data and physical knowledge, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method when executing the computer program.

[0024] The present invention has the following beneficial effects:

[0025] (1) The masked autoencoder can effectively learn and extract key feature information from complex flow field data. Its autoencoder structure enables it to learn meaningful representations between input and reconstruction, thereby achieving efficient data reconstruction and reducing the complexity of processing complex flow field data in traditional methods.

[0026] (2) The transfer learning method is used to perform fine-tuning in the downstream task stage, and physical mechanisms are introduced as constraints to improve the precision and accuracy of flow field reconstruction. By embedding physical equations such as fluid dynamics into the model, the model can be effectively guided to learn flow field representations that conform to physical laws, thereby reducing reconstruction errors caused by data noise or anomalies, and helping the model to learn and converge faster.

[0027] (3) The present invention only requires extremely sparsely arranged sensors and is not restricted by fixed sensor positions. It can cope with the situation where the sensor position moves and adapt to flow field reconstruction tasks in different scenarios and requirements. The trained model can reconstruct the real-time flow field by inputting the sensor data and position. Due to the characteristics of deep neural networks, this method has strong flexibility and scalability, and can be customized and optimized according to actual conditions to meet the needs of different application scenarios, thereby significantly reducing labor costs and resource consumption, and making the flow field reconstruction process more economical and efficient.

[0028] In summary, the present invention improves the generalization ability and overall performance of the model.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0031] Figure 1It is a schematic diagram of the overall process of training a flow field reconstruction model based on the fusion of sparse sensor data and physical knowledge disclosed in an embodiment of the present invention;

[0032] Figure 2 is a schematic diagram of the pre-training process disclosed in an embodiment of the present invention, wherein the number of sensors used is only 7;

[0033] Figure 3 is a schematic diagram of a position encoder and a network architecture of an encoder disclosed in an embodiment of the present invention;

[0034] Figure 4 It is a flow chart of a flow field reconstruction method based on the fusion of sparse sensor data and physical knowledge disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0036] Example 1

[0037] This embodiment discloses a flow field reconstruction method based on the fusion of sparse sensor data and physical knowledge. Figure 1 , including the following steps:

[0038] Step 1: Use professional numerical simulation software (such as ANSYS Fluent, OpenFOAM, etc.) to simulate the flow field. By setting the boundary conditions, initial conditions and required control parameters of the flow field, the control equation is the incompressible Navier-Stokes equation:

[0039]

[0040] in, They represent speed, time, density, pressure, and kinematic viscosity respectively; is the vector differential operator.

[0041] During the simulation, the flow field is sampled at set time intervals to ensure that a steady-state flow field snapshot is generated every 50 time steps. Each snapshot contains the complete state of the flow field at different time points, including parameters such as flow velocity and pressure. By saving these steady-state snapshots, the constructed data set has temporal continuity, and each snapshot represents an instantaneous flow field state. The data set contains flow field snapshots under different Reynolds numbers.

[0042] In this embodiment, the entire flow field is reconstructed using sparse sensor data. Specifically, the sensor's measurement value and its corresponding spatial position information are encoded into a potential representation through an encoder. This potential representation contains the key feature information of the flow field. Subsequently, the decoder decodes this potential representation to reconstruct the flow field information of the entire domain. The so-called "flow field snapshot" can be regarded as a storage snapshot corresponding to the flow field information at a certain moment.

[0043] Step 2: Pre-training stage, training is performed through the autoencoder.

[0044] Reference Figure 2 , the network structure includes an encoder and a decoder. The encoder maps the data observed by the sensor to a potential representation, and the decoder reconstructs the global flow field from the potential representation. Among them, the potential representation is the output result of the encoder, usually expressed as a multi-dimensional vector or tensor. Generally, the potential representation can be mathematically understood as a compact representation of the flow field information in a low-dimensional space, which captures the main features and dynamic characteristics of the flow field. The decoder uses this compact representation to reconstruct the complete flow field information.

[0045] The purpose of this step is to learn the potential representation of the system state from a small amount of sensor observation data at a given moment, and the decoder can reconstruct the entire flow field from the potential representation learned by the encoder. To this end, in this embodiment, an asymmetric design is adopted to allow the encoder to operate only on partially observed signals (without using masked tokens), and a lightweight decoder is used to reconstruct the complete flow field from the potential representation and masked tokens. Among them, "token" refers to a grid unit or a patch unit in the flow field. Each token can represent a physical quantity at a specific position in the flow field, such as velocity, pressure, or temperature. "Masked tokens" refers to flow field data points that are masked or unobserved in the input stage to assist the model in learning how to infer complete flow field information from limited observation data. The encoder processes known (unmasked) data points, while the decoder is responsible for reconstructing the entire flow field, including those that were originally masked. In this way, the spatial correlation and physical properties of the flow field are utilized, so that the entire flow field can be accurately reconstructed even with only partial observation data.

[0046] Thus, this step can be specifically divided into the following processing sub-steps:

[0047] Step 2.1 Mask

[0048] The flow field snapshot is divided into regular fixed-size patches, randomly sampled, and the unsampled patches are removed. This step can be regarded as only knowing the values ​​observed by the sensor.

[0049] Step 2.2 Position Encoder

[0050] Position encoding is a key component of this embodiment. By adding specific position encoding vectors, the model (including but not limited to Transformer, etc.) can accurately distinguish and process elements at different positions in the flow field. These position encoding vectors can be generated by a variety of mathematical methods, such as using a combination of sine and cosine functions, or other mathematical expressions that can uniquely identify spatial positions.

[0051] Preferably, the calculation of the position encoding vector can be implemented by the following general formula:

[0052] ;

[0053] in, represents the position index of the position encoding, represents the dimension index in the positional encoding vector, Indicates the dimension of encoding; Indicates frequency; is an integer index used to distinguish parity. Divided into two groups: odd and even. correspond Function, odd correspond function.

[0054] In other words, the above process can be simplified by the formula:

[0055] ;

[0056] in, represents the position code, Represents a position encoder, Indicates the sensor location.

[0057] By incorporating positional encoding, our method is able to better understand and exploit the spatial structure of the flow field, thereby maintaining spatial correlation and physical consistency during reconstruction, which is critical for accurately reconstructing complex flow field structures, especially when only sparse observation data are available.

[0058] In this step, the information of the input flow field is randomly masked, which can be understood as the sensor data information of the input random position. The encoder only processes the visible sensor data, which reduces the amount of calculation and memory usage, and can train a larger encoder at a lower cost; at the same time, the model can better capture the global characteristics of the flow field.

[0059] Step 2.3 Encoder

[0060] Reference Figure 3 , the encoder processes only sampled blocks without using mask tokens. The encoder input is composed of The data observed by the sensor at a certain moment ( Indicates the number of channels recorded by the sensor, such as 1 for temperature and 2 for velocity vector) and the position encoding of the sensor data. The encoder used in this application is ViT (Vision Transformer), which first maps the block into a lower-dimensional embedding space through linear projection and adds position encoding to capture the position information of the image block.

[0061] These vectors are then input into the encoder, which contains multiple feature extraction blocks (a single feature extraction block is a Transformer block well known to those skilled in the art, which includes a self-attention mechanism and a feedforward neural network. The number N of these feature extraction blocks is an adjustable hyperparameter that can be optimized according to specific task requirements), which compresses the input data into a potential representation containing the main features of the input data.

[0062] The main features and functions of the encoder are as follows:

[0063] 1. Sequence length preservation: The encoder keeps the input and output sequence lengths consistent. That is, the number of input sensors is the same as the length of the output features, which can cope with changes in the number of sensors.

[0064] 2. Hierarchical feature extraction: Multiple feature extraction blocks form a hierarchical structure, which enables the model to extract and abstract features layer by layer. Lower layers may focus on local features, while higher layers can capture more global features and relationships.

[0065] 3. Self-attention mechanism: The self-attention mechanism in each feature extraction block enables the model to dynamically focus on different parts of the sequence and effectively capture long-distance dependencies.

[0066] 4. Nonlinear transformation: The nonlinear activation function in the feature extraction block introduces nonlinear transformation to enhance the model's ability to model complex relationships.

[0067] 5. Residual connections and layer normalization: These mechanisms facilitate information flow and gradient propagation in deep networks, making it possible to train deeper models.

[0068] 6. Model capacity and expressiveness: Stacking multiple feature extraction blocks increases the number of parameters and expressiveness of the model, enabling it to learn more complex function mappings.

[0069] 7. Computational efficiency: Despite the increased model complexity, the parallel computing nature of the encoder enables efficient training and inference even for deep models.

[0070] It should be noted that the number of feature extraction blocks should be optimized according to the specific task and available computing resources. Increasing the number of layers does not always lead to performance improvement, so it is necessary to find a balance between model complexity, computing resources, and task requirements.

[0071] With this design, the encoder can effectively extract key features from sparse sensor data, laying the foundation for subsequent flow field reconstruction.

[0072] The above process can be given by the following formula:

[0073] ;

[0074] in, is the potential representation matrix, represents the encoder, Represents sensor data.

[0075] Step 2.4 Decoder

[0076] The input of the decoder consists of the latent representation obtained by the encoder and mask tokens, where mask tokens play an important placeholder role. During the encoding phase of the autoencoder, part of the input will be randomly masked, and mask tokens will be used to replace these masked parts. Specifically, mask tokens is a learnable parameter whose initial value is set to zero.

[0077] The decoder output is a vector representing the value of the block. Therefore, a linear projection is added to the last layer of the decoder, whose number of output channels is equal to the number of flow grids in the block. The decoder output is reassembled into a flow snapshot.

[0078] The above process can be given by the following formula:

[0079] ;

[0080] in, Represented as the value of the reconstructed flow field snapshot grid point, Indicates mask tokens, Represents a decoder, Represented as the position encoding of the flow field snapshot grid points.

[0081] In the pre-training stage, the loss function is the mean square error (MSE) of the grid points between the reconstructed flow field snapshot and the original flow field snapshot.

[0082] Step 3: In the downstream task stage, we add physical mechanisms and fine-tune the model for specific Reynolds numbers. In this stage, we introduce physical mechanisms and fine-tune the model for specific Reynolds numbers. We use the finite difference method to discretize the Navier-Stokes (NS) equations and embed them into the loss function. Through this step, we can ensure that the model satisfies the physical constraints of the fluid dynamics equations while reconstructing the flow field, thereby improving the accuracy and reliability of the model.

[0083] In this step, compared with step 1, it can be obtained that: the data set in step 1 contains the flow fields under different scenarios and different Reynolds numbers, for the convenience of description, it is called the first data set; and the data set in step 3 is constructed according to the specific Reynolds number corresponding to the target application scenario, which can be called the second data set.

[0084] For the partial derivatives in the NS equations, a gradient-free convolution filter is used to represent discrete numerical differentials to approximate the partial derivatives of interest. This step can use the existing finite difference-based filters, which are second-order and fourth-order central difference formulas. The convolution filter is shown below:

[0085] ;

[0086] ;

[0087] in, is the time conversion factor, is the position conversion factor, and Indicates the time step and the space step. Optionally, this step may also use other existing finite difference methods to discretize the Navier-Stokes equations and embed them into the loss function. Such a deformation is obvious to those skilled in the art and will not be described in detail.

[0088] In this step, the total loss function It is divided into two parts:

[0089] ;

[0090] For ease of description, is the data loss function, which corresponds to the loss function generated by the pre-training in the above steps; The new loss function added for the downstream task stage is called the physical loss function. and Indicates the weight of the corresponding item.

[0091] In this embodiment, the physical loss function may be specifically:

[0092] ;

[0093] in, , , denote the axial velocity, radial velocity, and pressure predicted by the model, respectively. Indicates the corresponding Reynolds number. Express The partial derivative of Express The partial derivative of Express The second-order partial derivative of Express The second-order partial derivative of .

[0094] Therefore, this step can be viewed as: adding a physical estimator, introducing physical prior knowledge, enforcing the Navier-Stokes equations with the help of convolution kernels and finite difference methods, and fine-tuning the model by continuously minimizing the loss function of the control equation.

[0095] Step 4: After the training model is completed, you only need to input sensor data and position to quickly reconstruct the flow field.

[0096] Generally, the flow field can be divided into velocity field, pressure field, etc. If the reconstruction is the velocity field, the supporting sensor that needs to be sparsely deployed can be a velocity sensor; if the reconstruction is the pressure field, the supporting sensor that needs to be sparsely deployed can be a pressure sensor. Based on the learning characteristics of deep neural networks, it is of course possible to train a pressure sensor to reconstruct the velocity field. If a hybrid sensor is trained to reconstruct the velocity field and the pressure field, then the hybrid sensor can also be used to reconstruct the velocity field and the pressure field. In this way, in some specific practical applications of this embodiment, seven or eight similar sensors at different positions can reconstruct a flow field with excellent performance, and the global refined process can be reconstructed through extremely sparse sensor deployment. In the model training process, according to the various flexible needs of different users, training and reconstruction can be performed based on a separate pressure field or velocity field, or the pressure field and velocity field can be trained and reconstructed based on the same model at the same time, so that the real-time data collected by the sparsely deployed sensors in the target application scenario are combined with the position of the sensor and input into the flow field reconstruction model to obtain a reconstructed global refined flow field.

[0097] In other words, the flow field reconstruction method of this embodiment has wide applicability and flexibility, and can be applied to the following multiple flow field types and sensor configurations:

[0098] A. Velocity field reconstruction: using sparsely deployed velocity sensors.

[0099] B. Pressure field reconstruction: using sparsely deployed pressure sensors.

[0100] C. Multi-physics field reconstruction: Mixed types of sensors can be used.

[0101] At the same time, based on the learning characteristics of deep neural networks, the method of this embodiment supports the combination of multiple sensor data and target flow fields. For example: a. Reconstructing the velocity field using pressure sensor data; b. Reconstructing the velocity field and pressure field simultaneously using mixed sensor data.

[0102] In practical applications, the method of this embodiment can achieve high-quality flow field reconstruction through extremely sparse sensor deployment, greatly improving the efficiency of sensor deployment. Practice has shown that in some specific application scenarios, only 7-8 similar sensors in different positions are needed to reconstruct a global refined flow field with excellent performance.

[0103] Moreover, the flexibility of the model training in this embodiment is reflected in the following aspects:

[0104] Customized training can be performed according to user needs: a. Single physical quantity training: training and reconstruction of a single pressure field or flow rate field; b. Multi-physical quantity joint training: training and reconstruction of the pressure field and flow rate field at the same time based on the same model.

[0105] In summary, the real-time application process of this embodiment can be summarized as follows: 1. Collect real-time data from sensors sparsely deployed in the target application scenario; 2. Combine the location information of the sensor; 3. Input the data into the trained flow field reconstruction model; 4. Output the reconstructed global refined flow field.

[0106] Through the above method, this embodiment can achieve high-precision, full-domain flow field reconstruction while minimizing hardware deployment, providing an efficient and economical solution for various fluid dynamics applications.

[0107] Example 2

[0108] Corresponding to the above embodiment, this embodiment discloses a flow field reconstruction system based on the fusion of sparse sensor data and physical knowledge, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following is achieved: Figure 4 The following series of steps are shown:

[0109] Step S1: construct a first data set, wherein the first data set includes flow fields under different scenes and different Reynolds numbers.

[0110] In this step, the specific information carrier of the flow field may be the flow field snapshot in the above embodiment, and the specific acquisition method thereof may be: in the numerical simulation software, the Navier-Stokes equation is used as the control equation, and simulation calculations are performed by setting the boundary conditions, initial conditions and required control parameters of the flow field; and during the simulation process, the flow field is sampled at set time intervals to obtain the flow field snapshots corresponding to different time points.

[0111] Step S2, pre-training is performed according to the data set to obtain an initial model; the initial model adopts an architecture in which the encoder and the decoder are decoupled; wherein, during the pre-training process, the encoder maps the remaining blocks of the flow field after mask processing together with the position encodings corresponding to the remaining blocks to a potential representation, and the decoder reconstructs the flow field according to the potential representation, and the corresponding loss function is a data loss function, and the goal of the data loss function is to minimize the error between the original flow field snapshot and the reconstructed flow field.

[0112] Among them, the potential representation is a compressed and abstract data form that captures the most important features of sensor data and helps the model process and understand it more effectively. It has the advantages of reducing computational complexity, improving generalization ability, and noise suppression. For details, please refer to the main features and functions of the above encoders, which will not be described in detail.

[0113] Step S3, fine-tuning the initial model for downstream tasks according to the Navier-Stokes equations and the second data set of the target application scenario to obtain a flow field reconstruction model; wherein, during the fine-tuning process, the Navier-Stokes equations and the Reynolds number corresponding to the flow field are embedded in the physical loss function, and then the physical loss function and the data loss function are multiplied by corresponding weights respectively to obtain the total loss function of the flow field reconstruction model; and in the process of embedding the Navier-Stokes equations and the Reynolds number into the physical loss function, numerical differentiation is performed to approximate the corresponding partial derivatives.

[0114] Preferably, the physical loss function for:

[0115] ;

[0116] in, , , denote the axial velocity, radial velocity, and pressure predicted by the model, respectively. Reynolds number; subscript Express The partial derivative of Express The partial derivative of Express The second-order partial derivative of Express The second-order partial derivative of .

[0117] Step S4: input the real-time data collected by the sensors sparsely deployed in the target application scenario into the flow field reconstruction model in combination with the relative position relationship between the position sensors to obtain a reconstructed real-time flow field.

[0118] In summary, the method and system of the present invention can be regarded as an advanced digital twin technology, which can accurately reconstruct the entire flow field state through a very small amount of sensor data. The core advantages and beneficial effects are as follows:

[0119] (1) The masked autoencoder can effectively learn and extract key feature information from complex flow field data. Its autoencoder structure enables it to learn meaningful representations between input and reconstruction, thereby achieving efficient data reconstruction and reducing the complexity of processing complex flow field data in traditional methods.

[0120] (2) The transfer learning method is used to perform fine-tuning in the downstream task stage, and physical mechanisms are introduced as constraints to improve the precision and accuracy of flow field reconstruction. By embedding physical equations such as fluid dynamics into the model, the model can be effectively guided to learn flow field representations that conform to physical laws, thereby reducing reconstruction errors caused by data noise or anomalies, and helping the model to learn and converge faster.

[0121] (3) The present invention only requires extremely sparsely arranged sensors and is not restricted by fixed sensor positions. It can cope with the situation where the sensor position moves and adapt to flow field reconstruction tasks in different scenarios and requirements. The trained model can reconstruct the real-time flow field by inputting the sensor data and position. Due to the characteristics of deep neural networks, this method has strong flexibility and scalability, and can be customized and optimized according to actual conditions to meet the needs of different application scenarios, thereby significantly reducing labor costs and resource consumption, and making the flow field reconstruction process more economical and efficient.

[0122] In summary, the present invention improves the generalization ability and overall performance of the model.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A flow field reconstruction method based on the fusion of sparse sensor data and physical knowledge, characterized in that: include: Step S1, constructing a first data set, wherein the first data set includes flow fields under different scenes and different Reynolds numbers; Step S2, pre-training is performed according to the data set to obtain an initial model; the initial model adopts a decoupled architecture of the encoder and the decoder; wherein, during the pre-training process, the encoder maps the remaining blocks of the flow field after mask processing together with the position encoding corresponding to each remaining block to a potential representation, and the decoder reconstructs the flow field according to the potential representation, and the corresponding loss function is a data loss function, and the goal of the data loss function is to minimize the error between the original flow field snapshot and the reconstructed flow field; Step S3, fine-tuning the initial model for downstream tasks according to the Navier-Stokes equations and the second data set of the target application scenario to obtain a flow field reconstruction model; wherein, in the fine-tuning process, the Navier-Stokes equations and the Reynolds number corresponding to the flow field are embedded in the physical loss function, and then the physical loss function and the data loss function are multiplied by corresponding weights respectively to obtain the total loss function of the flow field reconstruction model; and in the process of embedding the Navier-Stokes equations and the Reynolds number into the physical loss function, numerical differentiation is performed to approximate the corresponding partial derivatives; Step S4, inputting the real-time data collected by the sparsely deployed sensors in the target application scenario into the flow field reconstruction model in combination with the relative position relationship between the position sensors to obtain a reconstructed real-time flow field; Among them, the physical loss function for: ; in, , , denote the axial velocity, radial velocity, and pressure predicted by the model, respectively. Reynolds number; subscript Express The partial derivative of Express The partial derivative of Express The second-order partial derivative of Express The second-order partial derivative of .

2. The method according to claim 1, characterized in that In step S1, the flow field snapshot is obtained in the following manner: In the numerical simulation software, the Navier-Stokes equations are used as the control equations, and simulation calculations are performed by setting the boundary conditions, initial conditions and required control parameters of the flow field. During the simulation, the flow field is sampled at set time intervals to obtain snapshots of the flow field corresponding to different time points.

3. The method according to claim 1, characterized in that The gradient-free convolution filter is used to implement the processing of approximating the corresponding partial derivatives through numerical differentiation. During the processing, the time conversion factor and position conversion factor The values ​​are: ; ; in, and represent the time step and space step respectively.

4. A flow field reconstruction system based on the fusion of sparse sensor data and physical knowledge, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 3 is implemented.

Citation Information

Patent Citations

  • Turbulence super-resolution method based on Transform deep learning architecture

    CN116468602A

  • Plant protection unmanned aerial vehicle downwash wind field reconstruction method based on physical information neural network

    CN115239018A

  • Flow field feature extraction method based on deep convolution auto-encoder

    CN115311530A

  • Method for reconstructing flow field containing unknown parameters from passive scalar

    CN116050511A

  • Flow field time history deep learning prediction method considering flow control equation constraint

    CN116127844A