Flow field data compression and restoration method based on super-resolution reconstruction

Abstract

The application provides a flow field data compression and recovery method based on super-resolution reconstruction, which comprises the following steps: (1) collecting high-resolution flow fields at different times generated by experiments or simulations when the flow develops to a statistical stable stage; (2) determining the calculation geometry and boundary conditions of the high-precision flow field, and carrying out unsteady large eddy simulation to obtain low-resolution flow field data of the high-resolution flow field in time and space; (3) taking the low-resolution flow field in step (2) and the high-resolution flow field in step (1) as inputs at the same time, and generating a flow field restorer through unsupervised machine learning; (4) re-executing the large eddy simulation in step (2) to obtain low-resolution flow fields at different times, and obtaining high-resolution flow fields through the flow field restorer in step (3). In the test of isotropic turbulence, the compression ratio of the application is much better than that of other methods, and the data recovery process only needs to complete mapping through a neural network, and the calculation resource cost is lower than that of other methods.

Description

Technical Field

[0001] This invention relates to the field of flow field data processing technology, and specifically to a method for compressing and restoring flow field data based on super-resolution reconstruction. Background Technology

[0002] Compression and restoration of flow fields is an important technique for reducing the storage and transmission costs of databases. Modern high-precision simulation methods and high-resolution experiments yield massive amounts of physical field information: direct numerical simulations of fluid mechanics and the Large Hadron Collider generate petabytes of data per second. The storage cost of these massive amounts of data is extremely high, even exceeding the simulation and experiment costs themselves; managing and sharing them through data centers is also virtually infeasible. To overcome the challenges of storing, sharing, and reprocessing physical fields in industry and scientific research, it is necessary to compress the obtained physical fields themselves.

[0003] Current physical field compression techniques are mainly classified into two categories:

[0004] (1) Based on classic general data compression techniques. This type of compression scheme includes string compression techniques and image / video encoders; string compression techniques include bzip2 data compression techniques based on block sorting compression algorithms and rar compression techniques based on variable-length data blocks; image encoder schemes include lossy image compression techniques such as JPEG and PNG; video encoder schemes include video codecs such as HEVC and AV1; the aforementioned schemes are suitable for data types without constraints and are widely used in data centers; the main problems are: the compression ratio of general data compression techniques is limited; the flow field is usually binary files, which cannot utilize string-based compression techniques; and the compression of image and video encoders is around 10%.

[0005] (2) Compression based on the physical characteristics of the data. This type of compression scheme takes into account the characteristics of the physical field data itself and uses the partial differential equations of the physical field to restore the data; for example, some flow databases only retain the skeleton information in the computational domain, and the secondary skeleton region is restored by solving the Navier-Stokes equations. The main problem is that data decompression based on partial differential equations is very computationally expensive and time-consuming.

[0006] In conclusion, it is necessary to further innovate existing technologies. Summary of the Invention

[0007] To address the problems existing in the above-mentioned background technology, this invention proposes a flow field data compression and restoration method based on super-resolution reconstruction. The method is reasonable in concept and simple in process. In various tests of isotropic turbulence, the compression ratio can reach 1 / 64, which is far better than other methods. The data restoration process only needs to complete the mapping from the compressed flow field to the original flow field through a neural network, and the computational resource cost is less than other methods.

[0008] To address the aforementioned technical problems, this invention provides a method for compressing and restoring flow field data based on super-resolution reconstruction, which specifically includes the following steps:

[0009] (1) For specific flow problems, collect high-resolution flow fields at different times generated by experiments or simulations when the flow develops to the statistical steady stage;

[0010] (2) Based on flow physics, determine the computational geometry and boundary conditions of the high-precision flow field, and then use a fluid dynamics solver to carry out unsteady large eddy simulation to obtain the low-resolution flow field in time and space for the specific flow problem in step (1).

[0011] (3) Using the low-resolution flow field in step (2) and the high-resolution flow field in step (1) as input data, a flow field restorer based on super-resolution reconstruction is trained through unsupervised machine learning.

[0012] (4) Data restoration

[0013] Repeat the large eddy simulation of the specific computational geometry, boundary conditions and subgrid model in step (2) above to obtain the low-resolution flow field at different times. Then, obtain the high-resolution flow field through the flow field restorer based on super-resolution reconstruction in step (3) above.

[0014] The flow field data compression and restoration method based on super-resolution reconstruction includes: the computational geometry in step (2) includes wall geometry parameters and computational domain range; the wall geometry parameters are the corresponding walls of the high-resolution flow field; the computational domain range is selected to include the walls and ensure that the flow outside the domain has no effect on the results inside the domain; the boundary conditions are simplified from simulation conditions to non-penetrating and non-slip conditions or non-penetrating and slip conditions; the subgrid model uses the mature Vreman model.

[0015] The flow field data compression and restoration method based on super-resolution reconstruction, wherein: the subgrid model in step (2) can also adopt other subgrid models suitable for the current flow problem, including but not limited to WALE, Q criterion and dynamic Smagorinsky model.

[0016] The flow field data compression and restoration method based on super-resolution reconstruction, wherein the specific process of generating the super-resolution reconstructor through unsupervised machine learning in step (3) is as follows:

[0017] (3.1) Construct two neural network-based generators, namely the restoration generator SR for upscaling and the compression generator LR for downscaling; and the generators are constructed in the form of an enhanced super-resolution generative adversarial network.

[0018] (3.2) Next, construct two discriminators corresponding to the generator in step (3.1) above: a discriminator D for upscaling the result. SR Discriminator D with reduced resolution results LR The discriminator D SR The discriminator D is used to identify whether the restored result is authentic. LR Used to determine whether the compressed result is reasonable;

[0019] (3.3) Next, the two generative adversarial networks formed in steps (3.1) and (3.2) above are combined into a cyclic generative adversarial network;

[0020] (3.4) Next, the recurrent generative adversarial network obtained in step (3.3) is trained. Based on the high-resolution flow field in step (1) and the low-resolution flow field in step (2), the generator in step (3.1) and the discriminator in step (3.2) are evaluated by the loss function. The generator and discriminator are continuously optimized until the loss function is less than the preset value.

[0021] The flow field data compression and restoration method based on super-resolution reconstruction, wherein the specific training process of step (3.4) is as follows: the low-resolution flow field compressed in step (2) is first restored by the restoration generator SR, and then pseudo-compressed by the compression generator LR; the high-resolution flow field in step (1) is first restored by the compression generator LR, and then pseudo-restored by the restoration generator SR; finally, the objective function is calculated based on the difference between the pseudo-flow field and the real flow field, and the restoration generator SR, the compression generator LR, and the discriminator D are optimized. SR and D LR .

[0022] The flow field data compression and restoration method based on super-resolution reconstruction, wherein: the generator in step (3.1) can also be constructed in other forms besides enhanced super-resolution generative adversarial networks.

[0023] The flow field data compression and restoration method based on super-resolution reconstruction, wherein the loss function in the unsupervised machine learning process in step (3) includes the following three parts, specifically:

[0024] ① The adversarial relationship between the generator and the discriminator in the two generative adversarial networks in steps (3.1) and (3.2);

[0025] ② The difference between the high-resolution flow field in step (1) after being processed by the compression generator SR and the restoration generator LR and itself, and the difference between the low-resolution flow field in step (2) after being processed by the restoration generator LR and the compression generator LR and itself.

[0026] ③ Additional physical and non-physical constraints.

[0027] The flow field data compression and restoration method based on super-resolution reconstruction includes, in step ③, additional physical constraints that may include, but are not limited to, momentum conservation, energy conservation, smoothness, and other constraints on non-homogeneous data.

[0028] By adopting the above technical solution, the present invention has the following beneficial effects:

[0029] This invention presents a reasonable and simple flow field data compression and restoration method based on super-resolution reconstruction. In tests of various isotropic turbulent flows, the compression ratio reached 1 / 64, significantly better than other methods. The data restoration process only requires mapping through a neural network-based restoration generator. The computational resource cost of mapping through the restoration generator is lower than other methods. Specifically, the computational resources (floating-point operations) required for the restoration process from compressed flow field to high-precision flow field based on a trained neural network are less than those for flow field compression techniques based on partial differential equation solving and other general data compression techniques. In particular, it employs unsupervised machine learning-based super-resolution reconstruction technology to compress high-resolution flow field data generated in simulations and experiments; it uses two generative adversarial networks (GANs) to increase and decrease resolution to achieve unsupervised machine learning; during machine learning training, adding additional physical and non-physical constraints to the loss function yields better restoration results; and during machine learning training, constraints on non-homogeneous data can be added to the objective function, allowing for the simultaneous compression of non-homogeneous data sources (such as simulated flow fields across the entire computational domain and observational data from specific regions). Attached Figure Description

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the data compression process based on super-resolution reconstruction in the flow field data compression and restoration method based on super-resolution reconstruction of the present invention;

[0032] Figure 2 This is a schematic diagram of the data restoration process based on super-resolution reconstruction in the flow field data compression and restoration method based on super-resolution reconstruction of the present invention;

[0033] Figure 3 This is a schematic diagram of the unsupervised machine learning architecture used for super-resolution reconstruction in the flow field data compression and restoration method based on super-resolution reconstruction of the present invention. Detailed Implementation

[0034] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The present invention will be further explained below with reference to specific embodiments.

[0036] like Figure 1 As shown in the figure, the flow field data compression and restoration method based on super-resolution reconstruction provided in this embodiment specifically includes the following steps:

[0037] (1) For specific flow problems, collect high-resolution flow fields at different times generated by experiments or simulations when the flow develops to the statistically stable stage;

[0038] (2) Based on flow physics, determine the computational geometry and boundary conditions of the high-precision flow field; then, through the fluid dynamics solver, after giving the computational geometry and boundary conditions, select a suitable subgrid model and carry out unsteady large eddy simulation. After the flow field develops to the statistical steady stage, obtain the low-resolution flow field in time and space of the specific flow problem in step (1).

[0039] In step (2) above, the computational geometry should include wall geometry parameters and the computational domain. The wall geometry parameters are the corresponding walls in the high-resolution flow field (experimental or simulation results). The computational domain should include the walls and ensure that the flow outside the domain has no impact on the results inside the domain. The boundary conditions are simplified from simulation conditions to non-penetrating and non-slip conditions or non-penetrating and slip conditions. The mature Vreman model is selected for the subgrid model. In addition, other subgrid models suitable for the current flow problem can also be used in step (2) above, including but not limited to WALE, Q criterion and dynamic Smagorinsky model.

[0040] (3) Using the low-resolution flow field obtained from the large eddy simulation in step (2) and the high-resolution flow field generated by experiments or numerical methods in step (1) as input data, a flow field restorer based on super-resolution reconstruction is trained through unsupervised machine learning; such as Figure 3 As shown, the specific process of generating a super-resolution reconstructor includes the following steps:

[0041] (3.1) Construct two neural network-based generators: a restoration generator SR for upscaling and a compression generator LR for downscaling; the generators are constructed using an enhanced super-resolution generative adversarial network.

[0042] (3.2) Next, construct two discriminators corresponding to the generator in step (3.1) above: a discriminator D for upscaling the result. SR Discriminator D with reduced resolution results LR The discriminator D SR The discriminator D is used to identify whether the restored result is authentic. LR Used to determine whether the compressed result is reasonable;

[0043] (3.3) Next, the two generative adversarial networks formed in steps (3.1) and (3.2) above are combined into a cyclic generative adversarial network;

[0044] (3.4) Next, the recurrent generative adversarial network obtained in step (3.3) is trained. Based on a large number of high-resolution flow fields in step (1) and low-resolution flow fields in step (2), the generator in step (3.1) and the discriminator in step (3.2) are evaluated by the loss function. The generator and discriminator are continuously optimized until the loss function is less than the preset value.

[0045] The generator in step (3.1) above can also be constructed in other ways besides Enhanced Super-Resolution Generative Adversarial Network (ESRGAN).

[0046] The specific training process in step (3.4) above is as follows: the low-resolution flow field compressed in step (2) is first restored by SR, and then pseudo-compressed by LR; the high-resolution flow field in step (1) is first restored by LR, and then pseudo-restored by SR; finally, the objective function is calculated based on the difference between the pseudo-flow field and the real flow field, and the restoration generator SR, the compression generator LR, and the discriminator D are optimized. SR and D LR .

[0047] The loss function in step (3.4) above comprises three parts, specifically:

[0048] ① The adversarial relationship between the generator and the discriminator in the two generative adversarial networks in steps (3.1) and (3.2);

[0049] ② The difference between the high-resolution flow field in step (1) after being processed by the compression generator SR and the restoration generator LR and itself, and the difference between the low-resolution flow field in step (2) after being processed by the restoration generator LR and the compression generator LR and itself.

[0050] ③ Additional physical and non-physical constraints (such as divergence-free constraints); among which, additional physical constraints may also include, but are not limited to, other constraints such as momentum conservation, energy conservation, smoothness, and non-homogeneous data.

[0051] The above steps (1)-(3) are the specific process of data compression; in which, the high-precision flow field of the specific flow problem in step (1) has been compressed into the low-resolution flow field in step (2) and the flow field restorer based on super-resolution reconstruction obtained by training the cyclic generative adversarial network in step (3).

[0052] (4) Data restoration

[0053] The large eddy simulation of the specific computational geometry, boundary conditions and subgrid model in step (2) is re-executed to obtain the low-resolution flow field at different times. Then, the high-resolution flow field is generated by the flow field restorer based on super-resolution reconstruction in step (3).

[0054] This invention has a reasonable concept and a simple process. In various tests of isotropic turbulence, the compression ratio reached 1 / 64, which is far better than other methods. The data restoration process only needs to complete the mapping through the neural network weight matrix, and the computational resource consumption is better than other methods, making it suitable for promotion and application.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for compressing and restoring flow field data based on super-resolution reconstruction, characterized in that... Specifically, it includes the following steps: (1) For specific flow problems, collect high-resolution flow fields at different times generated by experiments or simulations when the flow develops to the statistical steady stage; (2) Based on flow physics, determine the computational geometry and boundary conditions of the high-precision flow field, and then use a fluid dynamics solver to carry out unsteady large eddy simulation to obtain the low-resolution flow field in time and space for the specific flow problem in step (1); (3) Using the low-resolution flow field in step (2) and the high-resolution flow field in step (1) as input data, a flow field restorer based on super-resolution reconstruction is trained through unsupervised machine learning; the specific process of generating the super-resolution reconstructor through unsupervised machine learning is as follows: (3.1) Construct two neural network-based generators, namely the restoration generator SR for upscaling and the compression generator LR for downscaling; and the generators are constructed in the form of an enhanced super-resolution generative adversarial network. (3.2) Next, construct two discriminators corresponding to the generator in step (3.1) above: a discriminator D for upscaling the result. SR Discriminator D with reduced resolution results LR The discriminator D SR The discriminator D is used to identify whether the restored result is authentic. LR Used to determine whether the compressed result is reasonable; (3.3) Next, the two generative adversarial networks formed in steps (3.1) and (3.2) above are combined into a cyclic generative adversarial network; (3.4) Next, the recurrent generative adversarial network obtained in step (3.3) is trained. Based on the high-resolution flow field in step (1) and the low-resolution flow field in step (2), the generator in step (3.1) and the discriminator in step (3.2) are evaluated by the loss function. The generator and discriminator are continuously optimized until the loss function is less than the preset value. (4) Data restoration: Re-execute the large eddy simulation of the specific computational geometry, boundary conditions and subgrid model in step (2) above to obtain the low-resolution flow field at different times, and then obtain the high-resolution flow field through the flow field restorer based on super-resolution reconstruction in step (3) above. The loss function in the unsupervised machine learning process in step (3) includes the following three parts: ① The adversarial relationship between the generator and the discriminator in the two generative adversarial networks in steps (3.1) and (3.2); ② The difference between the high-resolution flow field in step (1) after being processed by the compression generator SR and the restoration generator LR and itself, and the difference between the low-resolution flow field in step (2) after being processed by the restoration generator LR and the compression generator LR and itself; ③ Additional physical and non-physical constraints; The additional physical constraints in step ③ include momentum conservation, energy conservation, smoothness, and other constraints on non-homogeneous data.

2. The flow field data compression and restoration method based on super-resolution reconstruction as described in claim 1, characterized in that: The computational geometry in step (2) includes wall geometry parameters and computational domain range; the wall geometry parameters are the corresponding walls of the high-resolution flow field, and the computational domain range is selected to include the walls and ensure that the flow outside the domain has no effect on the results inside the domain; the boundary conditions are simplified from simulation conditions to non-penetrating and non-slip conditions or non-penetrating and slip conditions; the subgrid model uses the mature Vreman model.

3. The flow field data compression and restoration method based on super-resolution reconstruction as described in claim 1, characterized in that: The subgrid model in step (2) can be other subgrid models suitable for the current flow problem, including WALE, Q criterion and dynamic Smagorinsky model.

4. The flow field data compression and restoration method based on super-resolution reconstruction as described in claim 1, characterized in that... The specific training process of step (3.4) is as follows: the low-resolution flow field compressed in step (2) is first restored by the restoration generator SR, and then pseudo-compressed by the compression generator LR. The high-resolution flow field from step (1) is then processed by a compression generator (LR) to obtain a low-resolution flow field, and then by a restoration generator (SR) to obtain a pseudo-restored flow field. Finally, the objective function is calculated based on the difference between the pseudo-flow field and the real flow field, and the restoration generator (SR), compression generator (LR), and discriminator (D) are optimized. SR and D LR .

5. The flow field data compression and restoration method based on super-resolution reconstruction as described in claim 1, characterized in that: The generator in step (3.1) can be constructed in other ways besides enhanced super-resolution generative adversarial networks.

Images