A state estimation method based on GA-PCAE and LSTM

Abstract

This invention discloses a state estimation method based on GA-PCAE and LSTM, belonging to the field of deformable body state estimation technology, including the following steps: S1, establishing a transient database of deformable structures based on CFD; S2, constructing and training a PointNet-based geometric field joint autoencoder (GA-PCAE); S3, establishing a nonlinear mapping from sensor data to low-order representation based on an LSTM network; S4, reconstructing the flow field from sensor signals based on a GA-PCAE network decoder. This invention innovatively proposes a deformable body flow field reconstruction method based on GA-PCAE, a three-stage training method, and a state estimation method utilizing LSTM and GA-PCAE, solving the problem of high-precision reconstruction of complex boundaries in deformable bodies using traditional flow field reconstruction methods.

Description

Technical Field

[0001] This invention relates to the field of deformable body state estimation technology, and in particular to a state estimation method based on GA-PCAE and LSTM. Background Technology

[0002] Modeling, prediction, and control of complex fluid flows are of great significance for fields such as biomimetic propulsion, aerospace, and biomedicine. State estimation, which aims to reconstruct the complete flow field from sparse and noisy observations, is a key step in real-time monitoring and adaptive control.

[0003] To achieve state estimation of deformable bodies, deep learning techniques directly learn a nonlinear mapping from sparse, noisy observation sequences to "observation → latent state → full field," thus reducing the dependence on explicit time models and turbulent closure. However, in deformable / FSI scenarios, the domain shape and mesh change over time, often requiring sample interpolation to a uniform reference mesh before retraining. This approach easily introduces stepped walls and flow / solid domain ambiguity, weakening the consistency of near-wall gradients and boundary conditions. To realize a state estimation method for deformable bodies, it is necessary to develop a flow field reconstruction method and state estimation strategy that does not pre-set a mesh.

[0004] Deep learning-based state estimation techniques primarily utilize the nonlinear mapping relationship between sensors and physical fields. Neural networks are trained on different physical fields corresponding to various sensors to achieve physical field reconstruction and temporal estimation for sparse sensors. Currently, most state estimation research operates on fixed grids. State estimation for complex deformable structures using a small number of sensors is achieved by interpolating into a standard grid, but this interpolation can lead to boundary staircases in complex structures. Developing new flow field reconstruction methods and state estimation techniques to achieve state estimation for complex deformable coefficient sensors remains a challenge.

[0005] In the field of deep learning-based state estimation techniques, while existing technologies can be applied to state estimation (e.g., patent CN202510855412.1, "A Method for Predicting Unsteady Flow Fields in Turbines Under Unknown Operating Conditions," which uses mode decomposition to measure unsteady flow fields), the database still relies on a regular grid as an assumption, and complex boundaries are interpolated as stepped walls). Existing state estimation methods for handling moving boundaries (e.g., patent CN202210897466.0, "Data-Driven Method, Device, and Storage Medium for Reconstructing Flow Fields at Moving Boundaries") achieve state estimation using convolutional autoencoders and immersion boundary methods. However, the convolution behavior is still implemented on a regular grid, which still has limitations in handling complex boundaries.

[0006] In summary, existing state estimation techniques do not involve the reconstruction and prediction of multiphysics fields for complex deformable bodies. Summary of the Invention

[0007] The purpose of this invention is to provide a state estimation method based on GA-PCAE and LSTM to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a state estimation method based on GA-PCAE and LSTM, comprising the following steps: S1. Establish a transient database of deformable structures based on computational fluid dynamics (CFD); S2. A geometric field joint autoencoder based on PointNet point cloud neural network is constructed by building a GA-PCAE neural network and training it using a three-stage training method. S3. Based on the Long Short-Term Memory (LSTM) network, establish a nonlinear mapping from sensor data to low-order representation; S4. Based on the GA-PCAE neural network decoder, the sensor signal is reconstructed into the flow field.

[0009] Preferably, the specific steps of S1 are as follows: S11. Use ANSYS Fluent to perform CFD environment simulation and establish a transient database of deformable structures; S12. Divide the transient database into training set, test set and validation set according to a preset ratio.

[0010] Preferably, S11 uses a downflow wave equation to describe the motion of the airfoil centerline, and the specific formula of the downflow wave equation is as follows: ; in, A ( x ) is the floating amplitude function x The x-coordinate of the fish's body. The wavelength of the fish's body undulation. t For the swing time, T The oscillation period.

[0011] Preferably, the specific steps of S2 are as follows: The S21 and GA-PCAE neural networks encode point cloud data as a latent representation through encoders and decoders to reconstruct spatial coordinates and physical fields. S22. Initialize the weights and biases of the GA-PCAE neural network, and construct the loss function using geometric and physical terms; S23. The GA-PCAE neural network is trained using a three-stage training method with dual learning rates for both the geometric head and the physical head.

[0012] Preferably, the specific steps of S21 are as follows: S211. The input point cloud with shape B×N×5 is operated through the encoder; S212. Based on PointNet, apply each point number to the shared MLP; S213. Connect global max pooling and global average pooling to form a high-dimensional descriptor, and then linearly compress the high-dimensional descriptor into a latent vector. S214. Using the latent vector as the decoder input, the input latent vector is upgraded to an intermediate dimension by utilizing the shared backbone. S215. Using the intermediate dimension as input, and utilizing the geometry head and physics head, output the coordinates and physical quantities of the physical field.

[0013] Preferably, the specific steps of the three-stage training method in S23 are as follows: S231, Preheating Geometry: Geometric features and multiphysics features are covered and aligned by biased decomposition Chamfer driving. S232, Task Balancing: Combined with mild Laplace regularization, it suppresses oscillations in predicted physics values; S233, Cosine Annealing Refinement: Update the weights of the GA-PCAE neural network based on the gradient calculated from the loss function; S234. Repeat the above steps until the network converges or reaches the predetermined number of training rounds to obtain a converged GA-PCAE neural network.

[0014] Preferably, the specific steps of S3 are as follows: S31. Collect velocity and pressure values ​​at different locations during the CFD simulation process to simulate sensor signals; S32. Load the GA-PCAE neural network and its network weights and biases to obtain low-dimensional representations under different periods and phases; S33. Using continuous sensor signals as LSTM input and low-dimensional representations of future times as output, establish time series prediction and train the LSTM until convergence or the maximum preset number of rounds is reached.

[0015] Preferably, the specific steps of S4 are as follows: S41. Load the weights and biases of the LSTM and GA-PCAE neural networks, use the time-series signals from the sensor as the input to the LSTM network, and output a low-dimensional representation of the flow field at future time steps. S42. Using the low-dimensional representation of the LSTM output as the input of the GA-PCAE neural network, the output is the predicted flow field point cloud structure and its physical quantities of the deformable body, thereby achieving accurate reconstruction of the physical field of the deformable body.

[0016] Therefore, the state estimation method based on GA-PCAE and LSTM described above, as used in this invention, has the following beneficial effects: (1) This invention proposes a deformable body flow field reconstruction method based on GA-PCAE, which solves the problem of high-precision reconstruction of complex boundaries of deformable bodies by traditional flow field reconstruction methods.

[0017] (2) This invention proposes a three-segment neural network training method for flow field reconstruction of GA-PCAE, which significantly reduces the problems of multi-task gradient conflict and point cloud collapse.

[0018] (3) This invention proposes a sparse sensor state estimation method based on LSTM and GA-PCAE, which realizes accurate reconstruction of the future flow field through sensor time series signals.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of a state estimation method based on GA-PCAE and LSTM according to the present invention. Figure 2 This is a schematic diagram illustrating the implementation of a CFD simulation based on a state estimation method using GA-PCAE and LSTM according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of GA-PCAE, which is an embodiment of the state estimation method based on GA-PCAE and LSTM according to the present invention. Figure 4 This is a state estimation framework diagram using LSTM and GA-PCAE, which is an embodiment of the state estimation method based on GA-PCAE and LSTM according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0023] Example Please see Figures 1-4 This invention provides a state estimation method based on GA-PCAE and LSTM, comprising the following steps: S1. A transient database of deformable structures was established using computational fluid dynamics (CFD) methods, and simulations were performed using a CFD environment comprised of the fluid simulation software ANSYS Fluent.

[0024] Taking the undulating motion of a flexible fish as an example, such as Figure 2 The diagram shows a schematic of the flow structure around a flexible airfoil. A two-dimensional NACA0012 airfoil with chord L is placed in a uniform flow with velocity U in the x-direction. The leading edge of the airfoil is located at the origin of the coordinate system. The motion of the airfoil's centerline is described by the following traveling wave equation: ; in, A ( x ) is the floating amplitude function, generally defined as A ( x )= a 0+ a 1 x + a 2 x 2 In this embodiment, a 0= a 2=0, a 1 = 0.1, used to determine the airfoil of the wave; x The x-axis represents the fish's body. The wavelength of the fish's body undulation. t For the swing time, T The oscillation period.

[0025] To train the model, 1000 snapshots were generated using ANSYS Fluent as the training set, containing four cycles. 750 snapshots were used as the test set, and 250 snapshots as the validation set. We randomly downsampled the fluid domain of the original data to obtain 15004 points.

[0026] S2. Construct the GA-PCAE method based on PointNet, a point cloud neural network.

[0027] GA-PCAE encodes point cloud data into a latent representation and then reconstructs spatial coordinates and physical fields. Specifically, the encoder operates on an input point cloud of shape B×N×5, where B is the batch size, N is the number of points, and the five channels correspond to two-dimensional coordinates (x, y) and physical variables (u, v, p). Following the PointNet design, a shared MLP is applied to each point. Global max pooling and global average pooling are then concatenated to form a high-dimensional descriptor, which is linearly compressed into a latent vector.

[0028] In the decoder section, the latent vector is used as the decoder input. It first passes through a shared backbone to raise the dimensionality of the input latent vector to an intermediate dimension. Then, the intermediate dimension is used as input and passes through two task-specific head networks (geometric head and physical head, respectively). The geometric head is used to output the coordinates of the physical field, and the physical head is used to output the physical quantities of the physical field. This "shared backbone and dedicated head" design is driven by three considerations: (i) parameter efficiency and stability: the backbone captures common features of sample-level shape and field, preventing the two tasks from learning incompatible global structures from scratch; (ii) bias distribution: geometric and physical biases are aligned in the trunk but decoupled in the head, reflecting the training objective of "aligning geometry first, then regressing properties"; (iii) gradient coordination: by coordinating the head gradients through the shared backbone, soft competition between tasks is achieved, reducing single-task overfitting.

[0029] Initialize the weights and biases of the GA-PCAE neural network, using the deformable database obtained in step one as input. The reconstruction objective of GA-PCAE consists of both geometric and physical terms: the geometric term is used to constrain the alignment of the point set, using the chamfered Chamfer distance as a metric. Find the nearest point to each reconstructed point, and calculate the mean squared error between the reconstructed point and the original point. These two terms, in a certain proportion, constitute the loss function of the neural network, used to train the network.

[0030] During training, a three-stage curriculum-based scheduling approach is employed, using dual learning rates for both the geometry head and the physics head: "geometry warm-up" → "task balancing" → "cosine annealing refinement." A biased decomposition Chamfer is used for early coverage and alignment. In the mid-stage, a mild Laplacian regularization is applied to suppress oscillations in the predicted physics values. The weights of the neural network are updated based on the gradient calculated from the loss function. This process is repeated until the network converges or reaches the predetermined number of training epochs, resulting in a converged GA-PCAE neural network.

[0031] S3. Based on the Long Short-Term Memory (LSTM) network, establish a nonlinear mapping from sensor data to low-order representation.

[0032] Velocity and pressure values ​​at different locations during CFD simulations are collected to simulate sensor signals. A GA-PCAE neural network, along with its weights and biases, is loaded to obtain low-dimensional representations at different periods and phases. Each representation corresponds to a latent description of the flow at a specific time step. These latent states capture the fundamental geometric and physical characteristics of the watershed, providing an effective target space for subsequent temporal modeling. Using continuous sensor signals as input to an LSTM network and low-dimensional representations of future times as output, a temporal prediction from sensor signals to future low-dimensional representations is established. The neural network is trained until convergence or the maximum number of epochs is reached.

[0033] S4. Based on the GA-PCAE neural network decoder, the sensor signal is reconstructed into the flow field.

[0034] The GA-PCAE neural network and LSTM neural network obtained in the above steps are as follows: Figure 4 As shown. The weights and biases of the LSTM neural network are loaded. The weights and biases of the GA-PCAE neural network are loaded. Using the time-series signals from the test set sensors as input to the LSTM network, a low-dimensional representation of the flow field at future time points is output.

[0035] Furthermore, the low-dimensional representation output by the LSTM neural network is used as the input to the GA-PCAE neural network, which outputs the predicted flow field point cloud structure and its physical quantities of the deformable body. This achieves accurate reconstruction of the physical field of the deformable body.

[0036] Therefore, this invention adopts the above-mentioned state estimation method based on GA-PCAE and LSTM, and innovatively proposes a deformable body flow field reconstruction method based on GA-PCAE, a training method for GA-PCAE neural network using a three-stage training method, and a state estimation method using LSTM network and GA-PCAE neural network, thus solving the problem of high-precision reconstruction of complex boundaries of deformable bodies by traditional flow field reconstruction methods.

[0037] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A state estimation method based on GA-PCAE and LSTM, characterized in that, Includes the following steps: S1. Establish a transient database of deformable structures based on computational fluid dynamics (CFD); S2. A geometric field joint autoencoder based on PointNet point cloud neural network is constructed by building a GA-PCAE neural network and training it using a three-stage training method. S3. Based on the Long Short-Term Memory (LSTM) network, establish a nonlinear mapping from sensor data to low-order representation; S4. Based on the GA-PCAE neural network decoder, the sensor signal is reconstructed into the flow field.

2. The state estimation method based on GA-PCAE and LSTM according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Use ANSYS Fluent to perform CFD environment simulation and establish a transient database of deformable structures; S12. Divide the transient database into training set, test set and validation set according to a preset ratio.

3. The state estimation method based on GA-PCAE and LSTM according to claim 2, characterized in that, S11 uses a descending wave equation to describe the motion of the airfoil centerline. The specific formula of the descending wave equation is as follows: ； in, A ( x ) is the floating amplitude function. x The x-coordinate of the fish's body. The wavelength of the fish's body undulation. t For the swing time, T The oscillation period.

4. The state estimation method based on GA-PCAE and LSTM according to claim 1, characterized in that, The specific steps of S2 are as follows: The S21 and GA-PCAE neural networks encode point cloud data as a latent representation through encoders and decoders to reconstruct spatial coordinates and physical fields. S22. Initialize the weights and biases of the GA-PCAE neural network, and construct the loss function using geometric and physical terms; S23. The GA-PCAE neural network is trained using a three-stage training method with dual learning rates for both the geometric head and the physical head.

5. The state estimation method based on GA-PCAE and LSTM according to claim 4, characterized in that, The specific steps of S21 are as follows: S211. The input point cloud with shape B×N×5 is operated through the encoder; S212. Based on PointNet, apply each point number to the shared MLP; S213. Connect global max pooling and global average pooling to form a high-dimensional descriptor, and then linearly compress the high-dimensional descriptor into a latent vector. S214. Using the latent vector as the decoder input, the input latent vector is upgraded to an intermediate dimension by utilizing the shared backbone. S215. Using the intermediate dimension as input, and utilizing the geometry head and physics head, output the coordinates and physical quantities of the physical field.

6. The state estimation method based on GA-PCAE and LSTM according to claim 4, characterized in that, The specific steps of the three-stage training method in S23 are as follows: S231, Preheating Geometry: Geometric features and multiphysics features are covered and aligned by biased decomposition Chamfer driving. S232, Task Balancing: Combined with mild Laplace regularization, it suppresses oscillations in predicted physics values; S233, Cosine Annealing Refinement: Update the weights of the GA-PCAE neural network based on the gradient calculated from the loss function; S234. Repeat the above steps until the network converges or reaches the predetermined number of training rounds to obtain a converged GA-PCAE neural network.

7. The state estimation method based on GA-PCAE and LSTM according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. Collect velocity and pressure values ​​at different locations during the CFD simulation process to simulate sensor signals; S32. Load the GA-PCAE neural network and its network weights and biases to obtain low-dimensional representations under different periods and phases; S33. Using continuous sensor signals as LSTM input and low-dimensional representations of future times as output, establish time series prediction and train the LSTM until convergence or the maximum preset number of rounds is reached.

8. The state estimation method based on GA-PCAE and LSTM according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Load the weights and biases of the LSTM and GA-PCAE neural networks, use the time-series signals from the sensor as the input to the LSTM network, and output a low-dimensional representation of the flow field at future time steps. S42. Using the low-dimensional representation of the LSTM output as the input of the GA-PCAE neural network, the output is the predicted flow field point cloud structure and its physical quantities of the deformable body, thereby achieving accurate reconstruction of the physical field of the deformable body.

Images