Hydrodynamic estimation method based on pressure image, electronic equipment and storage medium

Abstract

The invention discloses a hydrodynamic estimation method based on a pressure image, electronic equipment and a storage medium. The method comprises the following steps: acquiring pressure distribution data of a moving target surface and corresponding hydrodynamic data; mapping the pressure distribution data to a two-dimensional image matrix to generate a pressure image of the moving target; taking the hydrodynamic data as labels to associate the corresponding pressure images, and constructing a data set; and inputting the data set into a pre-trained convolutional neural network model, extracting pressure features of the pressure image through the model, and outputting a hydrodynamic estimation value. According to the method, many technical problems existing in a traditional method are effectively solved, accurate estimation of the hydrodynamic force can be achieved in a complex swimming scene, real-time hydrodynamic force estimation can be achieved, and meanwhile the constructed model has physical interpretability.

Description

Technical Field

[0001] This application belongs to the field of hydrodynamic analysis technology, specifically relating to a method, electronic device and storage medium for hydrodynamic estimation based on pressure images. Background Technology

[0002] In numerous fields such as underwater vehicle development and biomimetic propulsion mechanism research, analyzing the mechanism of efficient cruising based on the water pressure distribution on the vehicle surface and accurately estimating hydrodynamic parameters (such as lift, drag, thrust, and torque) is crucial. This directly affects the performance, energy consumption, and safety of related equipment, and has significant academic value and far-reaching practical guiding significance for improving artificial propulsion technology. Hydrodynamic estimation aims to accurately obtain the forces acting on an object in a fluid, providing key basis for optimized design, performance prediction, and control strategy formulation.

[0003] Currently, traditional hydrodynamic estimation methods are mainly divided into three categories: theoretical modeling methods, numerical calculation methods, and data-driven methods.

[0004] Theoretical modeling methods are based on fluid mechanics theories, such as the Navier-Stokes equations and potential flow theory, to construct mathematical models and then calculate hydrodynamic parameters using analytical solutions or semi-empirical formulas. In practice, methods such as the element method and vortex lattice method are often used to discretize the surface of the object and solve for the flow field distribution by combining boundary conditions. However, this method has significant limitations. Because it requires assuming ideal fluids or neglecting viscous effects, it is difficult to accurately predict hydrodynamic conditions in complex turbulent or separated flow scenarios, thus greatly limiting its application scope.

[0005] Numerical simulation methods utilize computational fluid dynamics (CFD) techniques to discretize and numerically simulate flow fields, such as the finite volume method and finite element method, combined with turbulence models (e.g., k-ε, LES) to solve the Navier-Stokes equations. Commercial software like Fluent and OpenFOAM are typically used for mesh generation and iterative calculations. While this method provides relatively detailed and accurate flow field information, it demands extremely high computational resources. Achieving high-precision simulations often requires millions of meshes, with single calculations taking hours or even days, significantly limiting its rapid application and large-scale deployment in practical engineering.

[0006] Data-driven approaches rely on machine learning algorithms, such as neural networks and support vector machines, to construct mapping models between hydrodynamic forces and motion parameters (velocity, angle of attack, etc.), training them with historical data to achieve predictive capabilities. LSTM networks are commonly used when processing time-series data. However, this method is essentially a black-box model; while it can provide predictions, it lacks physical interpretability, making it difficult for engineers to deeply understand the physical mechanisms of hydrodynamic generation and hindering targeted improvements and optimizations to the model.

[0007] In summary, existing hydrodynamic estimation methods all have varying degrees of shortcomings and cannot meet the demands of practical engineering for efficient, accurate, and physically interpretable hydrodynamic estimation. Therefore, developing a new hydrodynamic estimation method is of significant practical importance. Summary of the Invention

[0008] This application aims to provide a method for hydrodynamic estimation based on pressure images. This method transforms pressure data into image form and utilizes convolutional neural networks for feature extraction to achieve accurate prediction of the hydrodynamic forces in complex tumbling swimming motions. This method effectively overcomes many technical challenges inherent in traditional methods, enabling both accurate and real-time estimation of hydrodynamic forces in complex swimming scenarios. Furthermore, the constructed model possesses physical interpretability.

[0009] To achieve the above technical objectives, this application specifically implements the following technical solution:

[0010] In one aspect of this application, a method for hydrodynamic estimation based on pressure images is provided, comprising the following steps:

[0011] S1: Acquire pressure distribution data and corresponding hydrodynamic data on the surface of the moving target;

[0012] S2: Map the pressure distribution data to a two-dimensional image matrix to generate a pressure image of the moving target;

[0013] S3: Use the hydrodynamic data as labels to associate with the corresponding pressure images to construct a dataset;

[0014] S4: Input the dataset into a pre-trained convolutional neural network model, extract the pressure features of the pressure image through the model, and output the hydrodynamic estimate.

[0015] In one implementation, step S1 includes: synchronously collecting pressure distribution data and corresponding hydrodynamic data on the surface of the moving target, using the moving target's motion cycle as the unit.

[0016] In one embodiment, the motion target includes a biomimetic propulsion device, and the hydrodynamic data includes lift, drag, thrust, or torque.

[0017] In one embodiment, the biomimetic thruster is selected from the structure of a sea lion flipper, the pressure distribution data is a pressure differential coefficient matrix, and the hydrodynamic data is a lift coefficient.

[0018] In one implementation, step S1 includes:

[0019] S11: Establish the motion trajectories of multiple moving targets, and calculate the motion parameters of each trajectory state point through inverse kinematics;

[0020] S12: Construct the target geometric model and obtain the hydrodynamic response in fluid experiments or simulation software based on the motion parameters;

[0021] S13: Extract the output lift coefficient and local surface element pressure difference coefficient, wherein the local surface element coincides with the spatial position of the preset pressure sensor.

[0022] In one implementation, step S2 includes:

[0023] S21: Normalize the pressure distribution data and map it to a first predefined numerical range;

[0024] S22: Construct a two-dimensional image matrix, map the shape of the moving target in the two-dimensional image matrix, and determine the influence area of ​​each pressure sensor in the two-dimensional image matrix according to the position of the pressure sensor; the pressure sensor is used to acquire the pressure distribution data;

[0025] S23: Fill the normalized pressure distribution data into all pixels within the corresponding sensor influence area, and fill the uncovered areas with zero values ​​to generate a pressure image.

[0026] In one implementation, step S3 includes:

[0027] S31: Normalize the acquired hydrodynamic data and map it to a second predefined numerical range;

[0028] S32: Use the normalized hydrodynamic data as labels and construct a dataset with a mapping relationship with the corresponding pressure image text.

[0029] In one implementation, the pre-trained convolutional neural network model includes:

[0030] Configure a convolutional neural network architecture and divide the dataset into a training set, a validation set, and a test set;

[0031] The model weight parameters are optimized using the training set, and the hyperparameters are adjusted based on the loss function values ​​of the validation set.

[0032] The model's prediction accuracy is evaluated using the test set, and the hyperparameters are dynamically and iteratively adjusted based on the loss function value and the prediction accuracy until the model output meets the preset convergence condition.

[0033] In another aspect of this application, an electronic device is provided, 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, it implements the steps of the method for hydrodynamic estimation based on pressure images as described above.

[0034] In another aspect of this application, a computer-readable storage medium is provided having an executable program stored thereon, which, when executed by a processor, implements the steps of the method for hydrodynamic estimation based on pressure images as described above.

[0035] The beneficial effects of this application are as follows:

[0036] Accurate and real-time hydrodynamic estimation: By visualizing pressure distribution data and extracting features using convolutional neural networks, this method can accurately estimate the hydrodynamic forces during complex tumbling movements, effectively solving the problem of traditional methods struggling to accurately estimate hydrodynamic forces during complex swimming processes. Furthermore, this method enables real-time hydrodynamic estimation, providing strong support for real-time monitoring and control in related fields. It delivers real-time hydrodynamic estimation results with high computational accuracy, significantly improving the timeliness and accuracy of hydrodynamic estimation.

[0037] Physically interpretable: This application borrows the principle of calculating forces using the pressure integral method, enabling the hydrodynamic estimation model based on convolutional neural networks to possess a certain degree of physical interpretability. This characteristic overcomes the limitation of unclear physical meaning in traditional methods, allowing the hydrodynamic estimation results to not only have numerical accuracy but also to be understood and analyzed from the perspective of physical principles, providing a theoretical basis for further model optimization and a deeper understanding of hydrodynamic phenomena.

[0038] This application opens up a new direction in hydrodynamic estimation: Compared to existing methods for estimating hydrodynamics based on motion parameters, this application takes a unique approach, estimating hydrodynamics from the perspective of pressure images, thus opening up a completely new research direction in the field of hydrodynamic estimation. This innovative method provides new ideas and technical means for solving complex hydrodynamic problems, and is expected to promote the application and development of hydrodynamic estimation technology in more fields, and promote the technological progress of related disciplines. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the hydrodynamic estimation method according to an embodiment of this application;

[0040] Figure 2 This is a schematic diagram of the pressure image and sensor distribution in an embodiment of this application. Detailed Implementation

[0041] The technical solution of this application will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art will understand that the embodiments described below are only some embodiments of this application, not all embodiments, and are only used to illustrate this application, and should not be regarded as limiting the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] In one embodiment, a method for hydrodynamic estimation based on pressure images is provided, including the following steps:

[0043] S1: Acquire pressure distribution data and corresponding hydrodynamic data on the surface of the moving target.

[0044] Specifically, pressure distribution data and corresponding hydrodynamic data on the surface of the moving target are collected synchronously, using the target's motion cycle as the unit. By acquiring surface pressure distribution data and corresponding hydrodynamic data within the complete motion cycle of the moving target, a foundational dataset is provided for subsequent pressure image modeling. Simultaneous acquisition of these two types of data ensures the spatiotemporal consistency of pressure characteristics and hydrodynamic response, resolving the prediction error problem caused by data separation in traditional methods.

[0045] The motion cycle is defined as a single complete motion cycle of the target (such as a single flapping motion of a bionic flipper), discretizing continuous motion into multiple periodic segments. For example, the flapping motion of a sea lion flipper consists of multiple flapping actions, and a complete cycle of one flapping action can be divided into "downward slapping stroke - horizontal slapping stroke - upward swing recovery".

[0046] Synchronous acquisition is achieved by aligning with hardware synchronization signals (such as external clock pulses) or software timestamps to ensure that the pressure sensor and the hydrodynamic measurement device start data recording at the same time, thus avoiding phase mismatch caused by time delay.

[0047] Bionic propulsion systems can be selected from flexible or rigid propulsion structures such as fish fins, sea lion flippers, and bionic hydrofoils, and their motion modes include oscillation, rotation, compound motion, or complex biological motion.

[0048] Hydrodynamic data can be selected from lift, drag, thrust, or torque. Specifically, lift is the force perpendicular to the wing surface of the biomimetic propeller (such as the upward force generated when the flippers flap); drag and thrust are forces parallel to the direction of motion (such as the obstruction of the flippers by water flow during forward movement, and the forward force exerted by the flippers on the vehicle during forward movement); torque is the torque about the axis of rotation (such as the yaw torque when the biomimetic tail fin swings). Pressure distribution data is the pressure difference between the upper and lower surfaces of various points on the moving target, acquired through a pressure sensor array to form a two-dimensional pressure field. Specifically, hydrodynamic data and pressure distribution data can be obtained through simulation software or real experiments.

[0049] In some embodiments, the biomimetic thruster is selected from the structure of a sea lion flipper, the pressure distribution data is a pressure differential coefficient matrix, and the hydrodynamic data is a lift coefficient.

[0050] Specifically, multiple motion trajectories of moving targets are established, and the motion parameters of each trajectory state point are calculated through inverse kinematics. Different actions are taken based on the method of obtaining the hydrodynamic response: when the hydrodynamic data is obtained from experiments, a physical model of the target should be made, and the hydrodynamic response should be obtained in a fluid experiment using pressure and force sensors based on the motion parameters; when the hydrodynamic data is obtained from simulation, a geometric model of the target is constructed, and the hydrodynamic response is solved in fluid simulation software based on the motion parameters; the output lift coefficient and local surface element pressure difference coefficient are extracted, wherein the local surface element coincides with the spatial position of the preset pressure sensor.

[0051] S2: The pressure distribution data is mapped to a two-dimensional image matrix to generate a pressure image of the moving target.

[0052] Specifically, the pressure distribution data is normalized and mapped to a first predefined numerical range; a two-dimensional image matrix is ​​constructed, in which the shape of the moving target is mapped, and the influence area of ​​each pressure sensor in the two-dimensional image matrix is ​​determined according to the position of the pressure sensor, wherein the pressure sensor is used to acquire the pressure distribution data; the normalized pressure distribution data is filled into all pixels within the influence area of ​​the corresponding sensor, and the uncovered areas are filled with zero values ​​to generate a pressure image.

[0053] Normalization can eliminate differences in measurement ranges between different sensors, unify data scales, and improve model training stability. It includes linear normalization and nonlinear normalization. For scenarios with large turbulent pressure fluctuations, logarithmic normalization or quantile normalization is used to preserve extreme pressure characteristics. In this embodiment, linear normalization is used to map the pressure distribution data to a first predefined numerical interval [0,1].

[0054] The principle for dividing the influence area of ​​a pressure sensor is that the influence area of ​​each sensor covers its actual measurement range (e.g., a square area with a side length of 2cm centered on the sensor).

[0055] This step transforms discrete pressure sensor data into structured two-dimensional images, providing spatially semantic input for CNN models and significantly improving the accuracy and interpretability of hydrodynamic estimation.

[0056] S3: Use the hydrodynamic data as labels to associate with the corresponding pressure images to construct a dataset.

[0057] Specifically, the acquired hydrodynamic data is normalized and mapped to a second predefined numerical range. The normalized hydrodynamic data is then used as labels to create a dataset that maps to the corresponding pressure images. By associating the pressure images with the corresponding hydrodynamic data, a structured dataset with mapping relationships is constructed, providing standardized input-label pairs for subsequent model training (such as deep learning).

[0058] The hydrodynamic data is normalized to integers within the range of 0 to 100 using linear normalization. If the data distribution is severely skewed, nonlinear normalization, such as logarithmic transformation or quantile normalization, can be used. The normalized hydrodynamic data (integers) are used as labels to establish a unique mapping relationship with the corresponding pressure images.

[0059] S4: Input the dataset into a pre-trained convolutional neural network model, extract the pressure features of the pressure image through the model, and output the hydrodynamic estimate.

[0060] The pre-trained convolutional neural network model includes: configuring a convolutional neural network architecture and dividing the dataset into a training set, a validation set, and a test set; optimizing the model weight parameters using the training set and adjusting the hyperparameters based on the loss function values ​​of the validation set; evaluating the model's prediction accuracy using the test set and dynamically iteratively adjusting the hyperparameters according to the loss function values ​​and prediction accuracy until the model output meets the preset convergence conditions.

[0061] The convolutional neural network architecture includes a pre-trained base model selected from at least one of Inception, ResNet, or EfficientNet; a feature extraction layer for extracting pressure features from the pressure image; and an output layer that uses a linear activation function to output a continuous hydrodynamic estimate.

[0062] Model training: Construct a convolutional neural network training model and divide the dataset into two groups, a and b. Group a is further divided into training dataset a1 and validation dataset a2, with a reference ratio of 8:1. Input the data into the model to generate the corresponding weights, bias terms, prediction accuracy, and loss function. Group b serves as the test dataset, and the data is input into the trained model to generate the model's prediction results.

[0063] Model hyperparameter tuning: Evaluate the model's performance and overfitting based on the prediction accuracy, loss function, and prediction results on the test dataset. If the training results meet expectations, the model construction is complete; if the training results do not meet expectations, adjust the hyperparameters set during model training and retrain the model.

[0064] In another aspect of this application, an electronic device is provided, including a memory for storing a computer program and a processor coupled to the memory for executing the computer program.

[0065] Specifically, the processor executes the computer program to perform the following steps: acquiring pressure distribution data and corresponding hydrodynamic data on the surface of the moving target; mapping the pressure distribution data to a two-dimensional image matrix to generate a pressure image of the moving target; associating the pressure image with the corresponding hydrodynamic data as a label to construct a dataset; inputting the dataset into a pre-trained convolutional neural network model, extracting the pressure features of the pressure image through the model, and outputting hydrodynamic estimates.

[0066] The pre-trained convolutional neural network model is trained in the following manner:

[0067] Configure the convolutional neural network architecture and divide the dataset into training, validation, and test sets;

[0068] The model weight parameters are optimized using the training set, and the hyperparameters are adjusted based on the loss function values ​​of the validation set.

[0069] The model's prediction accuracy is evaluated using the test set, and the hyperparameters are dynamically and iteratively adjusted based on the loss function value and the prediction accuracy until the model output meets the preset convergence condition.

[0070] When the processor executes the computer program, it also performs the following steps: normalizing the hydrodynamic data and mapping it to integers within a predefined numerical range [0, 100].

[0071] The convolutional neural network architecture includes: a pre-trained base model selected from at least one of Inception, ResNet, or EfficientNet; a feature extraction layer for extracting pressure features from the pressure image; and an output layer that uses a linear activation function to output continuous hydrodynamic estimates.

[0072] In another aspect of this application, a computer-readable storage medium is provided having an executable program stored thereon, which, when executed by a processor, implements the steps of the method for hydrodynamic estimation based on pressure images as described above.

[0073] Example

[0074] This embodiment uses the flipper of a sea lion as an example to illustrate the method of hydrodynamic estimation based on pressure images, referring to... Figure 1 As shown, pressure distribution data and corresponding hydrodynamic data of the surface of a moving target are acquired. The pressure distribution data is mapped to a two-dimensional image matrix to generate a pressure image of the moving target. The hydrodynamic data is used as labels to associate the corresponding pressure images, constructing a dataset. A convolutional neural network model is trained using the dataset. The trained convolutional neural network model is used to extract pressure features from the pressure images of the moving target and output hydrodynamic estimates. The specific steps are as follows:

[0075] A. Simulation data acquisition

[0076] A1. Use MATLAB to generate multiple flapping motion trajectories of sea lion flippers and obtain the coordinate information corresponding to each state point on the trajectory. Calculate the pitch angle, sweep angle and dihedral angle during the flipper movement through inverse kinematics.

[0077] A2. Establish a sea lion flipper model, obtain cross-sectional data of the sea lion flipper, restore the flipper shape in SOLIDWORKS, and input the shape data, motion parameters and flow field information into AVL (Athena Vortex Lattice) to obtain hydrodynamic data. The pressure difference coefficient is stored in .txt format and the lift coefficient is stored in .mat format matrix.

[0078] B. Data Restructuring

[0079] B1. Extract the differential pressure data and normalize it to the range of 0 to 1; extract the lift coefficient and normalize it to an integer range of 0 to 100. Convert both types of data into .mat matrices for easy subsequent use.

[0080] B2, Reference Figure 2As shown, the target shape is depicted in the .CSV table. The influence range of the pressure data is determined based on the distribution of the pressure sensors. The red dots in the figure represent the actual positions of the sensors on the target surface, and the color blocks of the same color within the area of ​​the red dots represent the assumed influence range of the sensor.

[0081] B3. Create a pressure data file containing pressure differential coefficients and conforming to the target shape. Fill the normalized pressure differential coefficients into the pixels within the corresponding influence range, ensuring that the pressure differential coefficient values ​​are the same in each influence range. Fill the remaining ranges without pressure differential coefficients with 0, and then save it as a .CSV table.

[0082] B4. The lift coefficient is calculated and converted into lift, then the lift is normalized and used as the label for the corresponding file. For example, if a pressure differential coefficient table corresponds to trajectory 1, state 1, and normalized lift 30, the file will be named "1-1-30.CSV", where 30 is the label value. Rename all files and construct the dataset.

[0083] C. Model Training

[0084] C1. Construct a convolutional neural network training model, adjust the parameters of the Dense layer, and add a Softmax layer after the Dense layer to make the output result conform to the probability distribution. Refer to the existing convolutional neural network structure to construct the model. Since the lift coefficient is normalized and adjusted to an integer in the range of 0 to 100, the parameters of the last Dense layer are adjusted to 101, and Softmax is used to activate this layer.

[0085] C2. Divide the dataset obtained in step B into two groups, a and b, in a 3:1 ratio, with group b serving as the test dataset. Further divide group a into a training dataset and a validation dataset in an 8:1 ratio.

[0086] C3. Initialize the model and input the data into the model to generate the corresponding weights and biases. Calculate and plot the model's prediction accuracy, loss function, and result prediction.

[0087] Adjust the filters, kernel size, padding, activation, and density of each layer, and initialize the model's batch size, epochs, and loss function. After inputting the dataset into the model, save the weights and biases to the checkpoint for later optimization. After training, output the prediction accuracy and loss function for each epoch. Based on this output data, plot line graphs of prediction accuracy and loss function, and retain the output of the last epoch as the model's training result. Since the output results conform to a probability distribution, weighted summation and inverse normalization are used to solve for the lift corresponding to the input pressure coefficient table, and this result is plotted together with the inversely normalized labels to create a prediction graph. Furthermore, variance and maximum absolute error are used to evaluate the prediction performance.

[0088] D. Model parameter tuning

[0089] When the prediction results do not meet expectations, return to the model and adjust the filters, kernel size, padding, activation, density, as well as the optimizer, learning rate, loss function, batch size, epochs=15, validation_freq and other hyperparameters. Among them, the optimizer is a more important part. When the prediction results are consistently not up to expectations, you can try changing the optimizer model.

[0090] E. Use the trained convolutional neural network model to extract the pressure features of the pressure image of the moving target and output the hydrodynamic estimate.

[0091] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and these are all within the scope of protection of this application.

Claims

1. A method for hydrodynamic estimation based on pressure images, characterized in that, include: S1: Acquire pressure distribution data and corresponding hydrodynamic data on the surface of the moving target; S2: Map the pressure distribution data to a two-dimensional image matrix to generate a pressure image of the moving target; S3: Use the hydrodynamic data as labels to associate with the corresponding pressure images to construct a dataset; S4: Input the dataset into a pre-trained convolutional neural network model, extract the pressure features of the pressure image through the model, and output the hydrodynamic estimate.

2. The hydrodynamic estimation method based on pressure images according to claim 1, characterized in that, Step S1 includes: synchronously collecting pressure distribution data and corresponding hydrodynamic data on the surface of the moving target, using the moving target's motion cycle as the unit.

3. The hydrodynamic estimation method based on pressure images according to claim 1, characterized in that, The motion target includes a biomimetic propulsion device, and the hydrodynamic data includes lift, drag, thrust, or torque.

4. The hydrodynamic estimation method based on pressure images according to claim 1, characterized in that, The biomimetic thruster is selected from the structure of a sea lion flipper, the pressure distribution data is a pressure difference coefficient matrix, and the hydrodynamic data is a lift coefficient.

5. The hydrodynamic estimation method based on pressure images according to claim 4, characterized in that, Step S1 includes: S11: Establish the motion trajectories of multiple moving targets, and calculate the motion parameters of each trajectory state point through inverse kinematics; S12: Construct the target geometric model and obtain the hydrodynamic response in fluid experiments or simulation software based on the motion parameters; S13: Extract the output lift coefficient and local surface element pressure difference coefficient, wherein the local surface element coincides with the spatial position of the preset pressure sensor.

6. The hydrodynamic estimation method based on pressure images according to claim 1, characterized in that, Step S2 includes: S21: Normalize the pressure distribution data and map it to a first predefined numerical range; S22: Construct a two-dimensional image matrix, map the shape of the moving target in the two-dimensional image matrix, and delineate the influence area of ​​each pressure sensor in the two-dimensional image matrix according to the position of the pressure sensor; the pressure sensor is used to acquire the pressure distribution data; S23: Fill the normalized pressure distribution data into all pixels within the corresponding sensor influence area, and fill the uncovered areas with zero values ​​to generate a pressure image.

7. The hydrodynamic estimation method based on pressure images according to claim 1, characterized in that, Step S3 includes: S31: Normalize the acquired hydrodynamic data and map it to a second predefined numerical range; S32: Use the normalized hydrodynamic data as labels to construct a dataset with a mapping relationship with the corresponding pressure image data.

8. The hydrodynamic estimation method based on pressure images according to claim 1, characterized in that, Pre-trained convolutional neural network models include: Configure a convolutional neural network architecture and divide the dataset into a training set, a validation set, and a test set; The model weight parameters are optimized using the training set, and the hyperparameters are adjusted based on the loss function values ​​of the validation set. The model's prediction accuracy is evaluated using the test set, and the hyperparameters are dynamically and iteratively adjusted based on the loss function value and the prediction accuracy until the model output meets the preset convergence condition.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the hydrodynamic estimation method based on pressure images according to any one of claims 1 to 8.

10. A computer-readable storage medium having an executable program stored thereon, which, when executed by a processor, implements the steps of the hydrodynamic estimation method based on pressure images according to any one of claims 1 to 8.

Images