A Method for Detecting Transition Points for Flow Field Analysis of Underwater Vehicles

Abstract

An embodiment of the present invention discloses a transition point detection method for flow field analysis of an underwater vehicle, which relates to the field of flow field feature detection, can expand the application scope, be applicable to more application scenarios, and solves the defect of single-point estimation of the transition point. The present invention includes: receiving the initial flow field information sent by the client and establishing a flow field model according to the initial flow field information; establishing a training data set and a test data set; performing encoding processing on the training data set and the test data set, and the obtained encoding results include: heat maps corresponding to the training data set and the test data respectively; establishing a convolutional neural network prediction model, training the deep convolutional neural network prediction model through the training data set, and updating the network weights of the convolutional neural network prediction model; in the test stage, inputting the test data into the trained convolutional neural network prediction model, and then decoding the output of the convolutional neural network prediction model to obtain the final transition point coordinates and returning them to the client.

Description

Technical Field

[0001] The present invention relates to the field of flow field feature detection, and particularly to a method for detecting transition points for flow field analysis of an underwater vehicle. Background Art

[0002] When an underwater submarine moves forward, a fluid structure of separation bubbles will be generated on its surface, and the transition point is a key point in the separation bubbles. The position of the transition point is related to the Reynolds number, surface curvature, turbulence intensity, angle of attack, etc. It is difficult for traditional models to cover all influencing factors. Analyzing and determining the position of this point is of great significance for airfoil design and submarine design.

[0003] The modeling methods of the transition point analysis model include empirical methods based on linear stability theory (LST), quadratic stability equation (PSE), large eddy simulation (LES), and direct numerical simulation (DNS). The method based on linear stability theory (LST) evaluates the local amplification rate of unstable waves and assumes that turbulence occurs after the unstable waves reach a sufficient large amplitude threshold. Since the degree of nonlinear instability is short, the linear instability region occupies most of the transition process. Therefore, the method based on LST has been widely studied.

[0004] However, since in most actual cases, the start of the transition is determined by empirical values, the model can only work in specific scenarios (specific working conditions and airfoils, etc.), and the obtained transition point is a single-point estimate, lacking the corresponding confidence level.

[0005] Therefore, how to optimize the method for detecting transition points during the process of flow field analysis, so as to be applicable to more scenarios and working conditions, rather than being limited to certain specific scenario conditions, has become a problem that needs to be studied and solved. Summary of the Invention

[0006] An embodiment of the present invention provides a method for detecting transition points for flow field analysis of an underwater vehicle, which can expand the application scope, be applicable to more application scenarios, and solve the defect of single-point estimation of the transition point.

[0007] To achieve the above object, the embodiment of the present invention adopts the following technical solutions:

[0008] Step 1, receiving the initial flow field information sent by the client, and establishing a flow field model according to the initial flow field information;

[0009] Step 2, establishing a training data set and a test data set;

[0010] Step 3, performing encoding processing on the training data set and the test data set, and the obtained encoding results include: heat maps corresponding to the training data set and the test data respectively.

[0011] Among them, the heat map is used as the true label of the data;

[0012] Step 4: Establish a convolutional neural network prediction model, and train the convolutional neural network prediction model through the training data set and the heat map corresponding to the training data set to update the network weights of the convolutional neural network prediction model;

[0013] Step 5: In the test phase, extract test data from the test data set and input it into the trained convolutional neural network prediction model. Then, decode the output of the convolutional neural network prediction model to obtain the final transition point coordinates and return them to the client.

[0014] The transition point detection method for underwater vehicle flow field analysis provided by the embodiments of the present invention enables each technician to use the information provided by the client to calculate the main body running on the server side. Establish a flow field model and calculate flow field data according to the flow field information set by the client; construct training data according to the streamline information; perform reasonable encoding on the data for training the convolutional neural network model; establish a convolutional neural network prediction model, train the deep convolutional neural network prediction model, and update the network weights; use the model to calculate and decode the data to be predicted, and finally return the obtained result to the client.

[0015] Compared with the solutions in the prior art, this embodiment proposes a transition point detection method based on a convolutional neural network, improves the data encoding, and designs a brand-new transition point representation method with confidence. Compared with the prior art, the main advantages of the present invention are as follows: fewer prior assumptions, such as some empirical thresholds, can be used, and it is applicable to more scenarios (working conditions), rather than being limited to specific scenario conditions. Moreover, the transition point representation method with confidence can give richer information compared with the traditional single-point estimation. In practical applications, it not only takes into account the allocation of computing power but also improves the accuracy and efficiency of flow field calculation, thus being able to take into account both the simulation calculation efficiency and accuracy in the process of flow field analysis and improve the efficiency of the entire flow field analysis system. Brief Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a schematic diagram of the method flow provided by the embodiments of the present invention;

[0018] Figure 2It is the overall flowchart of the transition point detection model based on the convolutional neural network in the specific example provided by the embodiment of the present invention;

[0019] Figure 3 It is the structural schematic diagram of the dataset production and encoding in the specific example provided by the embodiment of the present invention;

[0020] Figure 4 It is the schematic diagram of different encoding comparisons in the specific example provided by the embodiment of the present invention;

[0021] Figure 5 It is the schematic diagram of encoding conversion in the specific example provided by the embodiment of the present invention. Detailed implementation manners

[0022] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners. The implementation manners of the present invention will be described in detail below, and examples of the implementation manners are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The implementation manners described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention and cannot be construed as a limitation of the present invention. Those skilled in the art of the present technology can understand that, unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention means that there are the described features, integers, steps, operations, elements and / or components, but does not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The phrase "and / or" used herein includes any unit and all combinations of one or more related listed items. Those skilled in the art of the present technology can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which the present invention belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless defined as here.

[0023] The embodiment of the present invention provides a transition point detection method for the flow field analysis of an underwater vehicle, as Figure 1 shown, including:

[0024] Step 1: Receive the initial flow field information sent by the client, and establish a flow field model according to the initial flow field information.

[0025] Among them, the client sends the initial flow field information to the server to establish a flow field model for calculating the experimental data required for the experiment. In the implementation process of this embodiment, a certain amount of computing resources are required to train and test the model, and the flow field data in the database (such as the initial flow field information, the data used to establish the training dataset and the test dataset, etc.) needs to be called. Therefore, in practical applications, it is designed in the form of front-end and back-end interaction between the client and the server to facilitate deployment in scenarios such as scientific research institutions and the R & D departments of technology enterprises.

[0026] Step 2: Establish a training dataset and a test dataset.

[0027] Among them, according to the data obtained in Step 1, supplemented by streamlines, a training dataset and a test dataset are made.

[0028] Step 3: Perform encoding processing on the training dataset and the test dataset. The obtained encoding results include: the heat maps corresponding to the training dataset and the test data respectively.

[0029] Among them, the training dataset and the test dataset are encoded into heat maps. The heat map is used to label the true labels of the data so that the labeled data can participate in subsequent model training.

[0030] Step 4: Establish a convolutional neural network prediction model, and train the convolutional neural network prediction model through the training dataset to update the network weights of the convolutional neural network prediction model.

[0031] Step 5: In the test stage, input the test data into the trained convolutional neural network prediction model. Then, decode the output of the convolutional neural network prediction model to obtain the final transition point coordinates and return them to the client.

[0032] Specifically, before Step 1, it also includes: The server receives the login information and task information sent by the client. The task information includes the initial flow field information. Determine the maximum amount of computing resources available to the client according to the login information, and detect whether the computing resources corresponding to the task information exceed the maximum amount of computing resources. If not, extract computing resources from the resource pool and allocate them to the client.

[0033] Among them, the user account login operation can be performed by the client, and the login information is sent to the server. The server extracts computing resources from the resource pool according to the login information and allocates them to the account. Alternatively, after receiving the initial information of the flow field sent by the client, before performing the initialization process, the server estimates the required computing power according to the initial information, and extracts computing resources from the resource pool according to the estimation result and allocates them to the account.

[0034] In this embodiment, a transition point detection method based on a convolutional neural network as shown in Figure 2 is designed, in which the production of the dataset of the transition point, the model calculation prediction and the output decoding process can be completed. The experimental data in this embodiment is obtained by solving the flow field model, specifically referring to the incompressible N-S equation, and the specific explanation is as follows:

[0035] In step 1, a flow field model is established according to the initial information of the flow field, including:

[0036] Establish an incompressible Navier-Stokes flow field model:

[0037]

[0038]

[0039] where p represents pressure, u represents the velocity vector, F c (p, u) and represent the inviscid and viscous fluxes respectively, represents the gradient calculation, t represents time, represents the partial derivative of velocity with respect to time;

[0040] Specifically, in this embodiment, the high-order Discontinuous Galerkin (DG) algorithm can be substituted into the above formula to obtain the discretized version as follows:

[0041]

[0042]

[0043] where represents the basis function, N represents the number of basis functions, and are obtained by solving the following Riemann problem equation:

[0044]

[0045]

[0046]

[0047] Combined with the initial conditions:

[0048]

[0049] and

[0050]

[0051] Finally, the following discrete equations can be obtained:

[0052]

[0053] where W represents the overall vector of unknown degrees of freedom, and Mis represents the global block diagonal matrix. The above time-varying system can be solved by the following implicit relaxation conditions:

[0054] W (0) = W (n)

[0055]

[0056]

[0057] After obtaining the above data, streamline diagrams can be plotted using the velocity components, and the streamlines are used as the basis for judgment.

[0058] Furthermore, in this embodiment, the dataset annotation method as shown in Figure 3 can be adopted. Generally speaking: the separated bubble region is cropped from the original image with a bounding box. The independent closed streamline is the separated bubble, and the center enclosed by the streamline in the separated bubble is the location of the transition point. The separated bubble is marked with a square box, and the transition point in the separated bubble is marked with a dot. Then, the size of the input data of the convolutional neural network is unified. Then, the transition point is converted from the form of two-dimensional coordinates (x1, x2) to the form of a heat map.

[0059] Specifically applied in this embodiment, in step 2, it includes:

[0060] Step 2.1: Generate a streamline diagram using the established flow field model. Among them, the streamline diagram can be plotted according to the data obtained by solving the N-S equation.

[0061] Step 2.2: Search for the location of the transition point in the streamline diagram. Among them, the streamline is used as the judgment basis, the independent closed streamline is regarded as the separated bubble, and the center enclosed by the streamline in the separated bubble is the location of the transition point.

[0062] Step 2.3: Mark the separated bubble with a square box and mark the transition point in the separated bubble with a dot.

[0063] The input data size of the convolutional neural network needs to be unified. In this embodiment, the size of the input data is 768×192, and the average value of all data is obtained. The label in the form of a heat map adopted in this embodiment can make the model training easier. Specifically, in step 3, the encoding process includes:

[0064] Converting the labeled transition points from the form of two-dimensional coordinates to the form of a heat map, the process of converting the labeled transition points from the form of two-dimensional coordinates (x1, x2) to the form of a heat map.

[0065] Among them, a point on the heat map corresponds to the probability of a transition point appearing at that point, and the range of the probability is [0, 1]. The way to obtain the probability of each point on the heat map is:

[0066] μ represents the coordinates of the transition point, x represents the coordinates of the point on the heat map, and Σ represents the covariance matrix of the Gaussian distribution.

[0067] By observing the characteristics of the experimental data, it can be found that by changing the Σ matrix, the encoded heat map can be made more in line with the distribution law of the data. For example: by observing the characteristics of the experimental data, it is found that the transition points are distributed on the surface of the submarine, which is a "slender" distribution. The encoded heat map can be made more in line with the distribution law of the data by changing the Σ matrix. This application embodiment adopts an anisotropic Gaussian distribution to fit the distribution of the transition points. A comparison schematic diagram of the isotropic Gaussian distribution heat map and the anisotropic Gaussian distribution heat map is as Figure 4 shown. Performing a rotation transformation and a scale transformation on the isotropic Gaussian distribution can make the heat map distribution "fit" on the submarine surface, that is, there is a higher probability value along the submarine surface. The specific encoding conversion schematic diagram is as Figure 5 shown, the specific rotation transformation and scale transformation of the Σ matrix. Specifically, before converting from the form of two-dimensional coordinates to the form of a heat map, a rotation and scale transformation are performed on Σ, where:

[0068]

[0069]

[0070] R represents the rotation matrix, θ represents the rotation angle, S represents the scale transformation matrix, k represents the aspect ratio of the separation bubble, the transformed covariance matrix is Σ′, and Σ′ = TΣT T ; T = RS, T represents the transformation matrix obtained by multiplying the rotation matrix R and the scale matrix S, and T T represents the transpose of the transformation matrix.

[0071] After updating the network weights of the convolutional neural network prediction model using the stochastic gradient descent algorithm to obtain the training data and the corresponding encoded labels, a convolutional neural network prediction model is established, and the deep convolutional neural network prediction model is trained using the training data set, and the network weights are updated using the stochastic gradient descent algorithm.

[0072] In this embodiment, in step 4, it includes:

[0073] Step 4.1: After establishing the convolutional neural network prediction model, randomly initialize the initial weights of the neural network.

[0074] Step 4.2: Input the training data set into the convolutional neural network prediction model to output a predicted heatmap.

[0075] Step 4.3: Calculate the root mean square error RMSE between the predicted heatmap and the true heatmap, where,

[0076]

[0077] where, g i represents the probability value of the true label of the i-th point on the true heatmap, represents the predicted probability value of the corresponding point on the predicted heatmap, and N represents the total number of points on the predicted heatmap.

[0078] Step 4.4: Update the network weights of the convolutional neural network prediction model using the backpropagation algorithm according to the obtained mean square loss.

[0079] In this embodiment, in step 5, it includes:

[0080] Step 5.1: Input the test data into the trained convolutional neural network prediction model to output a test heatmap.

[0081] Step 5.2: Decode the test heatmap, convert it from the form of a heatmap to the form of two-dimensional coordinates (x1, x2), and obtain the two-dimensional coordinates.

[0082] Step 5.3: Use the obtained two-dimensional coordinates as the final transition point coordinates. Among them, the evaluation index can also be calculated according to the obtained two-dimensional coordinates to obtain the overall performance of the model.

[0083] Further, in the decoding of the test heatmap, it includes: offsetting at the position of the point with the maximum probability value on the test heatmap, where:

[0084] Decoding the test heatmap, this formula can be understood as making an offset on the basis of point m

[0085] Among them, l represents the decoding position, and m represents the position of the maximum probability value on the test heatmap. represents the heatmap output by the convolutional neural network prediction model. and represent the first-order derivative and the second-order derivative of the output heatmap at point m, respectively.

[0086] In the actual prediction stage, the heatmap output by the model is decoded or not decoded as needed, corresponding to single-point prediction and probability heatmap prediction with confidence, respectively. Here, the confidence is the probability value of the transition point.

[0087] The experimental results of this method on different coding methods and different backbone networks are shown in Table 1. Among them, Table 1 shows the experimental results on different coding methods and different backbone networks, AP, AP 50 , AP 75 , AR are indicators to evaluate the model performance. The larger the value, the better the performance. R50 and R101 represent the SimpleBaseline backbone networks of different model sizes in this paper, and HRNet-W32 and HRNet-w48 represent the HRNet backbone networks of different model sizes.

[0088] Table 1

[0089]

[0090] Observing the experimental results in Table 1, it can be found that the anisotropic Gaussian distribution coding method proposed in this paper, which is automatically adjusted according to the separation bubble size, can significantly improve the prediction accuracy and recall rate on different backbone networks, proving the effectiveness of the coding method.

[0091] Based on the above results, this method has the following four advantages compared with the traditional empirical model method in the transition point detection problem. First, it uses fewer prior assumptions, such as some empirical thresholds. Second, it is applicable to more scenarios (working conditions) and is not limited to specific scenario conditions. Third, it encodes with an anisotropic Gaussian distribution, which is more in line with the distribution of the transition point itself and also improves the final prediction performance. Fourth, it adopts a representation method of the transition point with confidence, which can give more abundant information compared with the traditional single-point estimation. Thus, this method has certain application value and prospects.

[0092] This embodiment relates to the field of extraction and detection of flow field characteristic structures, which can improve the generalization performance of the entire flow field transition point detection system and provide richer detection information. It runs on the server side using the information provided by the client. Among them, according to the flow field information set by the client, a flow field model is established and flow field data is calculated. Training data is constructed based on streamline information. The data is reasonably encoded for training a convolutional neural network model. A convolutional neural network prediction model is established, and the deep convolutional neural network prediction model is trained to update the network weights. The model is used to calculate and decode the data to be predicted, and the final result is returned to the client.

[0093] Compared with the solutions in the prior art, this embodiment proposes a transition point detection method based on a convolutional neural network and improves the data encoding. In addition, a brand-new transition point representation method with confidence is proposed. Compared with the prior art, the main advantages of the present invention are as follows: 1. Fewer prior assumptions, such as some empirical thresholds, can be used. 2. It is applicable to more scenarios (working conditions) and is not limited to specific scenario conditions. 3. Encoding with an anisotropic Gaussian distribution is more in line with the distribution of the transition point itself and also improves the final prediction accuracy. Using the transition point representation method with confidence can provide richer information compared with the traditional single-point estimation.

[0094] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the description of the method embodiment. The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A transition point detection method for flow field analysis of an underwater vehicle, characterized in that, Including: Step 1: Receive the initial flow field information sent by the client and establish a flow field model according to the initial flow field information; Step 2: Establish a training data set and a test data set; Step 3: Perform encoding processing on the training data set and the test data set, and the obtained encoding results include: heat maps corresponding to the training data set and the test data respectively; Step 4: Establish a convolutional neural network prediction model, and train the convolutional neural network prediction model through the training data set and the heat map corresponding to the training data set, and update the network weights of the convolutional neural network prediction model; Step 5: In the test stage, extract test data from the test data set and input it into the trained convolutional neural network prediction model. Then, decode the output of the convolutional neural network prediction model to obtain the final transition point coordinates and return them to the client; In Step 2, it includes: Step 2.1: Use the established flow field model to generate a streamline diagram; Step 2.2: In the streamline diagram, search for the position of the transition point. Among them, the streamline is used as the judgment basis, and the independent closed streamline is regarded as the separation bubble, and the center surrounded by the streamline in the separation bubble is the position of the transition point; Step 2.3: Mark the separation bubble with a box and mark the transition point in the separation bubble with a dot; In Step 3, the encoding processing performed includes: Convert the marked transition point from the form of two-dimensional coordinates to the form of a heat map. Among them, a point on the heat map corresponds to the probability of a transition point appearing at that point, and the range of the probability is [0,1]. The way to obtain the probability of each point on the heat map is: where μ represents the coordinates of the transition point, x represents the coordinates of the point on the heat map, and Σ represents the covariance matrix of the Gaussian distribution; It also includes: Before converting from the form of two-dimensional coordinates to the form of a heat map, perform rotation and scale transformation on Σ, where: R represents the rotation matrix, θ represents the rotation angle, S represents the scale matrix, k represents the aspect ratio of the separation bubble, and the transformed covariance matrix is Σ ′ , and Σ ′ = TΣT T ; T = RS, T represents the transformation matrix obtained by multiplying the rotation matrix R and the scale matrix S, and T T represents the transpose of the transformation matrix; Use the stochastic gradient descent algorithm to update the network weights of the convolutional neural network prediction model; In the decoding of the test heat map, it includes: Offset the position of the point with the maximum probability value on the test heat map, where: Decode the test heat map, where l represents the decoded position and m represents the position with the maximum probability value on the test heat map. represents the heat map output by the convolutional neural network prediction model. and respectively represent the first-order derivative and the second-order derivative of the output heat map at point m.

2. The method according to claim 1, wherein Before Step 1, it also includes: The server receives the login information and task information sent by the client, and the task information includes the initial flow field information; Determine the maximum amount of computing resources available to the client according to the login information, and detect whether the computing resources corresponding to the task information exceed the maximum amount of computing resources. If not, extract computing resources from the resource pool and allocate them to the client.

3. The method according to claim 1, wherein In Step 1, establishing a flow field model according to the initial flow field information includes: Establish an incompressible Navier-Stokes flow field model: where p represents pressure, u represents the velocity vector, F c (p, u) and represent the inviscid and viscous fluxes respectively, denotes the gradient calculation, t represents time, denotes the partial derivative of velocity with respect to time.

4. The method according to claim 1, characterized in that, In Step 4, it includes: Step 4.1: After establishing the convolutional neural network prediction model, randomly initialize the initial weights of the neural network; Step 4.2: Input the training data set into the convolutional neural network prediction model and output a predicted heat map; Step 4.3: Calculate the root mean square error RMSE between the predicted heat map and the true heat map, where, where g i represents the probability value of the true label of the i-th point on the true heatmap, represents the predicted probability value of the corresponding point on the predicted heatmap, and N represents the total number of points on the predicted heatmap; Step 4.4: Update the network weights of the convolutional neural network prediction model using the backpropagation algorithm based on the obtained mean squared loss.

5. The method according to claim 4, wherein In step 5, it includes: Step 5.1: Input the test data into the trained convolutional neural network prediction model to output the test heat map. Step 5.2: Decode the test heat map, convert it from the form of a heat map to the form of two-dimensional coordinates, and obtain the two-dimensional coordinates. Step 5.3: Use the obtained two-dimensional coordinates as the final transition point coordinates.

Images