Sound explosion prediction method and device based on deep learning, equipment and program product

Abstract

The embodiment of the invention provides a sound explosion prediction method and device based on deep learning, equipment and a program product. The method comprises the steps of obtaining to-be-predicted target data, wherein the target data comprises three-dimensional point cloud data representing physical attribute information of an aircraft profile and a surrounding flow field; the target data is converted into graph structure data, the graph structure data comprises node features and edge features, the node features comprise position information and physical attributes, and the edge features comprise topological relations among nodes; inputting the graph structure data into a sound explosion prediction model based on a graph neural network model to obtain a sound explosion prediction result; the acoustic explosion prediction model is used for learning a fluid dynamic relationship in the graph structure data through multiple rounds of message passing so as to output acoustic explosion physical characteristics. According to the method, the structured modeling capability of the graph neural network is utilized to replace traditional fluid mechanics simulation, sound explosion prediction is realized, and the prediction efficiency is greatly improved while the prediction accuracy is guaranteed.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and program product for predicting sonic booms based on deep learning. Background Technology

[0002] Supersonic aircraft generate powerful sonic booms when breaking the sound barrier, characterized by dramatic pressure changes in the surrounding airflow and the formation of shock waves similar to an explosion. These sonic booms can pose potential hazards to ground personnel, buildings, and the environment, thus hindering the development of supersonic flight. Therefore, predicting the distribution of sonic booms generated by aircraft and optimizing aircraft design to mitigate their impact has become a crucial research issue in the development of supersonic aircraft.

[0003] Currently, sonic boom prediction relies on computational fluid dynamics simulations, which theoretically simulate the generation and propagation process of sonic booms to predict them. However, this method is computationally very expensive, with a single simulation potentially taking hours to days, resulting in low prediction efficiency. Summary of the Invention

[0004] The deep learning-based sonic boom prediction method, device, equipment, and program products provided in this application utilize the structured modeling capabilities of graph neural networks and incorporate a multi-round message passing mechanism to accurately capture the fluid dynamics relationships between nodes. This replaces traditional computational fluid dynamics simulation, enabling sonic boom prediction and solving the problems of low computational efficiency and difficult modeling in sonic boom prediction. While ensuring prediction accuracy, it can significantly improve prediction efficiency.

[0005] In a first aspect, embodiments of this application provide a deep learning-based method for predicting sonic booms, comprising: acquiring target data to be predicted, the target data including three-dimensional point cloud data representing the physical properties of the aircraft shape and the surrounding flow field; converting the target data into graph structure data, the graph structure data including node features and edge features, the node features including location information and physical properties, and the edge features including the topological relationships between nodes; inputting the graph structure data obtained from the conversion of the target data into a sonic boom prediction model based on a graph neural network model to obtain a sonic boom prediction result; the sonic boom prediction model is used to learn the fluid dynamic relationships in the graph structure data through multi-round message passing to output the physical features of the sonic boom.

[0006] In one possible implementation, the target data is converted into graph structure data, specifically including: determining an initial sampling area by excluding areas in the target data that are not disturbed by sonic booms and areas that are affected by sonic booms to a less than preset level; performing downsampling within the initial sampling area; generating a mesh graph structure based on a triangulation algorithm according to the downsampled target data, deleting edges with lengths exceeding a preset threshold and removing isolated nodes and isolated subgraphs, and taking the largest connected subgraph to obtain the graph structure data corresponding to the target data.

[0007] In one possible implementation, downsampling is performed within the initial sampling area, specifically including: increasing the ratio of key region nodes to the total number of nodes by adjusting parameter settings to focus sampling on key regions; limiting the size of the point cloud after downsampling, and preserving the original sonic boom curve features through interpolation processing.

[0008] In one possible implementation, the sonic boom prediction model includes: an encoder for normalizing node features and edge features, and converting the node and edge features into implicit features of the nodes and edges; the node features include location information, pressure, density, velocity, and node type, and the edge features include direction vectors and magnitudes; a processor for learning the hydrodynamic relationships between nodes through multi-round message passing; and a decoder for converting the implicit node features into predicted physical features.

[0009] In one possible implementation, the method further includes: acquiring a sonic boom dataset, which includes three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields; constructing a sonic boom prediction model; training the constructed sonic boom prediction model based on the sonic boom dataset; evaluating and optimizing the sonic boom prediction model until the model performance meets preset performance requirements, thereby obtaining a trained sonic boom prediction model.

[0010] In one possible implementation, acquiring a sonic boom dataset includes: modifying the shape parameters of a standard aircraft model to obtain different aircraft shapes; calculating the sonic boom results for each aircraft shape based on fluid simulation technology, and generating three-dimensional point cloud data characterizing the physical properties of different aircraft shapes and the surrounding flow field.

[0011] In one possible implementation, evaluating the sonic boom prediction model specifically includes: setting a peak mean square error (RMSE) evaluation index; and evaluating the model performance of the sonic boom prediction model based on the RMSE and global mean square errors of the sonic boom curves output by the sonic boom prediction model and the actual sonic boom curves, combined with the RMSE evaluation index.

[0012] Secondly, embodiments of this application provide an apparatus, comprising: a data acquisition module for acquiring target data to be predicted, the target data including three-dimensional point cloud data characterizing the physical properties of the aircraft shape and surrounding flow field; a data processing module for converting the target data into graph structure data, the graph structure data including node features and edge features, the node features including location information and physical properties, and the edge features including the topological relationships between nodes; and a sonic boom prediction module for inputting the graph structure data obtained from the conversion of the target data into a sonic boom prediction model based on a graph neural network model to obtain a sonic boom prediction result; the sonic boom prediction model is used to learn the fluid dynamic relationships in the graph structure data through multi-round message passing to output the physical characteristics of the sonic boom.

[0013] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method provided in the first aspect above.

[0014] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method provided in the first aspect above.

[0015] Fifthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method provided in the first aspect above.

[0016] The deep learning-based sonic boom prediction method, apparatus, device, and program products provided in this application acquire three-dimensional point cloud data of the aircraft to be predicted and its surrounding flow field, and convert the three-dimensional point cloud data into graph structure data. Through the node and edge features in the graph structure data, three-dimensional modeling is achieved, solving the problem of difficult three-dimensional modeling in fluid dynamics simulation. Then, the graph structure data is input into a sonic boom prediction model based on a graph neural network model. Utilizing the structured modeling capabilities of the graph neural network, combined with a multi-turn message passing mechanism, the fluid dynamic relationships between nodes are accurately captured, achieving a precise mapping from the aircraft's shape to the physical characteristics of the sonic boom. The resulting sonic boom prediction is effective and accurate. Compared to fluid dynamics simulation, this application significantly reduces computational costs through data-driven modeling, achieves higher sonic boom prediction efficiency, and supports rapid design iteration, providing an efficient tool for the development of supersonic aircraft. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 A schematic diagram illustrating a scenario where a supersonic aircraft generates a sonic boom, provided as an embodiment of this application.

[0019] Figure 2 A flowchart illustrating a deep learning-based sonic boom prediction method provided in an embodiment of this application;

[0020] Figure 3 A flowchart illustrating another deep learning-based sonic boom prediction method provided in this application embodiment;

[0021] Figure 4 A schematic diagram of the architecture of a sonic boom prediction model provided in an embodiment of this application;

[0022] Figure 5 A flowchart illustrating the method for constructing and training a sonic boom prediction model provided in an embodiment of this application;

[0023] Figure 6 A schematic diagram of a deep learning-based sonic boom prediction device provided in an embodiment of this application;

[0024] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0025] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0027] First, let me explain the terms used in this application:

[0028] A sonic boom occurs when an object travels at a speed greater than or equal to the speed of sound. The sound waves generated by the compressed air cannot propagate ahead of the object, causing the air to be continuously compressed and superimposed, forming a high-temperature, high-pressure, and high-density shock wave in a specific area. This shock wave, also known as a blast wave, travels to the ground and produces a violent, thunderous explosion.

[0029] Figure 1 This is a schematic diagram illustrating a scenario where a supersonic aircraft generates a sonic boom, as provided in an embodiment of this application. Figure 1 As shown, when a supersonic aircraft breaks the sound barrier—that is, flies at a speed greater than or equal to the speed of sound—the sound waves emitted by the aircraft superimpose and accumulate around the aircraft fuselage, producing a powerful sonic boom. This manifests as a dramatic change in pressure within the airflow field surrounding the aircraft, forming a shock wave. After reaching the ground, this shock wave generates a certain impact force, potentially posing a threat to ground personnel, buildings, and the ecological environment. It may also lead to negative perceptions of supersonic flight, limiting its development.

[0030] By predicting the sonic booms generated by supersonic aircraft during flight and optimizing the aircraft's aerodynamic design, the impact of sonic booms on ground personnel, buildings, or the ecological environment can be controlled, thus promoting the research of supersonic aircraft.

[0031] Currently, sonic boom prediction mainly relies on computational fluid dynamics (CFD) simulations. Specifically, a three-dimensional geometric model is constructed based on the aircraft's shape, and a mesh is set to discretize the flow field region. Then, the Navier-Stokes equations are solved using the finite volume method or the finite element method to simulate the formation process of sonic booms in supersonic flow fields, thus completing the prediction of sonic booms.

[0032] However, in computational fluid dynamics simulations, the complex and irregular shape of aircraft makes constructing a three-dimensional geometric model difficult. Furthermore, it requires processing large amounts of mesh, and the solution process is time-consuming, with a single simulation potentially taking hours to days, resulting in extremely high computational costs. In practical applications, the aircraft shape is often optimized based on sonic boom predictions, and sonic boom predictions are then performed again on the optimized aircraft shape, forming an iterative design process. Traditional prediction methods are inefficient enough to meet the demands of rapid iterative design.

[0033] The deep learning-based sonic boom prediction method provided in this application, after acquiring the target data to be predicted, converts the target data into graph structure data including node features and edge features, realizing 3D modeling including the aircraft's shape, thus solving the problem of difficult 3D modeling in fluid dynamics simulation. Then, utilizing the structured feature modeling capability of graph neural networks, combined with a multi-turn message passing mechanism, the fluid dynamic correlation between nodes is learned to achieve accurate prediction of the flow field, and then the sonic boom is predicted based on the predicted flow field. By using deep learning technology to replace computational fluid dynamics simulation, the powerful ability of deep learning to solve high-dimensional, multi-scale problems is significantly reduced, thus significantly reducing the computational cost of sonic boom prediction while ensuring high accuracy.

[0034] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0035] Figure 2 This is a flowchart illustrating a deep learning-based sonic boom prediction method provided in an embodiment of this application. The method provided in this embodiment can be executed by an electronic device with corresponding processing capabilities. Figure 2 As shown, the method includes:

[0036] Step S201: Obtain the target data to be predicted.

[0037] The target data includes three-dimensional point cloud data that characterizes the physical properties of the aircraft's shape and the surrounding flow field.

[0038] The physical properties of the aircraft's shape include, but are not limited to: key parameters of the aircraft's shape (such as wing curvature, fuselage length, etc.), speed, pressure exerted on the aircraft's surface, and air density in the contact area with the aircraft's shape. The physical properties of the surrounding flow field include, but are not limited to: pressure, air velocity, and air density at various points in the surrounding flow field.

[0039] 3D point cloud data is a collection of 3D spatial data consisting of multiple discrete points (nodes). Each point contains geometric coordinates (such as x, y, z) and physical property information (such as pressure, density, and velocity).

[0040] In this step, three-dimensional point cloud data can be generated based on the physical property data of the aircraft's shape and the physical property data of the surrounding flow field.

[0041] Step S202: Convert the target data into graph structure data.

[0042] Graph structure data includes node features and edge features. Node features include location information and physical attributes, while edge features include the topological relationships between nodes.

[0043] Graph-structured data consists of nodes and edges connected by topological relationships, and is used to represent the correlation of unstructured point cloud data.

[0044] In this step, nodes are determined based on multiple discrete points in the 3D point cloud data included in the target data. The characteristics of each node include its location information (i.e., 3D coordinates) and the physical properties of the corresponding aircraft shape or the surrounding flow field. Edges are added between the nodes, and the edge characteristics include the connected nodes and the direction and distance between them. Nodes and edges constitute the graph structure data.

[0045] Step S203: Input the graph structure data obtained by converting the target data into the sonic boom prediction model based on the graph neural network model to obtain the sonic boom prediction result.

[0046] The sonic boom prediction model is built upon a graph neural network model. It is used to learn the fluid dynamic relationships in graph-structured data through multi-round message passing to output the physical characteristics of sonic booms.

[0047] Graph Neural Networks (GNNs) are a deep learning model architecture specifically designed for processing graph-structured data. They achieve information transmission and aggregation through node and edge features, ultimately uncovering the relationships between nodes.

[0048] Message passing refers to the process by which a node updates its own characteristics by aggregating information from its neighboring nodes.

[0049] This application uses a graph neural network model architecture to construct a sonic boom prediction model. By training the sonic boom prediction model, it can learn the hydrodynamic relationships in graph-structured data. Based on these hydrodynamic relationships, the sonic boom prediction model can then output sonic boom prediction results.

[0050] The sonic boom prediction results are the pressure changes at various locations within the area surrounding the aircraft.

[0051] Specifically, the graph-structured data obtained by transforming the target data is input into a pre-trained sonic boom prediction model. The model preprocesses the node and edge features in the graph-structured data, and then uses a multi-round message passing mechanism to pass the features of a node's neighbors to the node's features. Based on the final node features, the model determines the hydrodynamic relationships of the nodes. Based on these hydrodynamic relationships, the physical characteristics of the sonic boom are determined and output.

[0052] The sonic boom prediction results include the physical characteristics of the sonic boom.

[0053] The multi-round message passing mechanism is as follows: the edge feature in the (n+1)th round is generated by concatenating the current edge feature and the node features at both ends, followed by a nonlinear transformation and fusion. The node feature in the (n+1)th round is generated by summing and aggregating its own current feature and the updated edge features of all adjacent edges, followed by a nonlinear transformation and fusion.

[0054] In some embodiments, a multi-scale message passing mechanism can be introduced into the sonic boom prediction model, enabling the model to simultaneously learn local neighbor features and flow field correlations. For example, by setting different levels of neighbor sampling ranges (such as neighbor nodes including neighbor nodes connected to the target node through at least one node), fluid dynamic relationships can be learned based on the different levels of neighbor sampling ranges.

[0055] Specifically, during the message passing phase, neighbor information at multiple scales is aggregated for each node simultaneously. Features at different scales are weighted and fused using an attention mechanism to obtain multi-scale features. These multi-scale features are then processed in layers: lower layers focus on local details (such as peak shapes), while higher layers focus on global distribution (such as peak propagation paths).

[0056] By using a multi-scale message passing mechanism, specific fluid dynamic relationships and global structural features are constructed simultaneously, thus solving the problem of one-sided feature extraction caused by traditional graph neural network models relying solely on single-scale neighbor information.

[0057] This application provides a deep learning-based sonic boom prediction method. It acquires 3D point cloud data of the aircraft to be predicted and its surrounding flow field, converts this data into graph-structured data, and achieves 3D modeling through node and edge features, solving the problem of difficult 3D modeling in fluid dynamics simulations. Then, the graph-structured data is input into a sonic boom prediction model based on a graph neural network. Utilizing the structured modeling capabilities of the graph neural network, combined with a multi-turn message passing mechanism, it accurately captures the fluid dynamic relationships between nodes, achieving a precise mapping from the aircraft's shape to the physical characteristics of the sonic boom. The resulting sonic boom prediction is effective and accurate. Compared to fluid dynamics simulation, this application significantly reduces computational costs through data-driven modeling, achieves higher sonic boom prediction efficiency, and supports rapid design iteration, providing an efficient tool for the development of supersonic aircraft.

[0058] Figure 3 This is a flowchart illustrating another deep learning-based sonic boom prediction method provided in an embodiment of this application. The method provided in this embodiment is... Figure 2 Based on the provided embodiments, a more detailed explanation of step S202 is provided, such as... Figure 3 As shown, the method provided in this embodiment includes:

[0059] Step S301: Obtain the target data to be predicted.

[0060] The target data includes three-dimensional point cloud data that characterizes the physical properties of the aircraft's shape and the surrounding flow field.

[0061] In this embodiment, a Cartesian coordinate system can be established for the aircraft to be predicted.

[0062] Specifically, the origin of the coordinate system is the aircraft's nose; the aircraft's axis is the x-axis, i.e., the flight direction, with the positive direction of the x-axis extending from the nose to the tail; the aircraft's lateral axis is the y-axis, i.e., the horizontal direction perpendicular to the flight direction; and the aircraft's longitudinal axis is the z-axis, i.e., the vertical direction perpendicular to the flight direction, with the positive direction of the direction of increasing altitude. z < 0 corresponds to the area below the aircraft, and z > 0 corresponds to the area above the aircraft.

[0063] The plane corresponding to y=0 is the reference plane for sonic boom prediction in this application. This plane represents the region with the highest accuracy of sonic boom data, which is also the target prediction region of this application. In other words, this application is used to perform sonic boom prediction on the region corresponding to the y=0 plane.

[0064] Step S302: Determine the initial sampling area by excluding areas in the target data that are not disturbed by sonic booms and areas that are disturbed by sonic booms to a less than preset level.

[0065] Based on conventional experience, it is known that during supersonic aircraft flight, the area around the aircraft is not always affected by sonic boom disturbances, and the degree of disturbance varies across the affected areas. Sampling and processing the entire area would result in a massive amount of data, reducing the efficiency of sonic boom prediction. Therefore, this embodiment divides the area around the aircraft, sampling and analyzing only the areas affected by sonic boom disturbances and those with a high degree of impact from them.

[0066] In this step, the initial sampling area is defined as the area where the location of each node in the target data is not in the area unaffected by the sonic boom or in the area with a low degree of sonic boom disturbance, that is, the area affected by the sonic boom disturbance is less than the preset level of influence.

[0067] According to research in this field, sonic boom propagation exhibits a specific conical disturbance boundary. The region on one side of the cone is unaffected by the sonic boom, and its pressure change is zero. This conical surface can be defined by a formula. The formula can be expressed as:

[0068]

[0069]

[0070]

[0071] Where r is the distance from a point on the cone surface to the x-axis, and θ is the rotation angle.

[0072] Therefore, as seen in the plane with y=0, there are many regions on the left side of the cone where the pressure change is zero; these regions are unaffected by the sonic boom. These can be directly excluded, leaving only the areas near the cone and the rear side where the pressure change is significant.

[0073] The preset impact level is a pre-set configurable parameter that can be determined based on the required sonic boom prediction accuracy. In some embodiments, areas with denser grids are more affected by sonic boom disturbances, and the preset impact level can be determined by the grid density.

[0074] Since the target prediction region is the plane with y=0, that is, the physical characteristics of the sonic boom within the region corresponding to the y=0 plane, we can exclude regions in the target data that are less affected by sonic boom disturbances than a preset level. This can be done by excluding regions farther from the y=0 plane, as the impact of the sonic boom decreases with distance. The boundaries of these regions are the boundaries of those regions where the impact of the sonic boom disturbance is less than the preset level. For example, we can retain regions where 0.15 > y > -0.15 and exclude other regions. Similarly, since the region directly above the aircraft has minimal impact on ground sonic booms and is not relevant to the study, we can exclude this region. For example, we can retain regions where 0.1 > z > -1 and exclude other regions.

[0075] For example, by excluding regions in the target data that are not disturbed by sonic booms and regions that are affected by sonic booms to a less than preset level, the initial sampling region is determined to be 0.15 > y > -0.15, 0.1 > z > -1.

[0076] Step S303: Perform downsampling within the initial sampling area.

[0077] In this step, multiple nodes of the target data are uniformly downsampled within the initial sampling area.

[0078] Specifically, based on the nodes in the target data located in the initial sampling region, a uniform downsampling method is used to sample multiple nodes.

[0079] Uniform downsampling can preserve the density characteristics of discrete points in the aircraft shape and the surrounding flow field.

[0080] Optionally, downsampling can be performed within the initial sampling area, specifically including: increasing the ratio of key region nodes to the total number of nodes by adjusting parameter settings to focus sampling on key regions; limiting the size of the point cloud after downsampling, and preserving the original sonic boom curve features through interpolation processing.

[0081] In this embodiment, key area nodes include nodes on the aircraft's exterior. The total number of nodes is the number of nodes included in the initial sampling area.

[0082] Specifically, by adjusting parameter settings, the ratio of key area nodes to the total number of nodes is increased, thereby increasing the number of nodes sampled in the key area, i.e., focusing on sampling in the key area.

[0083] For example, the ratio of nodes on the aircraft shape to the total number of nodes is increased to 15% to 20% to increase the number of nodes sampled in the aircraft shape region.

[0084] Limit the size of the downsampled point cloud, i.e., the number of nodes obtained from sampling, to balance accuracy and processing efficiency. For example, limit the number of nodes obtained from sampling, i.e., limit the point cloud size to 30,000 nodes per sonic boom curve.

[0085] Furthermore, the original sonic boom curve characteristics are preserved through interpolation.

[0086] For example, preserving the original sonic boom curve characteristics through interpolation can be achieved by selecting an appropriate interpolation method (such as cubic splines or piecewise high-order polynomials) based on the physical properties of the sonic boom (such as peak overvoltage, waveform steepness, spectral distribution, etc.). Using the interpolation method, a smooth transition curve is constructed between the sampled nodes.

[0087] This sampling method can significantly compress the original massive point cloud, such as reducing 8 million points to the 500,000 level, while maintaining the original sonic boom curve characteristics. This reduces computational load while ensuring the representativeness of the sampled nodes. Furthermore, by increasing the ratio of nodes in key regions and focusing on sampling nodes in critical areas, key features can be accurately captured, computational resource allocation optimized, and the uniformity of data feature distribution ensured when performing sonic boom prediction in different scenarios, thus improving the generalization ability of sonic boom prediction.

[0088] In some embodiments, a dynamic adaptive sampling strategy can be introduced to dynamically determine the sampling density of key regions based on the sonic boom intensity distribution. For example, regions near the sonic boom peak and regions with large pressure gradients can be identified as key regions.

[0089] Specifically, based on the target data and conventional patterns, the sonic boom intensity (such as pressure change gradient and peak density) of each node is predicted. Regions with an average sonic boom intensity exceeding a preset intensity threshold are identified as critical regions. Based on the average sonic boom intensity of these critical regions, the ratio of critical region nodes to the total number of nodes is increased to prioritize sampling in these critical regions.

[0090] By dynamically adjusting the sampling density, the method retains detailed information on the key features of the sonic boom while avoiding the storage and computational overhead of redundant target data. At the same time, dynamic sampling can more accurately capture the distribution characteristics of the sonic boom, improving the adaptability of the method to complex flow fields.

[0091] Step S304: Based on the downsampled target data, generate a mesh graph structure using the triangulation algorithm, delete edges with lengths exceeding a preset threshold, remove isolated nodes and isolated subgraphs, and take the largest connected subgraph to obtain the graph structure data corresponding to the target data.

[0092] Triangulation is an algorithm that divides a set of points into non-overlapping triangles, suitable for generating mesh graph structures. In this application, the triangulation algorithm can be based on the 3D Delaunay triangulation algorithm.

[0093] An isolated node is a node in a graph structure that has no edges. An isolated subgraph is a subgraph in which nodes are only connected to nodes within the isolated subgraph by edges, and have no connections to nodes outside the isolated subgraph. Isolated nodes cannot perform message passing, and isolated subgraphs cannot perform message passing with other subgraphs. Furthermore, isolated nodes and isolated subgraphs may correspond to local noise; these nodes are not related to the global sonic boom physical characteristics and do not need to be retained.

[0094] Specifically, based on the nodes corresponding to the target data obtained after downsampling, a mesh graph structure is generated using a triangulation algorithm. The mesh graph structure consists of multiple triangles, with each triangle's vertex representing a node. The edges of all triangles in the mesh graph structure are traversed, and their lengths are calculated. Edges exceeding a preset threshold are deleted. After deleting edges exceeding the preset threshold, isolated nodes and isolated subgraphs are identified and removed. Based on the remaining maximum connected subgraph, the graph structure data corresponding to the target data is determined.

[0095] For example, the graph structure data corresponding to the target data can be represented as G=(V, E), where V is the set of node features and E is the set of edge features. The set of node features V can be represented as:

[0096] V = (x, y, z, p0, den0, v0, type)

[0097] Where x, y, and z are the three-dimensional coordinates of each node in the maximally connected subgraph, p0 is the pressure of each node, den0 is the air density of each node, v0 is the velocity of each node, and type is the category of each node. The categories of each node include boundary nodes, aircraft surface nodes, and nodes of the surrounding flow field.

[0098] The set of edges E can be represented as: E = (u1, u2), where u1 and u2 are the nodes connected by the edge.

[0099] The set of edges E can also be represented as E = (u, r). Here, u represents the direction vector of the edge, that is, the relative orientation of the two nodes connected by the edge in three-dimensional space, i.e., u = (x2−x1, y2−y1, z2−z1). (x1, y1, z1) are the three-dimensional coordinates of node u1, and (x2, y2, z2) are the three-dimensional coordinates of node u2. r is the magnitude between the two nodes.

[0100] Step S305: Input the graph structure data obtained by converting the target data into the sonic boom prediction model based on the graph neural network model to obtain the sonic boom prediction result.

[0101] The sonic boom prediction model is used to learn the fluid dynamic relationships in graph-structured data through multi-round message passing in order to output the physical characteristics of the sonic boom.

[0102] The physical characteristics of a sonic boom include the pressure, density, and air velocity at various points after the boom occurs.

[0103] For example, the sonic boom prediction results include sonic boom physical characteristics, which include the node pressure p, the node air density den, and the node velocity v.

[0104] In this application, downsampling and triangulation-based mesh graph structuring are used to solve the problems of data scarcity and difficulty in 3D modeling in traditional methods, providing high-quality input for deep learning models.

[0105] Figure 4 This is a schematic diagram of the architecture of a sonic boom prediction model provided in an embodiment of this application. Figure 4 As shown, the sonic boom prediction model provided in this embodiment includes an encoder, a processor, and a decoder.

[0106] The encoder is used to normalize node and edge features and convert node and edge features into implicit features of nodes and edges; node features include location information, pressure, density, velocity and node type, and edge features include direction vector and magnitude; the processor is used to learn the fluid dynamics relationship between nodes through multi-round message passing; the decoder is used to convert the implicit features of nodes into predicted physical features.

[0107] Node features include node location information (3D coordinates), node pressure, node density (air density), node velocity, and node type. Node types include boundary nodes, aircraft surface nodes, and nodes in the surrounding flow field. Edge features include the direction vectors of the two nodes connected by the edge and the magnitude between the two nodes.

[0108] The graph-structured data is input into a sonic boom prediction model based on a graph neural network. The encoder of the sonic boom prediction model normalizes the node and edge features and inputs them into a fully connected layer, converting them into implicit features of nodes and edges.

[0109] For example, the encoder normalizes the node features v = (x, y, z, p0, den0, v0, type) and edge features e = (u, r). Here, v represents the node features of each node in the node feature set V, and e represents the edge features of each edge in the edge feature set E. The normalized node features v and edge features e are then input into two different fully connected layers, respectively, to transform them into implicit features of nodes and edges with a specific dimension (e.g., 128 dimensions).

[0110] The implicit features of nodes and edges are input into the processor. The processor learns the fluid dynamic relationships between nodes through multiple rounds of message passing using the implicit features of nodes and edges.

[0111] For example, the processor learns the features of neighboring nodes and edges from the implicit features of nodes and edges through multiple rounds of message passing, updates its own features, and inputs the updated features into a fully connected layer to obtain features of a specific dimension (such as 128 dimensions). This feature of a specific dimension can represent the fluid dynamic relationship between nodes.

[0112] Features of a specific dimension are input into the decoder. The decoder then transforms these features into predicted physical features.

[0113] For example, the decoder inputs features of a specific dimension into a fully connected layer, and the output dimension of the fully connected layer is set to 5, predicting the sonic boom physical characteristics of pressure, density, and three axial velocities.

[0114] The sonic boom prediction model learns the three-dimensional dynamic relationship between nodes through a multi-round message passing mechanism and extracts high-order features by combining a fully connected layer, thus achieving accurate prediction of irregular flow fields.

[0115] In this application, in order for the sonic boom prediction model to accurately output sonic boom prediction results, it is also necessary to construct and train the sonic boom prediction model. Therefore, this application also provides a method for constructing and training a sonic boom prediction model.

[0116] Specifically, the deep learning-based sonic boom prediction method in this application also includes a method for constructing and training a sonic boom prediction model.

[0117] The method for constructing and training a sonic boom prediction model includes: acquiring a sonic boom dataset, which includes three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields; constructing a sonic boom prediction model and training the constructed sonic boom prediction model based on the sonic boom dataset; evaluating and optimizing the sonic boom prediction model until the model performance meets the preset performance requirements, thus obtaining a trained sonic boom prediction model.

[0118] In this embodiment, based on a standard aircraft model (such as the 69DWB standard aircraft model), three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields are generated, and a sonic boom dataset is constructed. Labels are added to the data in the sonic boom dataset, and the labels are the sonic boom physical features indicated by the corresponding standard aircraft model.

[0119] The construction of the sonic boom dataset can refer to the steps of converting the target data into graph structure data in the method provided in the above embodiments.

[0120] A sonic boom prediction model is constructed based on a graph neural network model architecture. The sonic boom prediction model is trained using a sonic boom dataset.

[0121] The sonic boom prediction model is evaluated using a loss function. If the loss function does not meet the preset conditions, the sonic boom prediction model is optimized. This optimization may involve adjusting the model's structure, training parameters, or data processing methods using a preset optimization algorithm until the model's performance meets the preset performance requirements, such as ensuring the loss function meets the preset conditions. If the loss function meets the preset conditions, the training or optimization of the sonic boom prediction model is complete, resulting in a well-trained sonic boom prediction model.

[0122] By training, evaluating, and optimizing the sonic boom prediction model, the transformation from data to functionality was achieved, enabling the sonic boom prediction model to effectively and reliably output sonic boom prediction results.

[0123] Optionally, a sonic boom dataset can be obtained, including: modifying the shape parameters of a standard aircraft model to obtain different aircraft shapes; calculating the sonic boom results for each aircraft shape based on fluid simulation technology, and generating three-dimensional point cloud data that characterizes the physical properties of different aircraft shapes and the surrounding flow field.

[0124] In this step, 3D point cloud data of a standard aircraft model is acquired, and the shape parameters of the standard aircraft model are modified to obtain different aircraft shapes. Based on fluid simulation technology, the sonic boom results for each aircraft shape are calculated. According to the aircraft shape and the physical property information used in the simulation, 3D point cloud data representing the physical properties of different aircraft shapes and the surrounding flow field are generated. The labels for the 3D point cloud data are the calculated sonic boom results.

[0125] The three-dimensional point cloud data constitutes the sonic boom dataset.

[0126] For example, based on the 3D point cloud data of a standard aircraft model using a 69DWB standard model, its shape is modified to obtain 100 aircraft models with different shapes. Fluid simulation technology is used to perform fluid simulations on the obtained aircraft models, such as calculating the sonic boom physical characteristics of each aircraft model using simulation software, i.e., the sonic boom results. Based on the modified aircraft shapes and the physical property information used in the simulation, 3D point cloud data representing the physical property information of different aircraft shapes and the surrounding flow field is generated. The sonic boom results are used as labels for the 3D point cloud data.

[0127] In some embodiments, different physical property information can be used to perform fluid simulation, generating multiple three-dimensional point cloud data.

[0128] By performing fluid simulation on a standard aircraft model, multiple three-dimensional point cloud data are generated, and a sonic boom dataset is constructed, which solves the problem of insufficient training samples in existing technologies.

[0129] Optionally, the sonic boom prediction model is evaluated, specifically including: setting a peak mean square error (MSE) evaluation index; and evaluating the model performance of the sonic boom prediction model based on the MSE and global mean square errors of the sonic boom curves output by the sonic boom prediction model and the actual sonic boom curves, combined with the MSE evaluation index.

[0130] The mean square error of the peak is an error metric for the position and intensity of the peak in a sonic boom waveform. The global mean square error is an evaluation metric that averages the prediction errors of all nodes.

[0131] The mean square error of the peak (rmse) is:

[0132]

[0133] Where n is the number of peaks present in both the actual sonic boom curve (label) and the predicted sonic boom curve (output of the sonic boom prediction model), and x p To predict the location of the peak in the sonic boom curve, x r The position of the peak in a real sonic boom curve, dp p To predict the pressure change at the peak of a sonic boom, dp r denoted as the pressure change at the peak in a real sonic boom curve, and m represents the number of peaks present only in a real sonic boom curve.

[0134] The peak mean square error evaluation index is a configurable parameter that can be set.

[0135] In this step, the positions of all peaks in the sonic boom curve are detected. Based on the spatial positions and physical properties of the predicted and actual peaks, the mean square error of the peaks is calculated. The prediction errors of all nodes are calculated to obtain the global mean square error. If the mean square error of the peaks does not meet the set evaluation index, the sonic boom prediction model is optimized until the model performance meets the preset performance requirements, resulting in a well-trained sonic boom prediction model.

[0136] If the global mean square error does not meet the preset global conditions, the sonic boom prediction model is optimized until the model performance meets the preset performance requirements, thus obtaining a well-trained sonic boom prediction model.

[0137] If the mean square error of the peak meets the set mean square error evaluation index and the global mean square error meets the preset global conditions, then the training of the sonic boom prediction model is considered complete.

[0138] While calculating the mean square error of all nodes, spatial error calculation is introduced for the peak position offset, which enhances the physical interpretability of the evaluation results. Furthermore, the peak-specific evaluation index guides the optimization of the sonic boom prediction model, ensuring the prediction accuracy of the sonic boom prediction model for key physical features (such as peak position) and improving the reliability of the sonic boom prediction model in practical applications.

[0139] Compared with traditional computational fluid dynamics simulation, the method proposed in this invention improves the running efficiency by 6-7 times under the same data scale and by up to 100 times under the same prediction accuracy.

[0140] Figure 5 This is a flowchart illustrating the method for constructing and training a sonic boom prediction model provided in this embodiment. To better understand the method provided in this embodiment, the method for constructing and training a sonic boom prediction model is described in detail. Figure 5 As shown, the method provided in this embodiment includes:

[0141] Step S501: Modify the external parameters of the standard aircraft model to obtain different aircraft shapes.

[0142] Step S502: Based on fluid simulation technology, the sonic boom results for each aircraft shape are calculated, and three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields are generated.

[0143] Step S503: Determine the initial sampling area by excluding areas in the target data that are not disturbed by sonic booms and areas that are disturbed by sonic booms to a less than preset level.

[0144] Step S504: Perform downsampling within the initial sampling area.

[0145] Step S505: Based on the downsampled target data, generate a mesh graph structure using the triangulation algorithm, delete edges with lengths exceeding a preset threshold, remove isolated nodes and isolated subgraphs, and take the largest connected subgraph to obtain the graph structure data corresponding to the target data.

[0146] Step S506: Construct a sonic boom prediction model and train the constructed sonic boom prediction model based on the sonic boom dataset.

[0147] Step S507: Evaluate the sonic boom prediction model.

[0148] Step S508: Optimize the sonic boom prediction model.

[0149] Figure 6 This is a schematic diagram of a deep learning-based sonic boom prediction device provided in an embodiment of this application. Figure 6 As shown, the deep learning-based sonic boom prediction device provided in this embodiment includes: a data acquisition module 601, a data processing module 602, and a sonic boom prediction module 603.

[0150] The data acquisition module 601 is used to acquire the target data to be predicted, which includes three-dimensional point cloud data representing the physical properties of the aircraft shape and the surrounding flow field. The data processing module 602 is used to convert the target data into graph structure data, which includes node features and edge features. The node features include location information and physical properties, and the edge features include the topological relationships between nodes. The sonic boom prediction module 603 is used to input the graph structure data obtained from the target data into a sonic boom prediction model based on a graph neural network model to obtain the sonic boom prediction result. The sonic boom prediction model is used to learn the fluid dynamic relationships in the graph structure data through multi-round message passing in order to output the physical features of the sonic boom.

[0151] Optionally, the data processing module 602 includes a sampling area determination unit, a sampling unit, and a graph structure data determination unit, wherein:

[0152] The sampling region determination unit is used to determine the initial sampling region by excluding regions in the target data that are not disturbed by sonic booms and regions that are affected by sonic booms to a less than preset level. The sampling unit is used to perform downsampling within the initial sampling region. The graph structure data determination unit is used to generate a mesh graph structure based on the downsampled target data using a triangulation algorithm, delete edges with lengths exceeding a preset threshold, remove isolated nodes and isolated subgraphs, and take the largest connected subgraph to obtain the graph structure data corresponding to the target data.

[0153] Optional, sampling unit, specifically used for:

[0154] By adjusting parameter settings, the ratio of key area nodes to the total number of nodes is increased to focus sampling on key areas; the size of the point cloud after downsampling is limited, and the original sonic boom curve features are preserved through interpolation.

[0155] Optionally, the sonic boom prediction model includes: an encoder for normalizing node and edge features and converting node and edge features into implicit features of nodes and edges; node features include location information, pressure, density, velocity, and node type, and edge features include direction vector and magnitude; a processor for learning the hydrodynamic relationships between nodes through multi-round message passing; and a decoder for converting the implicit features of nodes into predicted physical features.

[0156] Optionally, the deep learning-based sonic boom prediction device also includes a sonic boom prediction model maintenance module, which comprises a dataset acquisition unit, a construction unit, and an evaluation and optimization unit, wherein:

[0157] The dataset acquisition unit is used to acquire the sonic boom dataset, which includes three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields; the construction unit is used to construct the sonic boom prediction model and train the constructed sonic boom prediction model based on the sonic boom dataset; the evaluation and optimization unit is used to evaluate and optimize the sonic boom prediction model until the model performance meets the preset performance requirements, and obtain the trained sonic boom prediction model.

[0158] Optional, the dataset acquisition unit is specifically used for:

[0159] By modifying the shape parameters of a standard aircraft model, different aircraft shapes are obtained. Based on fluid simulation technology, the sonic boom results of each aircraft shape are calculated, and three-dimensional point cloud data representing the physical properties of different aircraft shapes and the surrounding flow field are generated.

[0160] Optional, evaluation and optimization unit, specifically used for:

[0161] Set the peak mean square error evaluation index; evaluate the model performance of the sonic boom prediction model based on the peak mean square error and global mean square error of the sonic boom curve output by the sonic boom prediction model and the actual sonic boom curve, combined with the peak mean square error evaluation index; optimize the sonic boom prediction model until the model performance meets the preset performance requirements, and obtain the trained sonic boom prediction model.

[0162] The deep learning-based sonic boom prediction device provided in this embodiment can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0163] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device of this embodiment may include: at least one processor 701; and a memory 702 communicatively connected to at least one processor; wherein the memory 702 stores instructions that can be executed by at least one processor 701, and the instructions are executed by at least one processor 701 to cause the electronic device to perform the method as described in any of the above embodiments.

[0164] Optionally, the memory 702 can be either standalone or integrated with the processor 701. When the memory 702 is set up independently, the device also includes a bus for connecting the memory 702 and the processor 701.

[0165] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0166] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0167] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0168] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0169] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0170] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0171] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0172] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0173] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0174] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0175] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0176] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A deep learning-based method for predicting sonic booms, characterized in that, include: Acquire target data to be predicted, the target data including three-dimensional point cloud data characterizing the physical properties of the aircraft shape and the surrounding flow field; The target data is converted into graph structure data, which includes node features and edge features. The node features include location information and physical attributes, and the edge features include the topological relationships between nodes. The graph structure data obtained by converting the target data is input into a sonic boom prediction model based on a graph neural network model to obtain sonic boom prediction results. The sonic boom prediction model is used to learn the fluid dynamics relationship in the graph structure data through multi-round message passing in order to output the physical characteristics of the sonic boom.

2. The method according to claim 1, characterized in that, The process of converting the target data into graph structure data specifically includes: The initial sampling area is determined by excluding areas in the target data that are not disturbed by sonic booms and areas that are affected by sonic booms to a less than preset level. Downsampling is performed within the initial sampling area; Based on the downsampled target data, a mesh graph structure is generated using a triangulation algorithm. Edges exceeding a preset threshold in length are deleted, and isolated nodes and isolated subgraphs are removed. The largest connected subgraph is then selected to obtain the graph structure data corresponding to the target data.

3. The method according to claim 2, characterized in that, The downsampling within the initial sampling area specifically includes: By adjusting parameter settings, the ratio of key area nodes to the total number of nodes is increased to focus sampling on the key areas; The size of the point cloud after downsampling is limited, and the original sonic boom curve features are preserved through interpolation.

4. The method according to claim 1, characterized in that, The sonic boom prediction model includes: An encoder is used to normalize node features and edge features, and to convert the features of nodes and edges into implicit features of nodes and edges; the node features include location information, pressure, density, velocity and node type, and the edge features include direction vector and magnitude. The processor is used to learn the fluid dynamics relationships between nodes through multi-round message passing; The decoder is used to convert the implicit features of nodes into predicted physical features.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: Acquire a sonic boom dataset, which includes three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields; A sonic boom prediction model is constructed, and the constructed sonic boom prediction model is trained based on the sonic boom dataset; Evaluate and optimize the sonic boom prediction model until its performance meets the preset performance requirements, thus obtaining a well-trained sonic boom prediction model.

6. The method according to claim 5, characterized in that, The acquisition of the sonic boom dataset includes: By modifying the external parameters of a standard aircraft model, different aircraft shapes can be obtained; Based on fluid simulation technology, the sonic boom results of each aircraft shape are calculated, and three-dimensional point cloud data representing the physical properties of different aircraft shapes and surrounding flow fields are generated.

7. The method according to claim 5, characterized in that, The evaluation of the sonic boom prediction model specifically includes: Set the peak mean square error evaluation index; Based on the peak mean square error and global mean square error of the sonic boom curve output by the sonic boom prediction model and the actual sonic boom curve, and combined with the peak mean square error evaluation index, the model performance of the sonic boom prediction model is evaluated.

8. A deep learning-based sonic boom prediction device, characterized in that, include: The data acquisition module is used to acquire the target data to be predicted, which includes three-dimensional point cloud data that characterizes the physical properties of the aircraft shape and the surrounding flow field. The data processing module is used to convert the target data into graph structure data, which includes node features and edge features. The node features include location information and physical attributes, and the edge features include the topological relationships between nodes. The sonic boom prediction module is used to input the graph structure data obtained by converting the target data into a sonic boom prediction model based on a graph neural network model to obtain the sonic boom prediction result; the sonic boom prediction model is used to learn the fluid dynamic relationship in the graph structure data through multi-round message passing in order to output the physical characteristics of the sonic boom.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images