CN-122021370-A - Forward design method and system for broadband transducer

Abstract

The invention discloses a forward design method and a forward design system of a broadband transducer, which comprise the steps of constructing a transducer structure topological graph, extracting node characteristics and adjacency relations to form a structure topological characteristic stream, extracting transducer physical field parameters including material properties, boundary conditions and excitation conditions to form a physical field parameter stream, designing a double-flow graph neural network architecture, respectively carrying out characteristic extraction and fusion on the structure topological characteristic stream and the physical field parameter stream, introducing a parameter enhancement mechanism, enhancing the characterization capability of the network on the acoustic characteristics of the transducer through weighting key physical parameters, constructing a forward prediction model, and realizing rapid prediction of a broadband transducer sound source level curve, impedance characteristics and an operating frequency band. The invention breaks through the efficiency bottleneck that the traditional design method relies on a large number of simulation iterations, realizes the collaborative modeling of structural topological features and physical field parameters, and remarkably improves the design efficiency and the prediction accuracy of the broadband transducer.

Inventors

• GAO BING
• She Yingsen
• Zhuo Jinliang
• LUO AN
• XU XIUXIAN
• Zhao Nengtong
• NING QIAN
• XU QIANMING
• WANG LEI
• GONG HEYAN
• YAN XU
• HUANG ZHONGSHENG

Assignees

• 湖南大学

Dates

Publication Date

20260512

Application Date

20260415

Claims (6)

1. 1. A method for forward designing a broadband transducer, comprising the steps of: S1, sampling a design space of a transducer and constructing a heterogeneous data set, namely acquiring structural parameter-grid topology-frequency response characteristic sample data covering the design space of the broadband transducer through finite element simulation, extracting macroscopic physical structural parameter characteristics and microscopic geometric topological characteristics from the structural parameter-grid topology-frequency response characteristic sample data, and constructing the heterogeneous data set; S2, constructing a parameter enhanced dual-flow graph neural network proxy model, namely designing a dual-flow architecture, wherein the dual-flow architecture comprises a structural topological feature flow and a physical field parameter flow; The method comprises the steps of using a graph neural network to extract structural connection relation and space layout information of a broadband transducer, wherein a structural topological feature stream is used for processing microscopic geometric topological features of the transducer, and the microscopic geometric topological features of the transducer comprise a macroscopic feature vector M and a microscopic topological graph G; The physical field parameter flow is used for processing physical field parameters, wherein the physical field parameters comprise key structural parameters, working conditions and excitation conditions, and the two flows are processed in parallel and then are subjected to multi-scale feature fusion to construct a parameter enhanced double-flow-chart neural network proxy model capable of simultaneously representing the influence of structural topology and physical field parameters on the output performance of the transducer; s3, training the parameter enhanced double-flow-chart neural network proxy model by adopting the heterogeneous data set to obtain a trained parameter enhanced double-flow-chart neural network model; s4, replacing finite element simulation with the trained parameter enhanced dual-flow graph neural network model, carrying out rapid performance prediction on the candidate structure scheme, screening the candidate structure parameter scheme meeting the broadband requirement through the target acoustic characteristics of the broadband transducer, continuously approaching to the global optimal solution through iterative optimization, and finally outputting the optimal broadband transducer structure.
2. 2. The wideband transducer forward design method of claim 1, wherein the heterogeneous data set is constructed as follows: S11, defining design variables, namely firstly defining the design variables, wherein the design variables comprise a range of a small end radius r 1 of a conical radiation shell, a range of a large end radius r 2 of the conical radiation shell, a range of a conical bottom length r 3 , a range of a conical joint radiation surface length r 4 , a range of a small end thickness d 1 of the conical radiation shell, a range of a large end thickness d 2 of the conical radiation shell, a range of a conical thickness m and a range of a transducer height h, and secondly, adopting an experimental design method of Latin hypercube sampling to generate N groups of parameter combinations in the design space, wherein each group of parameter combinations represents a candidate design scheme; S12, finite element modeling and simulation calculation, namely firstly, parameterizing finite element modeling, automatically establishing a corresponding transducer finite element model for each group of parameter combinations obtained by sampling, generating a geometric model according to structural size parameters, endowing each component with material properties according to material parameters, setting constraint and load according to boundary condition parameters, setting driving conditions according to excitation condition parameters, dividing grids, recording grid node coordinates, unit types and node connection relations, secondly, adopting finite element software to carry out multi-physical field coupling simulation, and solving acoustic characteristics of the transducer, and finally, extracting broadband acoustic characteristics of the transducer from simulation results, wherein the broadband acoustic characteristics comprise a sound source level curve, a sound pressure result and a working frequency bandwidth; S13, extracting macroscopic physical structure parameter characteristics from each group of parameter combinations to form a structural characteristic vector, wherein the macroscopic physical structure parameter characteristics comprise geometric characteristic parameters, physical field characteristics and excitation working conditions, the geometric characteristic parameters comprise r 1 、r 2 、r 3 、r 4 、d 1 、d 2 , M and H, the physical field characteristics comprise prestress sigma, a bias magnetic field H σ and water depth H water , the excitation working conditions comprise current I and coil turns n, and the macroscopic physical structure parameter characteristics are spliced to form a macroscopic characteristic vector M; S14, extracting microscopic geometric features from a finite element grid model, and representing the local geometric structure and connection relation of a transducer, wherein the method comprises node feature extraction and graph structure construction, wherein node feature extraction is used for extracting node coordinates (X, Y, Z) and node neighborhood curvature, X, Y and Z are respectively the X axis, the Y axis and the Z axis of a node, the graph structure construction is used for constructing a microscopic topological graph G by taking finite element grid nodes as nodes of the graph and taking a unit connection relation as edges of the graph, and finally, performing topological downsampling and simplification, namely downsampling the microscopic topological graph G by adopting a node sampling method, and obtaining a simplified microscopic topological graph G simple after downsampling, wherein the number of nodes is controlled within a range of 2000-5000 and is used as microscopic topological feature input for subsequent graph neural network processing; S15, constructing a structural parameter-grid topology-frequency response characteristic heterogeneous data set, namely firstly integrating a macroscopic feature vector M and a microscopic topological graph G to construct a heterogeneous data set covering a design space, firstly, enabling each sample to correspond to a group of design parameter combinations and comprise data components of the macroscopic feature vector M, a simplified microscopic topological graph G simple and a frequency response characteristic label Y, wherein the frequency response characteristic label Y comprises frequency response characteristics and sound source level curves, the frequency response characteristics comprise resonance frequencies and frequency bandwidths, secondly, dividing the data set, dividing all N samples into a training set, a verification set and a test set according to preset proportion, then, carrying out data standardization and preprocessing, carrying out Z-score standardization on the macroscopic feature vector M to enable the mean value of each feature to be 0 and the variance to be 1, interpolating the frequency response characteristics to unify frequency sampling points, carrying out smoothing treatment on the sound source level curves to remove numerical noise, and finally, executing storage of sample data, wherein the macroscopic feature and physical field parameters are NumPy array, the microscopic topological graph is stored as an adjacent matrix and a node matrix, the frequency response characteristic label is stored as a matrix NumPy, and the array data is stored in a matrix form.
3. 3. The method for forward designing a broadband transducer according to claim 1, wherein the data processing flow of the parameter enhanced dual-flow graph neural network proxy model is as follows: S21, taking the structural topology characteristic flow of the parameter enhanced double-flow graph neural network as a first branch, constructing a topology sensing encoder based on GraphSAG algorithm by the first branch, stacking the topology sensing encoder with the structure topology characteristic flow The layer GRAPHSAGE convolution module comprises the final characteristics of each node including topology information in a third-order neighborhood of the final characteristics, wherein shallow characteristics capture cell distortion and density degree of local grids, deep characteristics capture geometric structures in a larger range, wherein the geometric structures comprise integral curvature change of a radiation head and macroscopic layer number distribution of coils, and the node characteristic matrix containing abundant local geometric information is obtained after the transducer microscopic geometric topological characteristics are subjected to multi-layer convolution In order to obtain a fixed dimension vector characterizing the overall transducer geometry, a global averaging pooling layer is introduced as a Readout function: (6) In the formula, As a function of the aggregation function, As a vector of the topological feature, V represents any node in the node set; s22, taking the physical field parameter flow of the parameter enhanced double-flow graph neural network as a second branch, constructing a global parameter encoder based on a multi-layer perceptron MLP by the second branch, stacking 4 full-connection layers to form the global parameter encoder so as to fully extract the nonlinear relation of the feature combination, and enabling the global parameter encoder to be in a full-extraction characteristic combination Representation of MLP No. The output vector of a layer, the forward propagation process is defined as: (7) In the formula, And (3) with The weight matrix and the bias vector are respectively used for normalizing the characteristic distribution of each layer, preventing gradient from disappearing and accelerating training convergence of a deep network; The linear combination result of the k-th layer is shown, Representing the output feature vector of the k-1 layer, A network layer number index representing a multi-layer perceptron MLP; after 4 layers of nonlinear transformation, the physical field parameters are finally output to be of one dimension Macroscopic physical feature vectors of (a) : (8) Wherein, the Representing physical field parameters; S23, extracting topological feature vectors respectively according to a multi-scale feature fusion mechanism FiLM linear modulation and decoding strategy of the parameter enhanced double-flow-graph neural network proxy model And macroscopic physical feature vector After that, to And Fusion is carried out, comprising two steps: Modulation factor generation from macroscopic physical feature vectors using two lightweight fully connected layers In learning to generate scaling factors for feature channel levels And translation coefficient : (9) Representation generation of scaling coefficients Is the first lightweight fully-connected layer responsible for inputting macroscopic physical features Converted into a scaling weight matrix for affine transformation, Representing translation coefficients Is a second lightweight fully connected layer responsible for characterizing macroscopic physical characteristics Converting into a translation bias matrix for affine transformation; Feature affine transformation using scaling coefficients And translation coefficient For topological feature vector Performing element-by-element operations: (10) In the formula, Representing the Hadamard product; Representing the fused feature vector; fused feature vectors The parameter enhanced double-flow-graph neural network agent model designs a multi-layer decoder which consists of 3 full-connection layers, wherein the last layer is the sound source level of a linear output direct regression transducer; (11) representing a global parameter encoder based on a multi-layer perceptron MLP; Representing the sound source level of the transducer.
4. 4. The method of forward designing a broadband transducer of claim 1, wherein the total loss function when training the parametric enhanced dual flow graph neural network proxy model is ; (12) In the formula, For weighted mean square error, N is the frequency band width, Is a frequency domain weight vector; the real response value of finite element simulation on the ith frequency point, As the predicted response value of the neural network on the ith frequency point, As a weight coefficient for the peak loss, For the weight coefficients of the regularized term, For peak loss or resonance peak error, N represents the total number of sampling frequency points within the frequency band, For a maximum transmit response calculated for the finite element, Maximum emission response predicted for the parametric enhanced dual-flow graph neural network proxy model; as the Frobenius norm sum of squares of all weight matrices, The set contains GRAPHSAGE aggregate weights, MLP coding weights, fiLM generator weights, and decoder weights; The representation refers to a learnable weight parameter of a certain layer in the model.
5. 5. The method of forward designing a wideband transducer of claim 1, wherein the specific steps of step S4 are as follows: s41, initializing a population, namely, firstly, in a design space Internal random generation of initial population Each individual corresponds to a set of multidimensional design vectors All variables are limited within upper and lower bounds to ensure physical rationality; s42, the trained parameter enhancement type double-flow-chart neural network model carries out shunt processing on design parameters of each individual in the population, wherein the first branch is used for carrying out shunt processing according to the parameters Fast reconstruction of geometric meshes Input GRAPHSAGE branch to extract local topological feature, and second branch to extract parameters Direct input MLP branch to extract global nonlinear characteristic, and direct output of predicted TCR frequency response curve after FiLM fusion of local topological characteristic and global nonlinear characteristic ; S43, fitness evaluation and environment selection, wherein TCR frequency response curve is based on prediction The method comprises the steps of calculating three objective function values, namely-3 dB bandwidth, peak response and resonant frequency deviation, performing non-dominant sorting and environment selection operation based on reference points according to the three objective function values, and distinguishing the current Pareto optimal front edge; s44, updating the population and iterating, namely generating a child population through a simulated group intelligent optimization operator if the preset maximum iteration number is not reached, repeating the steps S42 and S43, and repeating the steps until the population converges to the global optimal solution set.
6. 6. A broadband transducer forward design system comprising one or more processors and a memory having one or more programs stored thereon that, when executed by the one or more processors, cause the one or more processors to implement the broadband transducer forward design method of any of claims 1-5.

Description

Forward design method and system for broadband transducer Technical Field The invention relates to the technical field of artificial intelligence and transducer design, in particular to a forward design method and system of a broadband transducer, which are particularly suitable for a scene requiring collaborative modeling of topological features of a transducer structure and multiple physical field parameters and realizing rapid prediction from a design target to acoustic performance. Background The design level of the broadband transducer serving as a core device in the fields of underwater acoustic communication, submarine detection and the like directly determines the upper performance limit of the whole system. The traditional transducer design method generally depends on experience accumulation and a large number of finite element simulation iterations of a designer, and finally the transducer structure meeting the performance requirements is obtained by repeatedly modifying structural parameters, analyzing simulation results and adjusting design schemes. However, this design mode has the following prominent problems: 1. The design period is long and the efficiency is low. In the transducer structural parameter optimization process, performance verification is usually performed by means of finite element simulation. Each parameter adjustment needs to be modeled and solved again, and particularly, the complex multi-physical field coupling problem is solved, and the calculation cost is huge. When facing to a high-dimensional structural parameter space, the traditional serial design mode of 'simulation-modification-re-simulation' often needs weeks or even months, the design period is long due to the high-intensity calculation overhead, the research and development efficiency of the novel transducer is severely restricted, and the requirement of engineering application on quick iteration is difficult to meet. 2. The traditional agent model has difficulty in considering the dual complexity of structural topological characteristics and physical field parameters. Some proxy model methods exist to try to replace finite element simulation by constructing response surfaces or reduced order models, but these methods can only handle limited geometric parameters generally, and cannot effectively characterize the complex structural topological features (such as radiation surface shape, spatial structure, etc.) of the transducer. Meanwhile, the acoustic performance of the transducer is also influenced by coupling of multiple physical field parameters such as material properties, boundary conditions, prestress, excitation conditions and the like, and the conventional method is difficult to uniformly model the different types of characteristics. 3. The ability to forward predict from design goals to structural parameters is lacking. The existing design flow is basically in a reverse mode of trial-error-verification, wherein a group of structural parameters are given, performance indexes are obtained through simulation, and if the performance does not reach the standard, the parameters are adjusted through experience to simulate again. This mode cannot build an end-to-end mapping from design targets (e.g., target bandwidth, sound source level, impedance characteristics) to structural parameters, resulting in a lack of directionality in the design process, which makes performance driven intelligent design difficult to implement. 4. The data feature extraction and fusion mechanism is insufficient. The structural topology of the transducer is essentially a graph of structural data, and the connection between nodes implies critical vibration transmission path information. However, existing approaches mostly reduce structural parameters to scalar or vector inputs, ignoring the important impact of topological relations on acoustic performance. Meanwhile, a complex nonlinear coupling relation exists between physical field parameters and structural topological features, and how to effectively extract and fuse the two heterogeneous features is one of core challenges facing the intelligent design of the current transducer. In summary, how to break through the efficiency bottleneck of the traditional design method, and establish an intelligent transducer design method capable of considering both structural topological features and physical field parameters and having forward prediction capability has become a key scientific problem for promoting the technical development of broadband transducers. Noun interpretation: PA-DSGNN Parameter-Augmented Dual-STREAM GRAPH Neural Networks is based on a parametric enhanced Dual-flow graph neural network model. Disclosure of Invention The invention aims to solve the problems that the design period is long, the traditional proxy model is difficult to consider the structural topological feature and the physical field parameter, the forward prediction capability is lacking and the like in the d