Deep learning-based transducer dynamic characteristic rapid analysis optimization method and system

Abstract

The application discloses a kind of based on deep learning's transducer dynamic characteristic fast analysis optimization method and system, comprehensively consider the transient behavior of transducer under the condition of multi-physical field coupling, construct the time-domain model of giant magnetostrictive transducer, make output characteristic more close to real situation, then establish the deep learning model of giant magnetostrictive transducer, the output characteristic of giant magnetostrictive transducer is calculated quickly, finally search the best working condition parameter combination of giant magnetostrictive transducer by means of particle swarm improved grey wolf optimization algorithm;Compared with the method of existing manual adjustment parameter and carrying out finite element calculation, the present application is more efficient, more accurate, not only can quickly predict output characteristic, and can provide best working condition parameter combination, meet the requirement of real-time simulation.

Description

Technical Field

[0001] This invention relates to the field of rapid analysis of transducer dynamic characteristics, and more specifically, to a method and system for rapid analysis and optimization of transducer dynamic characteristics based on deep learning. Background Technology

[0002] The design methods for giant magnetostrictive transducers mainly include the equivalent circuit method and the finite element analysis method. The equivalent circuit method has a clear physical meaning, while the finite element analysis method has significant advantages in calculating three-dimensional model problems. Both have their own advantages and disadvantages and are widely used. However, the energy conversion process of giant magnetostrictive transducers involves multi-physics coupling of electro-magnetism-mechanics-acoustics, and its input-output relationship has very complex nonlinear characteristics. Moreover, the frequency domain characteristics are difficult to reflect its dynamic characteristics under different operating conditions. Therefore, constructing a time-domain model of the giant magnetostrictive transducer is a prerequisite for transducer design. However, because the time-domain model of the giant magnetostrictive transducer involves multi-physics coupling of electro-magnetism-mechanics-acoustics, and the giant magnetostrictive material exhibits hysteresis, eddy currents, and magnetostriction processes under current excitation, the calculation time of the time-domain multi-physics coupling model of the giant magnetostrictive transducer is too long and the efficiency is too low. Especially in the process of optimizing the design of giant magnetostrictive transducers, it is necessary to optimize and analyze the influence of different operating conditions and different size parameters of the giant magnetostrictive transducer, resulting in a large number of simulation samples. In addition, relying on experience to adjust the parameter combination of the transducer under different operating conditions is not only inefficient but also has a lot of blindness, which further prolongs the entire transducer simulation analysis process. Summary of the Invention

[0003] The main objective of this invention is to propose a rapid analysis and optimization method for the dynamic characteristics of transducers based on deep learning, as well as an analysis and optimization system. It aims to provide an optimization design method for giant magnetostrictive transducers based on a combination of particle swarm optimization algorithm and U-net, which can quickly predict the output characteristics of giant magnetostrictive transducers, such as the output sound source level, thereby solving the above-mentioned technical problems.

[0004] To achieve the above objectives, this invention proposes a rapid analysis and optimization method for the dynamic characteristics of transducers based on deep learning, comprising the following steps:

[0005] S1. Establish a physical model of the supermagnetostrictive transducer;

[0006] S2. Establish the time-domain model of the super magnetostrictive transducer;

[0007] S3. Establish a sample set for the deep learning training model of the super magnetostrictive transducer;

[0008] S4. Establish a deep learning training network for the supermagnetostrictive transducer;

[0009] S5. Process the sample set data, train the sample set, and analyze the results.

[0010] S6. Optimize the supermagnetostrictive transducer using the PSO-GWO algorithm combined with the U-net deep learning model.

[0011] Preferably, the physical model includes an excitation module, a magnetic circuit module, and an output module. The excitation module includes a coil and a permanent magnet, which provide an alternating magnetic field and a bias magnetic field to the giant magnetostrictive rod, respectively. The magnetic circuit module includes a magnetically conductive block, a silicon steel sheet, a permanent magnet, the giant magnetostrictive rod, and a magnetic yoke. The output module includes a magnetically conductive block, a silicon steel sheet, a permanent magnet, the giant magnetostrictive rod, and a bending shell.

[0012] Preferably, step S2 specifically includes:

[0013] S2.1 Construct the electromagnetic field component of the supermagnetostrictive transducer; the weak form equation of the electromagnetic field of the supermagnetostrictive rod is shown in equation (1):

[0014]

[0015] Where H is the magnetic field strength of the giant magnetostrictive rod, and J s σ is the current density of the excitation coil. c Let V be the electrical conductivity of the rod, A be the magnetic vector potential, and V be the magnetic vector potential. A For the volume domain of the electromagnetic field, For a closed region of electromagnetic field, H×n is Magnetic boundary tangential field;

[0016] S2.2 Construct the solid mechanical field part of the super magnetostrictive transducer; the weak form equation of the solid mechanical field of the super magnetostrictive transducer is shown in equation (2):

[0017]

[0018] Where ρ is the density of the bar, ζ is the mechanical damping of the bar, u is the total displacement of the bar, σ is the stress tensor, and F v For a given mechanical volume force, V u Let be the volume domain of the solid mechanical field. Let σn be the closed boundary of the solid mechanical field, and let σn be the surface traction force of the mechanical field boundary.

[0019] S2.3. The coupling relationship of the super magnetostrictive transducer is derived using the Jiles-Atherton model, as shown in equation (3):

[0020]

[0021] Among them, Hsolid H is the magnetic field strength in a solid mechanical field. mf M is the magnetic field strength in the magnetic field. mf M is the magnetization in the magnetic field. solid The magnetization intensity in a solid mechanical field;

[0022] S2.4 Construct the pressure acoustic field part of the super magnetostrictive transducer. The weak form equation of the pressure acoustic field of the super magnetostrictive transducer is shown in equation (5):

[0023]

[0024] Among them, V a Represents the pressure acoustic domain. The symbol represents the boundary between solid mechanics and the sound field, and P represents the sound pressure.

[0025] Preferably, step S3 specifically includes: using the joint simulation function of MATLAB and COMSOL software, preset the input characteristic range of the super magnetostrictive transducer in MATLAB, wherein the parameters of the input characteristics include: permanent magnet residual magnetic flux density Br, pre-compression stress σ, and operating frequency f; taking different combinations of the parameter values ​​as the input of the super magnetostrictive transducer, and then calling COMSOL software through MATLAB to run the transient model of the super magnetostrictive transducer, and sequentially outputting the displacement and sound source level of the super magnetostrictive transducer to obtain several sets of sample data.

[0026] Preferably, step S4 specifically includes: setting the network input layer data structure; defining each convolutional layer, activation layer, max pooling layer, regularization layer, transposed convolutional layer, and depth concatenation layer in a certain order; defining the output layer; defining connection layer branches to create the training network, and using the connectLayers function to connect each layer branch to form the network.

[0027] Preferably, step S5 specifically includes:

[0028] S5.1 Import feature data; import the sample data generated in S3 into the deep learning model of the super magnetostrictive transducer.

[0029] S5.2 Initialize training sets P_train and T_train and test sets P_test and T_test; divide the feature data into training and test sets, then assign the feature input of the training set to P_train and the feature output to T_train, and assign the feature input of the test set to P_test and the feature output to T_test.

[0030] S5.3 Data normalization; The mapminmax function is used to normalize the training set and the test set, unifying the feature input and feature output to the range [0,1].

[0031] S5.4 Data tiling processing; Use the reshape function to tile P_train and T_train into a four-dimensional data structure of [5×1×1×3], consistent with the input layer data structure;

[0032] S5.5 Initialize training parameters; set the optimization function solverName, gradient descent method miniBatchSize, maximum number of training epochs MaxEpochs, initial learning rate InitialLearnRate, learning rate policy LearnRateSchedule, learning rate drop factor LearnRateDropFactor, learning rate drop period LearnRateDropPeriod, validation set data ValidationData, validation frequency ValidationFrequency, and training progress training-progress;

[0033] S5.6. Use the trainNetwork function to train the U-net deep learning network for the magnetostrictive transducer.

[0034] S5.7 Adjust the training parameters and train the deep learning network multiple times;

[0035] S5.8 Model Prediction; After the deep learning model has been trained, the predict function is used to predict the output results.

[0036] S5.9 Data Inverse Normalization: Inverse normalize the prediction results to the initial order of magnitude;

[0037] S5.10 Calculate relevant parameters; calculate mean absolute error and mean square error and plot relevant graphs.

[0038] Preferably, step S6 specifically includes:

[0039] S6.1. Set the residual magnetic flux density Br of the permanent magnet, the pre-compression stress σ, and the operating frequency f as variables to be optimized, and take the output sound source level of the super magnetostrictive transducer as the optimization target. The expression for the fitness value is shown in Equation (8):

[0040] Fitness = max(SL) (8)

[0041] Where Fitness is the fitness value and SL is the output sound source level of the giant magnetostrictive transducer;

[0042] S6.2 Initialize the gray wolf population and related parameters, set the number of gray wolves in the population m, the individual number of gray wolves i, the number of population evolutions k, and the current evolution generation d; set the convergence flag K, coefficient vectors A and C, the optimal solution corresponding to the current gray wolf individual Pbest(i), the optimal solution of the gray wolf population Gbest, the optimal fitness of the current gray wolf individual FitnessPbest(i), and the optimal fitness of the gray wolf population FitnessGbest; initialize the required parameters r1, r2, r3, c1, c2, c3;

[0043] S6.3 Initialize the position x(i) of gray wolves, and perform normalization processing on gray wolf individuals;

[0044] S6.4 Let d=0, K=0, initialize the convergence flag;

[0045] S6.5 Starting from the first gray wolf individual, let i=1, FitnessGbest=0;

[0046] S6.6 Invoke the deep learning model of said giant magnetostrictive transducer to calculate the initial fitness value Fitness(i) of the gray wolf individual, update FitnessPbest(i) by judging the magnitude relationship between the current objective function value Fitness(i) and FitnessGbest, and make FitnessPbest(i) d =Fitness(i), Pbest(i) d =x(i);

[0047] S6.7 If Fitness(i)>FitnessGbest, let FitnessGbest=Fitness(i); if Fitness(i)<FitnessGbest, let FitnessGbest=FitnessGbest;

[0048] S6.8 Update the global optimal position of gray wolves according to the update result of FitnessGbest; if Fitness(i)>FitnessGbest, then Gbest=x(i); if Fitness(i)<FitnessGbest, then Gbest=Gbest;

[0049] S6.9 Determine whether i=m is true. If yes, the calculation is completed, execute S6.10; otherwise, execute S6.9;

[0050] S6.10 Let i=i+1, and return to S6.5;

[0051] S6.11 Let i=1, d=d+1, start iteration;

[0052] S6.12. Introduce the particle swarm velocity update formula v(i). d =w(v(i) d-1 +c1r1(Pbest(1) d -x(i) d )+c2r2(Pbest(2) d -x(i) d )+c3r3(Pbest(3) d -x(i) d and the position update formula x(i) d+1 =x(i) d +vi d Update the location of individual gray wolves;

[0053] S6.13. Calculate the initial fitness value (Fitness(i)) of the individual gray wolf using the deep learning model of the giant magnetostrictive transducer, and let FitnessPbest(i). d =Fitness(i);

[0054] S6.14, If FitnessPbest(i) d >FitnessPbest(i) d-1 Then FitnessPbest(i) d =FitnessPbest(i) d ;like

[0055] FitnessPbest(i) d <FitnessPbest(i) d-1 Then FitnessPbest(i) d =FitnessPbest(i) d-1 ;

[0056] S6.15, Based on FitnessPbest(i) d Update the results, and update the optimal position Pbest(i) for the individual gray wolf. d If FitnessPbest(i) d >FitnessPbest(i) d-1 Then Pbest(i) d =Pbest(i) d FitnessPbest(i) d <FitnessPbest(i) d-1 Then Pbest(i) d =Pbest(i) d-1 ;

[0057] S6.16: Update the global optimal solution of the grey wolf population, if FitnessPbest(i) d >FitnessGbest, let FitnessGbest=FitnessPbest(i) d ; if FitnessPbest(i) d <FitnessGbest, let FitnessGbest=FitnessPbest(i) d ;

[0058] S6.17: Update the global optimal position of grey wolves according to the update result of FitnessGbest. If FitnessPbest(i) d >FitnessGbest, then Gbest=Pbest(i) d ; if FitnessPbest(i) d <FitnessGbest, then Gbest=Gbest;

[0059] S6.18: Determine whether i=m holds. If it holds, the calculation is completed, execute S6.20; otherwise execute S6.19;

[0060] S6.19: Let i=i+1, and return to S6.12;

[0061] S6.20: Determine whether d<k holds. If it holds, execute S6.21; otherwise execute S6.11;

[0062] S6.21: Determine whether the convergence requirements are satisfied: 1) K reaches a preset value; 2) the output result is within a preset range. If converged, output the global optimal solution Gbest of grey wolves; otherwise, let K=K+1, and return to S6.4;

[0063] S6.22: The finally output optimal working condition parameter combination is: Br=5T, σ=24Mpa, f=400Hz, and the giant magnetostrictive transducer is designed with these parameters.

[0064] The present invention also provides a rapid analysis and optimization system for dynamic characteristics of transducers based on deep learning, comprising one or more processors and a memory, wherein one or more programs are stored on the memory, and when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the steps of the above method.

[0065] This invention presents a rapid analysis and optimization method for the dynamic characteristics of transducers based on deep learning. This method comprehensively considers the transient behavior of the transducer under multi-physics coupling conditions, constructing a time-domain model of the giant magnetostrictive transducer to make its output characteristics closer to reality. Then, a deep learning model of the giant magnetostrictive transducer is established to rapidly calculate its output characteristics. Finally, an improved gray wolf optimization algorithm based on particle swarm optimization is used to search for the optimal combination of operating parameters for the giant magnetostrictive transducer. Compared to manually adjusting parameters and performing finite element analysis, this invention is more efficient and accurate, not only rapidly predicting output characteristics but also providing the optimal combination of operating parameters to meet the requirements of real-time simulation. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0067] Figure 1 This is a coupling diagram of the super magnetostrictive transducer of the present invention;

[0068] Figure 2 This invention relates to a deep learning network for a supermagnetic-strict transducer based on the U-net neural network.

[0069] Figure 3 This is a diagram illustrating the deep learning model training process for the super magnetostrictive transducer of this invention.

[0070] Figure 4 This is a diagram showing the prediction results of the deep learning model for the super magnetostrictive transducer of this invention.

[0071] Figure 5 This is a comparison chart of the prediction results and actual results of the deep learning model for the super magnetostrictive transducer of this invention;

[0072] Figure 6 This is a flowchart of the method for optimizing a supermagnetostrictive transducer based on the PSO-GWO algorithm and the U-net deep learning model according to the present invention.

[0073] Figure 7 This is a flowchart of the analysis and optimization method of the present invention.

[0074] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0076] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0077] This invention proposes a rapid analysis and optimization method for the dynamic characteristics of transducers based on deep learning. This method can quickly find the optimal sound source level of a giant magnetostrictive transducer and the corresponding optimal combination of operating parameters, thereby solving the problems of low efficiency of finite element analysis methods and blind selection of various operating parameter combinations in the design process of giant magnetostrictive transducers in the prior art.

[0078] Please refer to Figures 1 to 7 The method includes the following steps:

[0079] Step 1: Establish the physical model of the giant magnetostrictive transducer. This includes an excitation module, a magnetic circuit module, and an output module. The excitation module consists of a coil and a permanent magnet, which provide the alternating magnetic field and bias magnetic field to the giant magnetostrictive rod, respectively. The magnetic circuit module consists of a magnetic conductor, a silicon steel sheet, a permanent magnet, the giant magnetostrictive rod, and a magnetic yoke. The output module consists of a magnetic conductor, a silicon steel sheet, a permanent magnet, the giant magnetostrictive rod, and a bending shell.

[0080] Step 2: Establish the time-domain model of the giant magnetostrictive transducer. The specific electro-magnetic-mechanical-acoustic coupling relationship is shown in the attached figure. Figure 1 As shown.

[0081] Step 2.1, Electromagnetic field component of the giant magnetostrictive transducer. Under the excitation of an AC coil, the weak form equation of the electromagnetic field of the giant magnetostrictive rod is derived from Maxwell's equations using the weighted residual method, as shown in equation (1):

[0082]

[0083] Where H is the magnetic field strength of the giant magnetostrictive rod, and J s σ is the current density of the excitation coil. c Let V be the electrical conductivity of the rod, A be the magnetic vector potential, and V be the magnetic vector potential. A For the volume domain of the electromagnetic field, Let H×n be a closed region of electromagnetic field. Magnetic boundary tangential field.

[0084] Step 2.2, Solid mechanical field part of the giant magnetostrictive transducer. Under the action of pre-stress, the weighted residual method is applied to the stress balance equation of the giant magnetostrictive rod in the entire solid mechanical domain to obtain the weak form equation of the solid mechanical field of the giant magnetostrictive transducer, as shown in equation (2):

[0085]

[0086] Where ρ is the density of the bar, ζ is the mechanical damping of the bar, u is the total displacement of the bar, σ is the stress tensor, and F v For a given mechanical volume force, V u Let be the volume domain of the solid mechanical field. Let σn be the closed boundary of the solid mechanical field, and let σn be the surface traction force of the mechanical field boundary.

[0087] Step 2.3: Derive the coupling relationship of the giant magnetostrictive transducer using the JA model. (See attached diagram) Figure 1 As shown, the giant magnetostrictive rod generates strain λ under the combined action of the permanent magnet bias magnetic field Br and the coil AC excitation magnetic field Hmf. That is, the strain exists in the solid mechanical field, but the excitation comes from the magnetic field. The specific relationship is shown in equation (3):

[0088]

[0089] Among them, H solid H is the magnetic field strength in a solid mechanical field. mf M is the magnetic field strength in the magnetic field. mf M is the magnetization in the magnetic field. solid Let be the magnetization intensity in the solid mechanical field.

[0090] Using the JA model to represent the supermagnetostrictive rod can simulate the nonlinear process during the magnetization of the rod and explain the cause of hysteresis. The JA model can be expressed as shown in equation (4):

[0091]

[0092] Among them, H e.solid The effective magnetic field strength is given by α, the mean field coupling coefficient is given by λ. s Let M be the saturation magnetostriction coefficient, σ be the total internal stress of the rod, and M be the total internal stress of the rod. s M is the saturation magnetization. an.solid Where is the hysteresis-free magnetization, a is the shape factor, and M is the magnetization. rev and M irr δ represents the reversible and irreversible magnetization, respectively; c is the reversibility coefficient; δ is the direction coefficient; and k is the pinning coefficient.

[0093]

[0094] Step 2.4, Pressure Acoustic Field of the Giant Magnetostrictive Transducer. Based on the coupling of the magnetic field and the solid mechanical field, the sound pressure caused by the deformation of the transducer shell in the pressure acoustic fluid domain is calculated. By weighting the basic equation of the sound pressure wave in the fluid domain, its weak form equation can be obtained, as shown in equation (5):

[0095]

[0096] Among them, V a Represents the pressure acoustic domain. The symbol represents the boundary between solid mechanics and the sound field, and P represents the sound pressure.

[0097] Furthermore, the sound pressure of the supermagnetostrictive transducer at a distance of d meters can be obtained, as shown in equation (6):

[0098]

[0099] Where, ρ ω U is the density of water. r and A h These represent the vibration velocity and effective cross-sectional area of ​​the radiating surface of the supermagnetostrictive transducer, respectively.

[0100] Furthermore, the sound source level of the supermagnetostrictive transducer at a distance of 1 meter from the acoustic axis is shown in equation (7):

[0101] TCR=20log(p(1) / I0)+120 (7)

[0102] Where I0 is the effective value of the emission current, and TCR is the output sound source level.

[0103] Step 3: Establish a sample set for the deep learning training model of the giant magnetostrictive transducer. Using the co-simulation capabilities of MATLAB and COMSOL software, the input characteristic ranges of the giant magnetostrictive transducer are preset in MATLAB, including: permanent magnet remanent flux density Br, with a value range of [0.6T, 5T] and a step size of 0.4T; pre-stress σ, with a value range of [0MPa, 32MPa] and a step size of 4MPa; and operating frequency f, with a value range of [100Hz, 400Hz] and a step size of 100Hz. Different combinations of the above parameters are used as inputs to the giant magnetostrictive transducer. Then, MATLAB calls COMSOL software to run the transient model of the giant magnetostrictive transducer, sequentially outputting the displacement and sound source level of the giant magnetostrictive transducer. Therefore, a total of 12×9×4 sets of data samples, or 432 sets of samples, are obtained.

[0104] Step 4: Establish a deep learning training network for the giant magnetostrictive transducer. This invention utilizes the U-net deep learning neural network to achieve rapid calculation of the output sound source level of the giant magnetostrictive transducer under the coupling problem of electromagnetic field and solid mechanical field. The training network structure is shown in the attached figure. Figure 2 As shown.

[0105] Step 4.1: Set up the network input layer data structure. Use the imageInputLayer function to create the input layer, and set the data size to InputSize = [5, 1, 1].

[0106] Step 4.2: Define convolutional layer 1. Use the convolution2dLayer function to create convolutional layer 1, setting FilterSize = [3,1], NumFilters = 64, and Padding = same. That is, the kernel size is 3×1, the number of kernels is 64, and the data size is the same before and after convolution.

[0107] Step 4.3: Define activation layer 1. The ReluLayer layer performs a threshold operation on each element of the input, where any value less than zero is set to zero.

[0108] Step 4.4: Define convolutional layer 2, the same as convolutional layer 1.

[0109] Step 4.5: Define activation layer 2, the same as activation layer 1.

[0110] Step 4.6: Define max pooling layer 1. Use the maxPooling2dLayer function to create max pooling layer 1, let PoolSize = [2,1], Padding = [0,0,0,0], that is, the pooling size is 2×1, the step size is 1, the padding data is 0, and the data format becomes [4×1×64×1].

[0111] Step 4.7: Define convolutional layer 3. Use the convolution2dLayer function to create convolutional layer 3, set FilterSize = [3,1], NumFilters = 128, Padding = same, that is, the kernel size is 3×1, the number of kernels is 128, and the data size before and after convolution is the same, and the data format is [4×1×128×1].

[0112] Step 4.8: Define activation layer 3, the same as activation layer 1.

[0113] Step 4.9: Define convolutional layer 4, the same as convolutional layer 3.

[0114] Step 4.10: Define activation layer 4, the same as activation layer 1.

[0115] Step 4.11: Define max pooling layer 2. Use the maxPooling2dLayer function to create max pooling layer 2, let PoolSize = [2,1], Padding = [0,0,0,0], that is, the pooling size is 2×1, the step size is 1, the padding data is 0, and the data format becomes [3×1×128×1].

[0116] Step 4.12: Define convolutional layer 5. Use the convolution2dLayer function to create convolutional layer 5, set FilterSize = [3,1], NumFilters = 256, Padding = same, that is, the kernel size is 3×1, the number of kernels is 256, and the data size before and after convolution is the same, and the data format is [3×1×256×1].

[0117] Step 4.13: Define activation layer 5, the same as activation layer 1.

[0118] Step 4.14: Define convolutional layer 6, the same as convolutional layer 5.

[0119] Step 4.15: Define activation layer 6, the same as activation layer 1.

[0120] Step 4.16: Define max pooling layer 3. Use the maxPooling2dLayer function to create max pooling layer 3, let PoolSize = [2,1], Padding = [0,0,0,0], that is, the pooling size is 2×1, the step size is 1, the padding data is 0, and the data format becomes [2×1×256×1].

[0121] Step 4.17: Define convolutional layer 7. Use the convolution2dLayer function to create convolutional layer 7, set FilterSize = [3,1], NumFilters = 512, Padding = same, that is, the kernel size is 3×1, the number of kernels is 128, and the data size before and after convolution is the same, and the data format is [2×1×512×1].

[0122] Step 4.18: Define activation layer 7, the same as activation layer 1.

[0123] Step 4.19: Define convolutional layer 8, the same as convolutional layer 7.

[0124] Step 4.20: Define activation layer 8, the same as activation layer 1.

[0125] Step 4.21: Define regularization layer 1. Use the dropoutLayer function and set Probability = 0.5 to prevent overfitting.

[0126] Step 4.22: Define max pooling layer 4. Use the maxPooling2dLayer function to create max pooling layer 4, let PoolSize = [2,1], Padding = [0,0,0,0], that is, the pooling size is 2×1, the step size is 1, the padding data is 0, and the data format becomes [1×1×512×1].

[0127] Step 4.23: Define convolutional layer 9. Use the convolution2dLayer function to create convolutional layer 7, set FilterSize = [3,1], NumFilters = 1024, Padding = same, that is, the kernel size is 3×1, the number of kernels is 128, and the data size before and after convolution is the same, and the data format is [1×1×1024×1].

[0128] Step 4.24: Define activation layer 9, the same as activation layer 1.

[0129] Step 4.25: Define convolutional layer 10, the same as convolutional layer 9.

[0130] Step 4.26: Define activation layer 10, the same as activation layer 1.

[0131] Step 4.27: Define regularization layer 2, the same as regularization layer 1.

[0132] Step 4.28: Define transposed convolutional layer 1. Use the transposedConv2dLaye function to create transposed convolutional layer 1. Let FilterSize = [3,1], NumFilters = 512, Stride = [2,1], Padding = same, that is, the kernel size is 3×1, the number of kernels is 512, and the data size before and after convolution is the same, and the data format is [2×1×512×1].

[0133] Step 4.29: Define activation layer 11, the same as activation layer 1.

[0134] Step 4.30: Define depth concatenation layer 1. Use the depthConcatenationLayer function to concatenate the outputs of activation layer 11 and activation layer 8, with the data format being [2×1×1024×1].

[0135] Step 4.31: Define convolutional layer 11. Use the tconvolution2dLayer function to create transposed convolutional layer 11, with FilterSize = [3,1], NumFilters = 512, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 512, and the data size before and after convolution is the same, with the data format being [2×1×512×1].

[0136] Step 4.32: Define activation layer 12, the same as activation layer 1.

[0137] Step 4.33: Define convolutional layer 12, the same as convolutional layer 11.

[0138] Step 4.34: Define activation layer 13, the same as activation layer 1.

[0139] Step 4.35: Define transposed convolutional layer 2. Use the transposedConv2dLaye function to create transposed convolutional layer 2, with FilterSize = [3,1], NumFilters = 256, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 256, and the data size before and after convolution is the same, with the data format being [3×1×256×1].

[0140] Step 4.36: Define activation layer 14, the same as activation layer 1.

[0141] Step 4.37: Define depth concatenation layer 2. Use the depthConcatenationLayer function to concatenate the outputs of activation layer 14 and activation layer 6, with the data format being [3×1×512×1].

[0142] Step 4.38: Define convolutional layer 13. Use the tconvolution2dLayer function to create transposed convolutional layer 13, with FilterSize = [3,1], NumFilters = 256, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 256, and the data size before and after convolution is the same, with the data format being [3×1×256×1].

[0143] Step 4.39: Define activation layer 15, the same as activation layer 1.

[0144] Step 4.40: Define convolutional layer 14, the same as convolutional layer 13.

[0145] Step 4.41: Define activation layer 16, the same as activation layer 1.

[0146] Step 4.42: Define transposed convolutional layer 3. Use the transposedConv2dLaye function to create transposed convolutional layer 3, with FilterSize = [3,1], NumFilters = 128, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 256, and the data size before and after convolution is the same, with the data format being [4×1×128×1].

[0147] Step 4.43: Define activation layer 17, the same as activation layer 1.

[0148] Step 4.44: Define depth concatenation layer 3. Use the depthConcatenationLayer function to concatenate the outputs of activation layer 17 and activation layer 4, with the data format being [3×1×512×1].

[0149] Step 4.45: Define convolutional layer 15. Use the tconvolution2dLayer function to create transposed convolutional layer 15, with FilterSize = [3,1], NumFilters = 128, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 128, and the data size before and after convolution is the same, with the data format being [4×1×128×1].

[0150] Step 4.46: Define activation layer 18, the same as activation layer 1.

[0151] Step 4.47: Define convolutional layer 16, the same as convolutional layer 15.

[0152] Step 4.48: Define activation layer 19, the same as activation layer 1.

[0153] Step 4.49: Define transposed convolutional layer 4. Use the transposedConv2dLaye function to create transposed convolutional layer 4, with FilterSize = [3,1], NumFilters = 64, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 64, and the data size before and after convolution is the same, with the data format being [5×1×64×1].

[0154] Step 4.50: Define activation layer 20, the same as activation layer 1.

[0155] Step 4.51: Define depth concatenation layer 4. Use the depthConcatenationLayer function to concatenate the outputs of activation layer 20 and activation layer 2, with the data format being [5×1×128×1].

[0156] Step 4.52: Define convolutional layer 17. Use the tconvolution2dLayer function to create transposed convolutional layer 17, with FilterSize = [3,1], NumFilters = 64, Stride = [2,1], and Padding = same. That is, the kernel size is 3×1, the number of kernels is 64, and the data size before and after convolution is the same, with the data format being [5×1×64×1].

[0157] Step 4.53: Define activation layer 21, the same as activation layer 1.

[0158] Step 4.54: Define convolutional layer 18, the same as convolutional layer 17.

[0159] Step 4.55: Define activation layer 22, the same as activation layer 1.

[0160] Step 4.56: Define convolutional layer 19. Use the tconvolution2dLayer function to create transposed convolutional layer 19, with FilterSize = [1,1], NumFilters = 1, Stride = [7,1], and Padding = same. That is, the kernel size is 1×1, the number of kernels is 1, and the data size before and after convolution is the same, with the data format being [1×1×1×1].

[0161] Step 4.57: Define the output layer. The regressionLayer uses mean squared error as the loss function.

[0162] Step 4.58: Define connection layer branches and create the network. Use the connectLayers function to connect each layer branch to form the network. Connect activation layer 2 with max pooling layer 1 and deep concatenation layer 4; connect activation layer 4 with max pooling layer 2 and deep concatenation layer 3; connect activation layer 6 with max pooling layer 3 and deep concatenation layer 2; connect activation layer 8 with regularization layer 1 and deep concatenation layer 1; connect activation layer 11 with deep concatenation layer 1; connect activation layer 14 with deep concatenation layer 2; connect activation layer 17 with deep concatenation layer 3; connect activation layer 20 with deep concatenation layer 4.

[0163] Step 5: Process, train, and analyze the sample data. The 432 sets of sample data for giant magnetostrictive transducers generated in Step 3 are processed and trained. The specific steps are as follows:

[0164] Step 5.1: Import feature data. Import the 432 sets of sample data of the giant magnetostrictive transducer generated in Step 3 into the deep learning model of the giant magnetostrictive transducer.

[0165] Step 5.2: Initialize training sets P_train and T_train, and test sets P_test and T_test. Divide the feature data into training and test sets, then assign the feature inputs of the training set to P_train and the feature outputs to T_train, and assign the feature inputs of the test set to P_test and the feature outputs to T_test.

[0166] Step 5.3: Data Normalization. The mapminmax function is used to normalize the training and test sets, unifying the feature inputs and outputs to the range [0,1]. This reduces the training burden on deep learning networks and accelerates network convergence.

[0167] Step 5.4: Data tiling. Use the reshape function to tile P_train and T_train into a four-dimensional data structure of [5×1×1×3], consistent with the input layer data structure.

[0168] Step 5.5: Initialize training parameters. Set the optimization function solverName, gradient descent method miniBatchSize, maximum training epochs MaxEpochs, initial learning rate InitialLearnRate, learning rate policy LearnRateSchedule, learning rate drop factor LearnRateDropFactor, learning rate drop period LearnRateDropPeriod, validation set data ValidationData, validation frequency ValidationFrequency, and training progress training-progress.

[0169] Step 5.6: Use the trainNetwork function to train the U-net deep learning network for the supermagnetostrictive transducer.

[0170] Step 5.7: Adjust training parameters and train the deep learning network multiple times. The learning rate (InitialLearnRate) and the learning rate drop factor (LearnRateDropFactor) have a significant impact on the training effect. Through multiple training sessions, the learning rate (InitialLearnRate) was finally determined to be 0.9, and the learning rate drop factor (LearnRateDropFactor) was determined to be 0.001.

[0171] Step 5.8, Model Prediction. After the deep learning model is trained, the predict function is used to predict the output.

[0172] Step 5.9: Data Inverse Normalization. Inverse normalize the prediction results to the initial order of magnitude.

[0173] Step 5.10: Calculate relevant parameters. Calculate the mean absolute error and mean square error, and plot the relevant graphs. (See attached image) Figure 3 Appendix Figure 4 Appendix Figure 5 As shown, the training process of the deep learning network for the giant magnetostrictive transducer converges quickly, and the predicted value of the output sound source level is consistent with the true value, indicating that the deep learning model can quickly and accurately predict the output sound source level of the giant magnetostrictive transducer, with good results.

[0174] Step 6: Optimize the giant magnetostrictive transducer using the PSO-GWO algorithm and the U-net deep learning model. The giant magnetostrictive transducer is optimized using a swarm intelligence optimization algorithm combined with a deep learning model. The specific process is shown in the attached figure. Figure 6 As shown, the detailed optimization steps are as follows:

[0175] Step 6.1: Set the residual magnetic flux density Br, preload stress σ, and operating frequency f of the permanent magnet as variables to be optimized. Set the range of Br to [0.6T, 5T], the range of σ to [0MPa, 32MPa], and the range of operating frequency f to [90Hz, 160Hz]. Take the output sound source level of the transducer as the optimization target. The expression for the fitness value is shown in Equation (8):

[0176] Fitness = max(SL) (8)

[0177] Where Fitness is the fitness value and SL is the output sound source level of the giant magnetostrictive transducer.

[0178] Step 6.2: Initialize the gray wolf population and related parameters. Set the gray wolf population size m, gray wolf individual number i, population evolution number k, current evolution generation d; convergence flag K, coefficient vectors A and C, the optimal solution Pbest(i) corresponding to the current gray wolf individual, the optimal solution Gbest of the gray wolf population, the optimal fitness of the current gray wolf individual FitnessPbest(i), and the optimal fitness of the gray wolf population FitnessGbest; initialize the required parameters r1, r2, r3, c1, c2, and c3.

[0179] Step 6.3: Initialize the gray wolf position x(i) and normalize the individual gray wolves;

[0180] Step 6.4: Set d = 0 and K = 0 to initialize the convergence flag.

[0181] Step 6.5: Starting from the first gray wolf individual, let i = 1 and FitnessGbest = 0.

[0182] Step 6.6: Calculate the initial fitness value (Fitness(i)) of the individual gray wolf using the deep learning model of the giant magnetostrictive transducer. Update FitnessPbest(i) by comparing the current objective function value Fitness(i) with FitnessGbest, and set FitnessPbest(i) to the desired value. d = Fitness(i), Pbest(i) d =x(i).

[0183] Step 6.7: If Fitness(i)>FitnessGbest, let FitnessGbest=Fitness(i); if Fitness(i)<FitnessGbest, let FitnessGbest=FitnessGbest.

[0184] Step 6.8: Update the global optimal position of gray wolves according to the update result of FitnessGbest. If Fitness(i)>FitnessGbest, then Gbest=x(i); if Fitness(i)<FitnessGbest, then Gbest=Gbest.

[0185] Step 6.9: Determine whether i=m holds. If yes, the calculation is completed, proceed to step 10); otherwise, proceed to step 9).

[0186] Step 6.10: Let i=i+1, and return to step 5).

[0187] Step 6.11: Let i=1, d=d+1, and start the iteration.

[0188] Step 6.12: Introduce the particle swarm velocity update formula v(i) d =w(v(i) d-1 +c1r1(Pbest(1) d -x(i) d )+c2r2(Pbest(2) d -x(i) d )+c3r3(Pbest(3) d -x(i) d )) and the position update formula x(i) d+1 =x(i) d +vi d to update the position of individual gray wolves.

[0189] Step 6.13: Call the deep learning model of the giant magnetostrictive transducer to calculate the initial fitness value Fitness(i) of the individual gray wolf, and set FitnessPbest(i) d =Fitness(i).

[0190] Step 6.14: If FitnessPbest(i) d >FitnessPbest(i) d-1 , then FitnessPbest(i) d =FitnessPbest(i) d ; if FitnessPbest(i) d<FitnessPbest(i) d-1 , then FitnessPbest(i) d =FitnessPbest(i) d-1 .

[0191] Step 6.15, update the result according to FitnessPbest(i) d to update the individual optimal position Pbest(i) of grey wolves d . If FitnessPbest(i) d >FitnessPbest(i) d-1 , then Pbest(i) d =Pbest(i) d ; when FitnessPbest(i) d <FitnessPbest(i) d-1 , then Pbest(i) d =Pbest(i) d-1 .

[0192] Step 6.16, update the global optimal solution of the grey wolf population, if FitnessPbest(i) d >FitnessGbest, then set FitnessGbest=FitnessPbest(i) d ; if FitnessPbest(i) d <FitnessGbest, then set FitnessGbest=FitnessPbest(i) d .

[0193] Step 6.17, update the result based on FitnessGbest, and update the global optimal position of grey wolves. If FitnessPbest(i) d >FitnessGbest, then Gbest=Pbest(i) d ; if FitnessPbest(i) d <FitnessGbest, then Gbest=Gbest.

[0194] Step 6.18, determine whether i=m holds. If it holds, the calculation is completed, step 20) is executed; otherwise, step 19) is executed.

[0195] Step 6.19, set i=i+1, and return to step 12).

[0196] Step 6.20, determine whether d<k holds. If it holds, step 21) is executed; otherwise, step 11) is executed.

[0197] Step 6.21: Determine if the convergence requirements are met: 1) K reaches the preset value; 2) The output result is within the preset value range. If convergence is achieved, output the global optimal solution Gbest; otherwise, let K = K + 1 and return 4.

[0198] Step 6.22: The final optimal operating parameter combination is: Br = 5T, σ = 24Mpa, f = 400Hz, and the super magnetostrictive transducer is designed based on these parameters.

[0199] Furthermore, this invention also provides a rapid analysis and optimization system for transducer dynamic characteristics based on deep learning, including one or more processors and a memory. The memory stores one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the steps of the method described above. The specific steps of this method are as described in the above embodiments. Since this analysis system adopts all the technical solutions of all the above embodiments, it possesses at least all the beneficial effects brought about by the technical solutions of the above embodiments, and will not be elaborated upon further here.

[0200] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent transformations made based on the inventive concept of the present invention and the contents of the specification and drawings of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for rapid analysis and optimization of transducer dynamic characteristics based on deep learning, characterized in that, Includes the following steps: S1. Establish a physical model of the supermagnetostrictive transducer; S2. Establish the time-domain model of the supermagnetostrictive transducer; wherein, step S2 specifically includes: S2.1 Construct the electromagnetic field component of the supermagnetostrictive transducer; the weak form equation of the electromagnetic field of the supermagnetostrictive rod is shown in equation (1): (1) Where H is the magnetic field strength of the giant magnetostrictive rod, and J s To the current density of the excitation coil, σ c Let be the electrical conductivity of the rod, and A be the magnetic vector potential. V A For the volume domain of the electromagnetic field, ∂V A It is a closed region of electromagnetic field. H × n for ∂V A Magnetic boundary tangential field; S2.2 Construct the solid mechanical field part of the super magnetostrictive transducer; the weak form equation of the solid mechanical field of the super magnetostrictive transducer is shown in equation (2): (2) in, ρ The density of the bar. ζ For the mechanical damping of the bar, u This represents the total displacement of the bar. σ For stress tensor, F v For a given mechanical volume force, V u Let be the volume domain of the solid mechanical field. ∂V u For the closed boundary of the solid mechanical field, σn The surface traction force at the boundary of the mechanical field; S2.

3. The coupling relationship of the super magnetostrictive transducer is derived using the Jiles-Atherton model, as shown in equation (3): (3) in, Let be the magnetic field strength in a solid mechanical field. The magnetic field strength is in the magnetic field. The magnetization intensity in the magnetic field. The magnetization intensity in a solid mechanical field; S2.4 Construct the pressure acoustic field part of the super magnetostrictive transducer. The weak form equation of the pressure acoustic field of the super magnetostrictive transducer is shown in equation (5): (5) in, V a Represents the pressure acoustic domain. ∂V sa Represents the boundary between solid mechanics and the sound field, where P represents sound pressure; and S3. Establish a sample set for the deep learning training model of the super magnetostrictive transducer; S4. Establish a deep learning training network for the supermagnetostrictive transducer; S5. Process the sample set data, train the sample set, and analyze the results. S6. The supermagnetostrictive transducer is optimized using the PSO-GWO algorithm combined with the U-net deep learning model; wherein... The physical model includes an excitation module, a magnetic circuit module, and an output module. The excitation module includes a coil and a permanent magnet, which provide an alternating magnetic field and a bias magnetic field to the giant magnetostrictive rod, respectively. The magnetic circuit module includes a magnetically conductive block, a silicon steel sheet, a permanent magnet, the giant magnetostrictive rod, and a magnetic yoke. The output module includes a magnetically conductive block, a silicon steel sheet, a permanent magnet, the giant magnetostrictive rod, and a bending shell.

2. The method for rapid analysis and optimization of transducer dynamic characteristics based on deep learning as described in claim 1, characterized in that, Step S3 specifically includes: using the joint simulation function of MATLAB and COMSOL software, preset the input characteristic range of the supermagnetic-strict transducer in MATLAB. The parameters of the input characteristics include: the remanent magnetic flux density of the permanent magnet. Br Prestress σ and operating frequency f Different combinations of the parameters are used as inputs to the super magnetostrictive transducer. Then, the transient model of the super magnetostrictive transducer is run by calling COMSOL software through MATLAB, and the displacement and sound source level of the super magnetostrictive transducer are output sequentially to obtain several sets of sample data.

3. The method for rapid analysis and optimization of transducer dynamic characteristics based on deep learning as described in claim 2, characterized in that, Step S4 specifically includes: setting the network input layer data structure; defining each convolutional layer, activation layer, max pooling layer, regularization layer, transposed convolutional layer, and depth concatenation layer in a certain order; defining the output layer; defining connection layer branches to create the training network, and using the connectLayers function to connect each layer branch to form the network.

4. The method for rapid analysis and optimization of transducer dynamic characteristics based on deep learning as described in claim 3, characterized in that, Step S5 specifically includes: S5.1 Import feature data; import the sample data generated in S3 into the deep learning model of the supermagnetostrictive transducer; S5.2 Initialize training sets P_train and T_train and test sets P_test and T_test; divide the feature data into training and test sets, then assign the feature input of the training set to P_train and the feature output to T_train, and assign the feature input of the test set to P_test and the feature output to T_test; S5.3 Data normalization; The mapminmax function is used to normalize the training set and the test set, unifying the feature input and feature output to the range [0,1]. S5.4 Data tiling processing; Use the reshape function to tile P_train and T_train into a four-dimensional data structure of [5×1×1×3], consistent with the input layer data structure; S5.5 Initialize training parameters; set the optimization function solverName, gradient descent method miniBatchSize, maximum number of training epochs MaxEpochs, initial learning rate InitialLearnRate, learning rate policy LearnRateSchedule, learning rate drop factor LearnRateDropFactor, learning rate drop period LearnRateDropPeriod, validation set data ValidationData, validation frequency ValidationFrequency, and training progress training-progress; S5.

6. Use the trainNetwork function to train the U-net deep learning network for the magnetostrictive transducer. S5.7 Adjust the training parameters and train the deep learning network multiple times; S5.8 Model Prediction; After the deep learning model has been trained, the predict function is used to predict the output results. S5.9 Data Inverse Normalization: Inverse normalize the prediction results to the initial order of magnitude; S5.10 Calculate relevant parameters; calculate mean absolute error and mean square error and plot relevant graphs.

5. The method for rapid analysis and optimization of transducer dynamic characteristics based on deep learning as described in claim 4, characterized in that, Step S6 specifically includes: S6.1, the residual magnetic flux density of the permanent magnet Br The prestress σ The operating frequency f The output sound source level of the super magnetostrictive transducer is set as the variable to be optimized, and the expression for the fitness value is shown in Equation (8): (8) Where Fitness is the fitness value. SL This is the output sound source level of the giant magnetostrictive transducer; S6.2 Initialize the gray wolf population and related parameters, setting the gray wolf population size m, gray wolf individual number i, population evolution number k, current evolution generation d; convergence flag K, coefficient vectors A and C, the optimal solution Pbest(i) corresponding to the current gray wolf individual, the optimal solution Gbest of the gray wolf population, the optimal fitness of the current gray wolf individual FitnessPbest(i), and the optimal fitness of the gray wolf population FitnessGbest; initialize the required parameters r1, r2, r3, c1, c2, c3; S6.

3. Initialize the position x(i) of the gray wolf and normalize the individual gray wolves; S6.

4. Set d=0, K=0, and initialize the convergence flag; S6.5 Starting from the first individual gray wolf, let i=1 and FitnessGbest=0; S6.

6. Calculate the initial fitness value (Fitness(i)) of the individual gray wolf using the deep learning model of the magnetostrictive transducer. Update FitnessPbest(i) by determining the relationship between the current objective function value (Fitness(i)) and FitnessGbest, and set FitnessPbest(i) to... d =Fitness(i), Pbest(i) d = x(i); S6.7 If Fitness(i) > FitnessGbest, then let FitnessGbest = Fitness(i); if Fitness(i) < FitnessGbest, then let FitnessGbest = FitnessGbest. S6.

8. Based on the FitnessGbest update result, update the global optimal position of the gray wolf; if Fitness(i)>FitnessGbest, then Gbest = x(i); if Fitness(i)<FitnessGbest, then Gbest = Gbest; S6.9 Determine if i=m is true. If true, the calculation is complete, and execute S6.10; otherwise, execute S6.

9. S6.

10. Let i = i + 1, and return to S6.5; S6.

11. Let i=1, d=d+1, and begin the iteration; S6.

12. Introduce the particle swarm velocity update formula v(i). d =w(v(i) d-1 +c1r1(Pbest(1) d -x(i) d )+c2r2(Pbest(2) d - x(i) d +c3r3 (Pbest(3)) d - x(i) d and the position update formula x(i) d+1 = x(i) d + vi d Update the location of individual gray wolves; S6.

13. Calculate the initial fitness value (Fitness(i)) of the individual gray wolf using the deep learning model of the giant magnetostrictive transducer, and let FitnessPbest(i). d =Fitness(i); S6.14, If FitnessPbest(i) d >FitnessPbest(i) d-1 Then FitnessPbest(i) d =FitnessPbest(i) d If FitnessPbest(i) d <FitnessPbest(i) d-1 Then FitnessPbest(i) d =FitnessPbest(i) d-1 ; S6.15, Based on FitnessPbest(i) d Update the results, and update the optimal position Pbest(i) for the individual gray wolf. d If FitnessPbest(i) d >FitnessPbest(i) d-1 Then Pbest(i) d = Pbest(i) d FitnessPbest(i) d <FitnessPbest(i) d-1 Then Pbest(i) d = Pbest(i) d-1 ; S6.

16. Update the global optimal solution for the gray wolf population. If FitnessPbest(i) d > FitnessGbest, then let FitnessGbest= FitnessPbest(i) d If FitnessPbest(i) d < FitnessGbest, then let FitnessGbest= FitnessPbest(i) d ; S6.

17. Based on the FitnessGbest update result, update the global optimal position of the gray wolf; if FitnessPbest(i) d >FitnessGbest, then Gbest =Pbest(i) d If FitnessPbest(i) d < FitnessGbest, then Gbest =Gbest; S6.18, determining whether i=m is satisfied; if yes, the calculation is completed, and step S6.20 is executed; otherwise, step S6.19 is executed; S6.19, setting i=i+1, and returning to step S6.12; S6.20, determining whether d<k is satisfied; if yes, step S6.21 is executed; otherwise, step S6.11 is executed; S6.21, determining whether a convergence requirement is satisfied: 1) K reaches a preset value; 2) an output result is within a preset value range; if convergence is achieved, outputting a gray wolf global optimal solution Gbest; otherwise, setting K=K+1, and returning to step S6.4; S6.22, The final optimal combination of operating parameters is: Br =5T, σ =24 MPa f =400Hz, and the super magnetostrictive transducer was designed using this parameter.

6. A rapid analysis and optimization system for the dynamic characteristics of transducers based on deep learning, characterized in that, comprising: one or more processors; a memory, having one or more programs stored thereon, where when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the steps of the method according to any one of claims 1 to 5.

Images