MEMS sensor intelligent design method and system based on generative network

Abstract

The invention discloses an MEMS sensor intelligent design method and system based on a generative network, which are used for efficiently generating a plurality of feasible structures under given sensor performance requirements, and the method comprises the following steps: obtaining MEMS sensor experimental data based on finite element simulation; constructing and training a verification network to realize forward design from a sensor structure to performance parameters; constructing and training a generative network CVAE-GAN to realize reverse design from performance to structure parameters; and constructing a multi-solution generator G-MS model for efficiently generating multiple groups of structure parameters meeting performance requirements. Compared with an MEMS sensor design method based on finite element simulation, time-consuming multi-physical field simulation and manual parameter adjustment are not needed, and the method has the advantages of being high in speed, high in precision and intelligent; compared with other MEMS sensor intelligent design methods, the method has the advantages that the efficiency and quality of generating a plurality of different design schemes under given performance requirements are remarkably improved, the design period is shortened, and rapid iterative design is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of MEMS sensor design, and in particular relates to a MEMS sensor intelligent design method and system based on a generative network. Background Art

[0002] MEMS (Micro-Electro-Mechanical System) sensors are functional devices that realize micro-electromechanical system perception and signal processing. MEMS sensors are manufactured using micro-nano processing technology and work based on the special effects of sensitive materials. They have the advantages of small size, low power consumption, high sensitivity, and high integration. They are widely used in smart phones, automobiles, medical equipment and other fields.

[0003] The traditional design process of MEMS sensors often frequently uses time-consuming finite element simulations and requires experienced designers to participate in the parameter optimization process, which greatly consumes manpower and time costs. In recent years, deep learning has been applied to the design process of MEMS sensors, which has improved the design efficiency to a certain extent.

[0004] Faced with the need to generate multiple design solutions during the MEMS sensor design process, researchers usually use the following two methods: create large data sets and cluster them to find different solutions, or traverse the inputs to find similar outputs. The first method is relatively expensive and can often only search for a small number of different solutions, or no different solutions; the second method can theoretically find all feasible solutions, but it is extremely time-consuming. Summary of the invention

[0005] In view of the defects existing in the prior art, the present invention proposes a MEMS sensor intelligent design method and system based on a generative network, which are used to efficiently solve the problem of multiple solution generation in the MEMS sensor design process.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A generative network-based intelligent design method for MEMS sensors is used to efficiently generate multiple feasible structures under given sensor performance requirements, including:

[0008] Step S1: Acquire MEMS sensor experimental data through finite element simulation;

[0009] Step S2: construct a verification network to realize the forward design from sensor structure to performance parameters, and use MEMS sensor experimental data for training;

[0010] Step S3: construct a generative network CVAE-GAN to achieve reverse design from performance to structural parameters, and use MEMS sensor experimental data for training;

[0011] Step S4: Based on the verification network and the generative network CVAE-GAN, a multi-solution generator G-MS model is constructed to generate multiple sets of structural parameters that meet performance requirements.

[0012] Furthermore, the MEMS sensor experimental data in step S1 includes structural parameters X and performance parameters Y; wherein the structural parameters X include continuous variables describing the sensor geometry with a dimension of m, and Y includes continuous variables describing the sensor performance with a dimension of n; the verification network in step S2 is an artificial neural network with an input dimension of n and an output dimension of m, and the mean square error is used to statistically characterize the error level of the network.

[0013] Furthermore, in step S3, the generative network CVAE-GAN includes an encoding network E, a generating network G and a discriminative network D; the number of input units of the encoding network E is n+m, and the number of output nodes is 2; the number of input units of the generating network G is the dimension m of the performance parameter plus the dimension of the noise Z, and the number of output units is n; the number of input units of the discriminative network D is n+m, the output node is 1, and the output unit is activated using the Sigmoid function.

[0014] Furthermore, the total loss function of the generative network CVAE-GAN is:

[0015]

[0016] Among them, L CVAE-GAN is the overall error value of the CAVE-GAN model, and L KL is the reconstruction loss and KL divergence, L G is the loss function of the generated network G.

[0017] Furthermore, the loss function L of the generated network G is G for:

[0018]

[0019] Where Z is the noise, G(Z i ,Y i ) is the structural parameter predicted by the generative network G, D(G(Z i ,Y i ),Y i ) is the discriminant probability value of the discriminant network D for the predicted structural parameters.

[0020] Furthermore, the reconstruction loss and KL divergence L KL for:

[0021]

[0022] Among them, σ is the standard deviation, μ is the mean, To predict structural parameters.

[0023] Furthermore, the training using MEMS sensor experimental data in step S3 specifically includes:

[0024] S3-1, the structural parameters X and performance parameters Y in the MEMS sensor experimental data are input into the encoding network E, which maps the input data to the Gaussian distribution in the latent space, extracts the features of the data, and obtains the mean μ and variance σ; noise Z is generated by sampling from μ and σ;

[0025] S3-2, input the noise Z and performance parameter Y into the generation network G to generate the predicted structural parameters

[0026] S3-3, the performance parameter Y is compared with the predicted structural parameter of the output of the generative network G Combined, used as an input parameter to determine whether the network D label is false; Combined with the structural parameter X and the performance parameter Y in the MEMS sensor experimental data, used as an input parameter to determine whether the network D label is true;

[0027] S3-3, repeat steps S3-1 to S3-3, so that the encoding network E searches for the distribution of input data, and the discriminant network D sequentially receives pseudo data from the generating network G and real data of the training set for iterative training; through the back propagation algorithm, the difference between the original data and the reconstructed data is minimized, so that the encoding network E can map the input data to the Gaussian distribution in the latent space, and the generating network G can reach the expected prediction ability.

[0028] Furthermore, the G-MS model in step S4 is a multi-solution generation network, which is composed of the verification network trained in step S2 and the generation network G trained in step S3. By updating the noise Z, the generation network G outputs multiple solutions that meet the expected performance Y. * result.

[0029] Furthermore, by updating the noise Z, the generator network G outputs multiple outputs that meet the expected performance Y * The results include:

[0030] Step S4-1, initialize the noise Z0 by Gaussian noise; set the expected performance parameter Y * , number of iterations I0, learning rate η and equalization factor γ; (Z0,Y * ) is input into the pre-trained generative network G to obtain X(1) ; Establish multiple solution sets And X (1) deposited therein;

[0031] Step S4-2, initialize the minimum loss function L min is 9999, and the current noise Z is synthesized by Gaussian noise c ; Z c Assign the value to the noise Z corresponding to the minimum loss m ;

[0032] Step S4-3, (Z0, Y * ) is input into the generative network G to obtain the current solution X (c) , and calculate X (c) With multiple solution sets The discrimination d of the solutions in c :

[0033]

[0034] Where d is X (c) With {X (1) ,X (2) ,…,X (M)}The average of the sum of Euclidean distances, M is the number of old solutions, M>1;

[0035] Step S4-4, X (c) Input into the validation network trained in step S2 to obtain And the current loss is calculated by the following formula:

[0036]

[0037] Among them, L c is the current loss, γ is the balancing factor for controlling weights; ε is used to avoid division by zero; if L c <L min , then L min The value is L c , Z m The value is Z c ;

[0038] Step S4-5, update Z using the gradient descent strategy c Make L c To reach minimum:

[0039]

[0040] Step S4-6, repeat steps S4-3 to S4-510-1 times to determine Z m , (Z m ,Y *) is input into the generative network G to obtain X (i) ; X (i) Store multiple solution sets middle;

[0041] Step S4-7, repeat steps S4-3 to S4-6 i times, and finally get i+1 solutions.

[0042] A MEMS sensor intelligent design system based on a generative network, comprising:

[0043] Data acquisition unit, which obtains MEMS sensor experimental data through finite element simulation;

[0044] Verification network construction training unit, used to build a verification network to achieve forward design from sensor structure to performance parameters, and use MEMS sensor experimental data for training;

[0045] Generative network CVAE-GAN construction training unit is used to construct the generative network CVAE-GAN to achieve reverse design from performance to structural parameters, and use MEMS sensor experimental data for training;

[0046] The multi-solution generator G-MS model construction unit builds the multi-solution generator G-MS model based on the verification network and the generative network CVAE-GAN, and generates multiple sets of structural parameters that meet performance requirements.

[0047] Compared with the prior art, the present invention has the following significant advantages:

[0048] (1) Extremely high design efficiency: Based on the CVAE-GAN model, the present invention establishes an end-to-end intelligent design model from MEMS sensor performance parameters to structural parameters, avoiding frequent finite element simulation and manual parameter adjustment processes, effectively shortening the design cycle, and having the advantages of fast speed, high precision and intelligence;

[0049] (2) High-quality different design schemes: Based on the noise updating strategy, the present invention establishes a multi-solution generator G-MS, which can quickly and high-quality generate a variety of different design schemes that meet given performance requirements, help shorten the design cycle, achieve rapid iterative design, and provide easy-to-process, low-cost alternative design schemes. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 is a flow chart of a MEMS sensor intelligent design method based on a generative network provided in an embodiment of the present invention;

[0052] Figure 2 It is a schematic diagram of the CVAE-GAN model and the G-MS model provided in an embodiment of the present invention.

[0053] Figure 3 It is a structural diagram of a verification network model provided by an embodiment of the present invention.

[0054] Figure 4 It is a structural diagram of the CVAE-GAN model provided by an embodiment of the present invention.

[0055] Figure 5 It is a flow chart of the G-MS model algorithm provided by an embodiment of the present invention.

[0056] Figure 6 It is a comparison diagram between the embodiment of the present invention and the traditional multi-solution generation scheme. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with its embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0058] This embodiment proposes a MEMS sensor intelligent design method based on a generative network. Figure 1 and Figure 2 , the specific steps are as follows:

[0059] Step S1: Acquire MEMS sensor experimental data based on finite element simulation.

[0060] 2000 sets of piezoresistive accelerometer experimental data were obtained by finite element simulation method, and their descriptive statistics are shown in Table 1. The experimental data include structural parameters: side length L of the mass block m , thickness T m and the length L of the cantilever beam b , Width L w , Thick T-shirt b And performance parameters: the sensor's range R, sensitivity K and resonant frequency f, can be expressed as (x1, x2, x3, x4, x5) and (y1, y2, y3) respectively. Among them, the structural parameter L m , T m , L b , L w and T bThe values ​​of are random integers in the ranges of 800 to 1000 μm, 150 to 300 μm, 20 to 50 μm, 2500 to 3000 μm and 400 to 500 μm respectively.

[0061] Preferably, in order to prevent the weight parameter ω from changing drastically during the training process of the neural network and to accelerate the convergence of the model, the experimental data can be normalized, and the normalization calculation method is as follows:

[0062]

[0063] where x norm is the normalized sensor parameter value, x is the original data, and x max is the maximum value of each characteristic parameter in the original data, x min is the minimum value of each characteristic parameter in the original data.

[0064] The experimental data is divided into training set, test set and validation set in the ratio of 6:2:2.

[0065] Step S2: Construct and train a verification network to implement forward design from sensor structure to performance parameters.

[0066] Take the sensor structure parameters X as input and predict the performance parameters As output, the mean square error loss function is used for error calculation, and its formula is as follows:

[0067]

[0068] Among them, L V To verify the error value of the network, N is the capacity of the data set, and Y is the actual performance parameter. It is the performance parameter for verifying network prediction.

[0069] Specifically, Figure 3 As shown in the figure, the input dimension of the verification network is 5 and the output dimension is 3. The structure of the verification network is determined by the ten-fold cross validation method, and the optimal neuron combination of 10-90-20-30 is finally obtained. At the same time, it is determined that the network performance is best when the batch size Batch Size value is 10. After the structure and parameters of the verification network are finally determined, it is fully trained. The average relative errors of the indicators (R, K, f) are 2.00%, 0.79% and 0.39% respectively when evaluated on the test data set.

[0070] Step S3: Construct and train the generative network CVAE-GAN to achieve reverse design from performance to structural parameters.

[0071] like Figure 4, the CVAE-GAN network model consists of an encoding network E, a generation network G, and a discriminant network D, as well as a network for calculating the reconstruction loss Verification network V; the number of input units of network E is n+m, and the number of output nodes is 2; the dimension of noise Z is also n, so the number of input units of network G is also n+m, and the number of output units is n. The number of input units of network D is n, and the output is a binary value (0 / 1) representing true or false.

[0072] The overall error calculation method of the CAVE-GAN model is as follows:

[0073]

[0074] Among them, LCVAE-GAN is the overall error value of the CAVE-GAN model, and LKL is the reconstruction loss and KL divergence, L G To generate the loss of network G, the calculation methods are as follows:

[0075]

[0076] Among them, σ is the standard deviation, μ is the mean, is the predicted structural parameters, Z is the noise, G(Z i ,Y i ) is the structural parameter predicted by the network G,

[0077] The CVAE-GAN model in step S3 simultaneously trains network E, network G, and network D based on the data set in step S1. The training process is as follows:

[0078] S3-1, the structural parameters X and performance parameters Y in the training set data are input into the network E, the network E maps the input data to the Gaussian distribution in the latent space, extracts the features of the data, and obtains the mean μ and variance σ; noise Z is generated by sampling from μ and σ;

[0079] S3-2, input the noise Z and performance parameter Y into the network G to generate the predicted structural parameters

[0080] S3-3, the performance parameter Y is compared with the predicted structural parameter of the output of the network G Combined, used as input parameters for network D label being false; Combined with the structural parameter X and performance parameter Y in the training set data as input parameters for judging whether the network D label is true;

[0081] S3-4, repeat steps S3-1 to S3-3, so that network E searches for the distribution of input data, and network D sequentially receives pseudo data from network G and real data of the training set for iterative training; through the back propagation algorithm, the difference between the original data and the reconstructed data is minimized, so that network E can map the input data to the Gaussian distribution in the latent space, and network G can achieve the expected prediction ability.

[0082] Furthermore, the ten-fold cross validation method was used to determine the structures of network E, network G, and network D, and the optimal neuron combinations of network E, network G, and network D were finally obtained as 10-90-20-30, 30-30-70, and 100-100-100-100, respectively. In addition, the value of Batch Size was determined to be 10.

[0083] After all parameters are determined, the CVAE-GAN model is retrained with a larger number of epochs, and the test loss finally reaches the minimum value of 2.08×10 at 8712 epochs. -5 , the corresponding training loss is 2.41×10 -5 The average relative errors of the indicators (R, K, f) are 1.66%, 1.76%, and 0.63%, respectively.

[0084] Step S4: Construct a multi-solution generator G-MS model to simultaneously generate multiple sets of structural parameters that meet performance requirements.

[0085] It consists of the verification network V trained in step S2 and the generation network G trained in step S3, which updates the noise Z so that the network G outputs multiple values ​​that meet the expected performance Y * Results: The G-MS model calculates the newly generated solution X by the following formula (i) With the old solution {X (1) ,X (2) ,…,X (M) Distinction of}:

[0086]

[0087] Where d is X (i) With {X (1) ,X (2) ,…,X (M)}The average of the sum of Euclidean distances, M (M>1) is the number of old solutions.

[0088] like Figure 5 As shown, the G-MS model is a multi-solution generation network that updates the noise Z through the following process:

[0089] S5-1, initialize noise Z0 through Gaussian noise; set target performance parameter Y *, number of iterations I0, learning rate η and equalization factor γ; (Z0,Y * ) is input into the pre-trained network G to obtain X (1) ; Establish multiple solution sets And X (1) deposited therein;

[0090] S5-2, initialize the minimum loss function L min is 9999, and the current noise Z is synthesized by Gaussian noise c ; Z c Assign the value to the noise Z corresponding to the minimum loss m ;

[0091] S5-3, (Z0,Y * ) is input into the network G to get the current solution X (c) , and calculate X (c) With multiple solution sets The discrimination d of the solutions in c ;

[0092] S5-4, X (c) Input into the verification network V to get And the current loss is calculated by the following formula:

[0093]

[0094] Among them, L c is the current loss, γ is the balancing factor for controlling weights, m is the dimension of Y; ε is used to avoid division by zero and its value is 10 -8 ; If L c <L min , then L min The value is L c , Z m The value is Z c ;

[0095] S5-5, update Z using the gradient descent strategy c Make L c To achieve the minimum, the implementation principle is shown in the following formula:

[0096]

[0097] S5-6, repeat steps S5-3 to S5-5 10-1 times to determine Z m , (Z m ,Y * ) is input into the network G to get X (i) ; X (i) Store multiple solution sets middle;

[0098] S5-7, repeat steps S5-3 to S5-6 i times, and finally get i+1 solutions.

[0099] To explore the impact of the learning rate η on the G-MS model, perform the following operations:

[0100] (1) Generate 2000 sets of data through the pre-trained verification network Treat X as The first solution X (1) ; Then each As the target performance parameter Y * Input into G-MS, then each group of Y * We get the second solution X (2) .

[0101] (2) Adjust η in the range of 0.01 to 10 and record the 2000 groups of Y generated by G-MS * Corresponding X (2) The average number of iterations required The standard deviation std of the average Euclidean distance d and the average value of d And X (2) Input into the network to get calculate With Y * The mean square error (MSE) of .

[0102] As η increases, After a slow decline, it stabilizes. When η>1, MES gradually increases and std also increases sharply. Taking various factors into consideration, η is taken as 0.5.

[0103] Furthermore, a threshold distance d may be set T , d T When <0.15, it can be considered that no new solution has been found.

[0104] The advantages of the present invention are further verified by comparison with the traditional method.

[0105] To design a piezoresistive accelerometer by traversal, the 5-dimensional structural parameter X needs to be traversed at intervals of 0.01, and 10 attempts are required to generate each set of data. 10 times, which takes about 10 4 hours, which is highly impractical, and creating large datasets and clustering them to find different solutions is relatively expensive.

[0106] Figure 6 100,000 sets of structural parameters X are generated for each set of performance parameters Y, and the value with the largest average Euclidean distance d is identified; when d <dT It is recorded as no new solution found. Figure 6 As shown in the figure, using the traditional method, when given performance requirements, 58.70% of the requirements failed to find the second solution even after 100,000 iterations, and the remaining requirements required an average of 78,289 iterations, a total of 13.09±5.34s, and d was only 0.10±0.14; in contrast, using the G-MS model, 99.65% of the requirements can find the second solution within 100,000 iterations, with an average of 4,709 iterations, 3.60±2.46s, and d of 0.75±0.19, which are 3.63 times and 7.19 times higher than the traditional method, respectively.

[0107] A MEMS sensor intelligent design system based on a generative network, comprising:

[0108] Data acquisition unit, which obtains MEMS sensor experimental data through finite element simulation;

[0109] Verification network construction training unit, used to build a verification network to achieve forward design from sensor structure to performance parameters, and use MEMS sensor experimental data for training;

[0110] Generative network CVAE-GAN construction training unit is used to construct the generative network CVAE-GAN to achieve reverse design from performance to structural parameters, and use MEMS sensor experimental data for training;

[0111] The multi-solution generator G-MS model construction unit builds the multi-solution generator G-MS model based on the verification network and the generative network CVAE-GAN, and generates multiple sets of structural parameters that meet performance requirements.

[0112] The CVAE-GAN model of the present invention combines the advantages of the CVAE model and the CGAN model. The structured latent space of CVAE promotes the diversity of samples, and uses the L2 distance between real data and predicted data as the loss function, which effectively enhances the ability of the generative network G to generate high-quality data and reduces the risk of mode collapse of the CGAN model to a certain extent. The adversarial loss of CGAN can continuously narrow the difference between the feature distribution of pseudo data output by the generative network G and the feature distribution of real data.

[0113] The present invention is based on a generative model and uses the performance parameters of a piezoresistive accelerometer as characteristic values ​​to predict structural parameters. Compared with the finite element simulation method, this method avoids complex multi-physics field calculations and shortens the design cycle of researchers. At the same time, compared with traditional machine learning, the CVAE-GAN model in this method has the ability to solve non-uniqueness problems, and the G-MS model can generate a variety of feasible solutions with high discrimination.

[0114] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for intelligent design of MEMS sensors based on generative networks, characterized in that: include: Step S1: Acquire MEMS sensor experimental data through finite element simulation; Step S2: construct a verification network to realize the forward design from sensor structure to performance parameters, and use MEMS sensor experimental data for training; Step S3: construct a generative network CVAE-GAN to achieve reverse design from performance to structural parameters, and use MEMS sensor experimental data for training; Step S4: Based on the verification network and the generative network CVAE-GAN, a multi-solution generator G-MS model is constructed to generate multiple sets of structural parameters that meet performance requirements.

2. The MEMS sensor intelligent design method based on generative network according to claim 1 is characterized in that: The MEMS sensor experimental data in step S1 includes structural parameters X and performance parameters Y; wherein the structural parameters X include continuous variables describing the sensor geometry with a dimension of m, and Y includes continuous variables describing the sensor performance with a dimension of n; the verification network in step S2 is an artificial neural network with an input dimension of n and an output dimension of m, and the mean square error is used to count and characterize the error level of the network.

3. The MEMS sensor intelligent design method based on generative network according to claim 1 is characterized in that: In step S3, the generative network CVAE-GAN includes an encoding network E, a generating network G and a discriminative network D; the number of input units of the encoding network E is n+m, and the number of output nodes is 2; the number of input units of the generating network G is the dimension m of the performance parameter plus the dimension of the noise Z, and the number of output units is n; the number of input units of the discriminative network D is n+m, the output node is 1, and the output unit is activated using the Sigmoid function.

4. The method for intelligent design of MEMS sensors based on generative networks according to claim 3 is characterized in that: The total loss function of the generative network CVAE-GAN is: Among them, L CVAE-GAN is the overall error value of the CAVE-GAN model, and L KL is the reconstruction loss and KL divergence, L G is the loss function of the generated network G.

5. The method for intelligent design of MEMS sensors based on generative networks according to claim 4, characterized in that: The loss function L of the generated network G G for: Where Z is the noise, G(Z i ,Y i ) is the structural parameter predicted by the generative network G, D(G(Z i ,Y i ),Y i ) is the discriminant probability value of the discriminant network D for the predicted structural parameters.

6. The MEMS sensor intelligent design method based on generative network according to claim 4 is characterized in that: The reconstruction loss and KL divergence L KL for: Among them, σ is the standard deviation, μ is the mean, To predict structural parameters.

7. The MEMS sensor intelligent design method based on generative network according to claim 3 is characterized in that: The training using MEMS sensor experimental data in step S3 specifically includes: S3-1, the structural parameters X and performance parameters Y in the MEMS sensor experimental data are input into the encoding network E, which maps the input data to the Gaussian distribution in the latent space, extracts the features of the data, and obtains the mean μ and variance σ; noise Z is generated by sampling from μ and σ; S3-2, input the noise Z and performance parameter Y into the generation network G to generate the predicted structural parameters S3-3, the performance parameter Y is compared with the predicted structural parameter of the output of the generative network G Combined, used as an input parameter to determine whether the network D label is false; Combined with the structural parameter X and the performance parameter Y in the MEMS sensor experimental data, used as an input parameter to determine whether the network D label is true; S3-3, repeat steps S3-1 to S3-3, so that the encoding network E searches for the distribution of input data, and the discriminant network D sequentially receives pseudo data from the generating network G and real data of the training set for iterative training; through the back propagation algorithm, the difference between the original data and the reconstructed data is minimized, so that the encoding network E can map the input data to the Gaussian distribution in the latent space, and the generating network G can reach the expected prediction ability.

8. The method for intelligent design of MEMS sensors based on generative networks according to claim 7, characterized in that: The G-MS model in step S4 is a multi-solution generation network, which is composed of the verification network trained in step S2 and the generation network G trained in step S3. By updating the noise Z, the generation network G outputs multiple solutions that meet the expected performance Y. * result.

9. The method for intelligent design of MEMS sensors based on generative networks according to claim 8, characterized in that: By updating the noise Z, the generated network G outputs multiple Y that meet the expected performance * The results include: Step S4-1, initialize the noise Z0 by Gaussian noise; set the expected performance parameter Y * , number of iterations I0, learning rate η and equalization factor γ; (Z0,Y * ) is input into the pre-trained generative network G to obtain X (1) ; Establish multiple solution sets And X (1) deposited therein; Step S4-2, initialize the minimum loss function L min is 9999, and the current noise Z is synthesized by Gaussian noise c ; Z c Assign the value to the noise Z corresponding to the minimum loss m ; Step S4-3, (Z0, Y * ) is input into the generative network G to obtain the current solution X (c) , and calculate X (c) With multiple solution sets The discrimination d of the solutions in c : Where d is X (c) With {X (1) ,X (2) ,…,X (M) }The average of the sum of Euclidean distances, M is the number of old solutions, M>1; Step S4-4, X (c) Input into the validation network trained in step S2 to obtain And the current loss is calculated by the following formula: Among them, L c is the current loss, γ is the balancing factor for controlling weights; ε is used to avoid division by zero; if L c <L min , then L min The value is L c , Z m The value is Z c ; Step S4-5, update Z using the gradient descent strategy c Make L c To reach minimum: Step S4-6, repeat steps S4-3 to S4-510-1 times to determine Z m , (Z m ,Y * ) is input into the generative network G to obtain X (i) ; X (i) Store multiple solution sets middle; Step S4-7, repeat steps S4-3 to S4-6 i times, and finally get i+1 solutions.

10. A MEMS sensor intelligent design system based on generative networks for implementing any of the methods described in claims 1-9, characterized in that: include: Data acquisition unit, which obtains MEMS sensor experimental data through finite element simulation; Verification network construction training unit, used to build a verification network to achieve forward design from sensor structure to performance parameters, and use MEMS sensor experimental data for training; Generative network CVAE-GAN construction training unit is used to construct the generative network CVAE-GAN to achieve reverse design from performance to structural parameters, and use MEMS sensor experimental data for training; The multi-solution generator G-MS model construction unit builds the multi-solution generator G-MS model based on the verification network and the generative network CVAE-GAN, and generates multiple sets of structural parameters that meet performance requirements.

Images