A piezoelectric metamaterial self-adaptive vibration suppression method based on deep learning reverse design
By reverse-engineering nonlinear piezoelectric metamaterials using deep learning, adaptive vibration control was achieved, solving the problems of complexity and lack of flexibility in traditional designs. This improved vibration suppression and design efficiency, making it suitable for vibration control of precision mechanical devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2025-04-28
- Publication Date
- 2026-07-07
AI Technical Summary
Traditional piezoelectric metamaterials are complex to design and lack flexibility, making it difficult to efficiently achieve the desired vibration suppression performance, and the design process relies on prior knowledge and experience.
A nonlinear piezoelectric metamaterial is constructed using a deep learning-based reverse design method. By acquiring circuit parameters and vibration response signal datasets, a neural network model is established to achieve reverse design of the vibration suppression characteristics of the piezoelectric metamaterial. Adaptive vibration control is achieved by adjusting the shunt circuit parameters using a microcontroller.
It significantly improves the flexibility and adaptability of vibration suppression, expands the vibration suppression range by 1.36 times, improves design efficiency, achieves a prediction accuracy of 98.916%, and reduces reliance on prior knowledge and debugging experiment time.
Smart Images
Figure CN120473040B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vibration control technology, specifically to an adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design. Background Technology
[0002] In applications requiring high precision, such as precision machining and measurement, space remote sensing and observation, mechanical devices may be subject to vibrations from the external environment or internal components, leading to a significant decrease in the precision and accuracy of the devices. To ensure the normal operation of the structure, effective vibration control is necessary. Common vibration suppression methods include passive isolation, semi-active vibration suppression, and active vibration suppression. Semi-active vibration suppression technology based on the bandgap characteristics of piezoelectric metamaterials has attracted considerable research attention due to its greater flexibility and lower power consumption.
[0003] Traditional piezoelectric metamaterials are typically based on analog shunt circuits. Analog circuit designs are complex, and modifying circuit parameters after completion is inconvenient and inflexible. Furthermore, designers need prior knowledge and experience to ensure the effectiveness of the design. Therefore, how to design piezoelectric metamaterials more efficiently and conveniently to achieve the desired vibration damping performance has become a pressing problem for those skilled in the art. Summary of the Invention
[0004] This invention addresses the technical problems existing in the prior art by utilizing the bandgap characteristics of piezoelectric metamaterials to provide an adaptive vibration control method for piezoelectric metamaterials based on deep learning reverse design. By acquiring the vibration response signals of nonlinear piezoelectric metamaterials under different circuit parameters, a neural network model corresponding to the circuit parameters and response signals is constructed to achieve reverse design of the vibration suppression characteristics of nonlinear piezoelectric metamaterials and to suppress structural vibration.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive vibration control method for piezoelectric metamaterials based on deep learning reverse design, comprising: constructing a nonlinear piezoelectric metamaterial; acquiring a dataset of circuit parameters and vibration response signals of the nonlinear digital piezoelectric metamaterial; the dataset of circuit parameters and vibration response signals including a circuit parameter vector and a bandgap label vector of the nonlinear digital piezoelectric metamaterial; constructing a deep learning-based reverse design network model; training the deep learning-based reverse design network model according to the circuit parameter and vibration response signal dataset; and acquiring the circuit parameters of the nonlinear digital piezoelectric metamaterial for vibration suppression based on the trained deep learning-based reverse design network model.
[0006] As an improvement of the present invention, the nonlinear digital piezoelectric metamaterial includes: local resonant units, an excitation signal sensor, and a response signal sensor. The local resonant units are arranged periodically, and two or more one-dimensional periodically arranged local resonant units form multiple vibration control channels. The two or more vibration control channels and their parameters are adjusted to attenuate the vibration waves of the metamaterial at a specified frequency. It should be noted that when the number of integrated local resonant units increases, the vibration suppression capability of the nonlinear digital piezoelectric metamaterial will be further enhanced. In the embodiment constructed by the present invention, the nonlinear digital piezoelectric metamaterial can achieve a vibration wave attenuation effect of up to 37.75 dB.
[0007] As an improvement of the present invention, the local resonant unit includes a piezoelectric transducer and a nonlinear shunt circuit. The local resonant units are periodically arranged on the substrate, and the piezoelectric transducers in the local resonant units are attached to one side surface of the substrate. This helps to simplify the structural layout, reduce mechanical coupling interference that may be caused by double-sided arrangement, thereby improving the stability and consistency of the system. At the same time, this single-sided arrangement can achieve directional coupling control of the structural vibration direction, enhance the ability to regulate specific modal vibrations, and further improve vibration suppression performance.
[0008] As an improvement of the present invention, the piezoelectric transducers in the local resonant unit are connected to a nonlinear shunt circuit, which includes a digital part and an analog part. The analog part is built using operational amplifiers, capacitors, and resistors, while the digital part is built using a microcontroller. The analog and digital parts are connected using a digital-to-analog converter module. The upper and lower surfaces of the piezoelectric transducers are connected to the output port of the nonlinear shunt circuit and the ground wire, respectively. Through the above structural design, the nonlinear term is effectively integrated into the piezoelectric metamaterial system. Utilizing the programmable characteristics of the microcontroller, the parameters of the nonlinear term can be flexibly adjusted, thereby giving the system a higher nonlinear response control capability. After adding the nonlinear term, the piezoelectric metamaterial exhibits characteristics such as a broadened vibration suppression frequency band, a shift in the vibration transmission rate curve jump point, and a change in the vibration wave attenuation degree in experiments, significantly enhancing the flexibility and adaptability of the vibration suppression system.
[0009] As an improvement of this invention, the digital part of the nonlinear shunt circuit is implemented using a microcontroller. The digital part transforms the nonlinear circuit into a transfer function in the s-domain and further discretizes it into a difference equation, which is:
[0010]
[0011] Where y[n] is the output value at the current time, y[nj] is the output value at the previous j sampling times, x[n] is the input value at the current time, x[ni] is the input value at the previous i sampling times, m is the order of the numerator, n is the order of the denominator, and b i a is the numerator coefficient. j K1 represents the coefficient of the denominator, K2 represents the coefficient of the nonlinear quadratic term, and K3 represents the coefficient of the nonlinear cubic term. Since the analysis method in the s-domain is difficult to directly handle nonlinear terms, the above design allows the nonlinear control strategy to be written into the microcontroller in differential form and implemented through programming. This effectively integrates the strategy into the piezoelectric metamaterial system, improving the adjustability and response flexibility of the vibration control process.
[0012] As an improvement of the present invention, the piezoelectric transducer on the substrate is any one of the following shapes: circular, elliptical, rectangular, rhomboid, triangular, or hexagonal.
[0013] As an improvement of the present invention, the shape of the substrate in the local resonant unit is any one of rectangle, square or hexagon, and the substrate material is metal.
[0014] As an improvement of the present invention, the number of vibration control channels is two or more.
[0015] As an improvement of the present invention, the number of local resonant units in each vibration control channel is one.
[0016] As an improvement of this invention, the deep learning-based reverse engineering network model consists of two network models: a pre-trained network and a reverse engineering network. The circuit parameter vector is used as the input to the pre-trained network, and the bandgap label vector is used as its output. Similarly, the bandgap label vector is used as the input to the reverse engineering network, and the circuit parameter vector is used as its output. Based on this, the reverse engineering network is used for reverse engineering from vibration response to circuit parameters, while the pre-trained network is used to verify whether the predicted circuit parameters can produce the corresponding vibration response, thus constructing a network model that integrates design and verification functions. This approach not only improves the accuracy and reliability of reverse engineering but also enables the network model to self-verify the design results.
[0017] As an improvement of this invention, the training of the deep learning-based serial pipeline model includes: training the pre-trained network and the reverse engineering network separately using a pipeline training method, the process of which is divided into the following two stages. In the first stage, the pre-trained network is trained to realize the mapping from the circuit parameters of the nonlinear piezoelectric metamaterial to the bandgap characteristics; in the second stage, the reverse engineering network and the pre-trained network are constructed into a serial pipeline. First, the reverse engineering network completes the mapping from the bandgap characteristics to the circuit parameters of the nonlinear piezoelectric metamaterial, and then the output of the reverse engineering network is used as the input of the pre-trained network to generate new target characteristics. The reverse engineering network is a Tran-conv network combining a 1DCNN neural network and a Transformer network; the Adam optimizer is used during backpropagation, with a learning rate of 0.0001.
[0018] The loss function of this concatenated network is:
[0019]
[0020] Where L is the data length, To reverse engineer the network's predictions, Let λ be the predicted value of the pre-trained network and λ be the weights assigned to the reverse-engineered network. This multi-objective loss function design helps improve the stability and prediction accuracy of the network in practical applications. In the embodiment constructed in this invention, the reverse-engineered network achieves an accuracy of up to 98.916% on the test set.
[0021] Compared with existing technologies, the beneficial effects of this invention are as follows: The nonlinear piezoelectric metamaterial provided by this invention can achieve adaptive vibration control. By implementing the shunt circuit using a microcontroller, the parameters of the linear part in the shunt circuit can be flexibly adjusted, thereby adjusting the frequency range of vibration suppression, etc., adding programmability to the metamaterial and significantly improving the flexibility of vibration suppression. Furthermore, by adjusting the coefficients of the nonlinear terms in the digital shunt circuit to control the nonlinear response of the metamaterial, a wider range of vibration suppression effects can be provided than linear piezoelectric metamaterials. The jump points of vibration transmission rate and the degree of vibration wave attenuation can also be flexibly adjusted, further improving the flexibility and adaptability of the metamaterial. In the embodiment constructed by this invention, the vibration suppression range can be widened by up to 1.36 times, demonstrating the superior performance of this metamaterial in vibration control. The neural network model established by this invention, in which the pre-trained network can replace the finite element method to accurately predict the structural vibration response, and the reverse-engineered network can accurately predict the circuit parameters of the desired response. In the embodiment constructed by this invention, the prediction accuracy can reach 98.916%. This method significantly improves the efficiency of the reverse-engineering process of metamaterial structural parameters. Compared to traditional design methods, this approach reduces reliance on prior knowledge and the time spent on debugging experiments, making metamaterial design simpler and more intelligent. The adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design provided by this invention combines nonlinear digital piezoelectric metamaterials with a deep learning reverse design model. This not only achieves effective vibration suppression but also improves design efficiency, realizing efficient and accurate intelligent reverse design. It enhances the application value of nonlinear piezoelectric metamaterials under complex working conditions and provides a new approach for constructing intelligent vibration control systems. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the nonlinear piezoelectric metamaterial structure of the present invention;
[0023] Figure 2 This is a schematic diagram of the nonlinear piezoelectric metamaterial single piezoelectric transducer and its programmable shunt circuit of the present invention;
[0024] Figure 3 This is a structural diagram of the deep learning-based reverse design network pipeline training model of the present invention;
[0025] Figure 4 An iterative graph of the loss function during the pipeline training network process;
[0026] Figure 5 This is a diagram illustrating the vibration suppression effect of the reverse design of a nonlinear piezoelectric metamaterial.
[0027] In the figure: 1, 2, and 3 are piezoelectric transducers; 4 is a local resonant unit; 5 is an operational amplifier; 6 is a resistor; 7 is an analog-to-digital converter module; 8 is a microcontroller; 9 is a digital-to-analog converter module; 10 is the target bandgap label vector; 11 is a convolutional block; 12 is a fully connected layer; 13 is the prediction circuit parameter vector; 14 is a Transformer block; 15 is the predicted bandgap label vector; 16 is the loss value; 17 is the training set loss value; 18 is the test set loss value; and 19 is the number of iterations. Detailed Implementation
[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and modifications without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.
[0029] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0030] To illustrate the technical solution described in this application, specific embodiments will be described below.
[0031] Example: This example provides a vibration-adaptive suppression method for piezoelectric metamaterials based on deep learning reverse design, including the following steps:
[0032] Step 1: Construct a nonlinear piezoelectric metamaterial. For example... Figure 1 As shown, a monolithic square piezoelectric transducer is periodically fixed to one side surface of the substrate using epoxy resin adhesive, and the piezoelectric transducer is connected to a nonlinear shunt circuit via wires. Figure 2 As shown, the piezoelectric transducer is connected to the analog section of the nonlinear shunt circuit via wires, and a digital shunt circuit is implemented through microcontroller programming. This allows for real-time interaction between the digital-to-analog converter (DAC) and analog-to-digital converter (ADC) with the analog circuit signals. For example... Figure 1 As shown, an excitation signal is applied through piezoelectric transducer 1, and a response signal is obtained through piezoelectric transducer 3.
[0033] Step 2: Obtain the circuit parameters and response signal dataset of the piezoelectric metamaterial. This dataset includes the nonlinear circuit parameter vector and bandgap label vector of the piezoelectric metamaterial. Using methods such as... Figure 1 The vibration response of the piezoelectric metamaterial shown is used as sample data for design. Its circuit parameters are represented by three parameters x1, x2, and x3, which respectively represent the coefficients and frequencies of the quadratic and cubic terms, and are then normalized.
[0034] The response signal corresponding to the circuit parameters is divided into target frequency ranges [f] min ,f max The data is divided into N segments, where N = 2000. Then, a corresponding label vector y = (y1, y2, ..., y...) is created. N This involves assigning values to a label vector to indicate whether the vibration wave is suppressed. When y i =1 means the frequency f i It is allowed to bring anything else; otherwise, it is prohibited.
[0035] Step 3: Construct the network model. In this example, to address the issue of piezoelectric metamaterials exhibiting identical response signals for the same circuit parameters, the overall network consists of two parts: a pre-trained network and a reverse-engineered network. The pre-trained network takes the circuit parameter vector as input and outputs the bandgap label vector. The reverse-engineered network takes the bandgap label vector as input and outputs the circuit parameter vector.
[0036] Step 4: Model Training. This example uses a pipeline training method to train the pre-trained network and the reverse-engineered network separately. The process is divided into two stages: First, the pre-trained network is trained to map circuit parameters to response signals; second, the reverse-engineered network and the pre-trained network are connected in series to form a pipeline-like model. First, the response signal is input into the reverse-engineered network to obtain the predicted circuit parameters. Then, the predicted circuit parameters are input into the trained pre-trained network to obtain the predicted response signal. The Adam optimizer is selected during network training, with a learning rate of 0.0001. The loss function of this connected network is:
[0037]
[0038] Where L is the sample size, To reverse engineer the network's predictions, λ represents the predicted values of the pre-trained network, and λ represents the weights assigned to the reverse-engineered network.
[0039] The dataset obtained in step 2 is divided into a training set and a test set in a 9:1 ratio to train the network. Figure 4 The graph shows the iterative process of the loss function of the concatenated network during training. It can be seen that the network's loss function gradually decreases until it converges, and the training set and test set are well matched, indicating that the network training is in good condition.
[0040] Step 5: After training the cascaded network in Step 4, input the target bandgap vector into the trained reverse engineering network to obtain the circuit parameters that satisfy the design. Then, use an in-circuit programming microcontroller to obtain the actual response of the piezoelectric metamaterial under this circuit. Compare the actual bandgap vector of the response signal with the target bandgap vector to verify the accuracy of the neural network. A specific example is provided. Figure 5 As shown, the gray area represents the suppressed vibration wave within this frequency band. It can be seen that the target suppression range and the actual suppression range are basically consistent, indicating that the reverse-engineered network performs well in predicting the bandgap.
[0041] The piezoelectric metamaterial designed in this invention utilizes the structural vibration response to a nonlinear shunt circuit to effectively achieve adaptive vibration suppression. Simultaneously, the deep learning-based reverse engineering model established in this invention can accurately achieve automatic reverse engineering of circuit parameters, simplifying the design process and improving design efficiency.
[0042] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. An adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design, characterized in that, include: Acquiring data sets of circuit parameters and vibration response signals of nonlinear digital piezoelectric metamaterials; The circuit parameter and vibration response signal dataset includes circuit parameter vectors and bandgap label vectors of nonlinear digital piezoelectric metamaterials; a deep learning-based reverse design network model is constructed; and the deep learning-based reverse design network model is trained based on the circuit parameter and vibration response signal dataset. The circuit parameters of the nonlinear digital piezoelectric metamaterial are obtained from the trained deep learning-based reverse engineering network model. In the local resonant unit, the piezoelectric transducers are connected to the nonlinear shunt circuit, which includes a digital part and an analog part. The analog part is built using operational amplifiers, capacitors and resistors, while the digital part is built using a microcontroller. The analog and digital parts are connected using a digital-to-analog converter module. The upper and lower surfaces of the piezoelectric transducers are connected to the output port of the nonlinear shunt circuit and the ground wire, respectively. The digital part of the nonlinear shunt circuit is implemented using a microcontroller, which converts the nonlinear circuit into... The transfer function of the domain is further discretized into a difference equation, which is: in This is the output value at the current moment. For the front The output value at each sampling time. This is the input value at the current moment. For the front The input value at each sampling time point, The molecular order is... The order of the denominator is . The numerator coefficient, The coefficient of the denominator is . The coefficients of the nonlinear quadratic term are denoted as . The coefficients of the nonlinear cubic term; The deep learning-based reverse design network model consists of two network models: a pre-trained network and a reverse design network. The circuit parameter vector is used as the input of the pre-trained network, and the bandgap label vector is used as the output of the pre-trained network. The bandgap label vector is used as the input to the reverse design network, and the circuit parameter vector is used as the output of the reverse design network. Training a deep learning-based cascaded pipeline model involves training both a pre-trained network and a reverse-engineered network using a pipeline training method. The process is divided into two stages: First, the pre-trained network is trained to map the circuit parameters of the nonlinear piezoelectric metamaterial to its bandgap characteristics. Second, the reverse-engineered network and the pre-trained network are constructed as a cascaded pipeline. The reverse-engineered network first completes the mapping from bandgap characteristics to the circuit parameters of the nonlinear piezoelectric metamaterial, and then the output of the reverse-engineered network is used as the input to the pre-trained network to generate new target characteristics. The pre-trained network is a 1DCNN network; the reverse-engineered network is a combination of a 1DCNN neural network and a Transformer network, containing a complete 1DCNN branch and a Transformer branch. The Adam optimizer is used during backpropagation with a learning rate of 0.0001. The loss function for a concatenated network is: Among them, For data length, To reverse engineer the network's predictions, These are the predictions from the pre-trained network. The weights assigned to the reverse-engineered network.
2. The adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design as described in claim 1, characterized in that: The nonlinear digital piezoelectric metamaterial includes local resonant units, excitation signal sensors, and response signal sensors. The local resonant units are arranged periodically, and two or more one-dimensional periodically arranged local resonant units form multiple vibration control channels. The two or more vibration control channels and their parameters are adjusted to attenuate the vibration waves of the metamaterial at a specified frequency.
3. The adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design as described in claim 2, characterized in that: The local resonant unit includes a piezoelectric transducer and a nonlinear shunt circuit. The local resonant units are arranged periodically on the substrate, and the piezoelectric transducer in the local resonant unit is attached to one side surface of the substrate.
4. The adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design as described in claim 3, characterized in that: The piezoelectric transducer on the substrate is any one of the following shapes: circular, elliptical, rectangular, rhomboid, triangular, or hexagonal. The substrate in the local resonant unit can be rectangular, square, or hexagonal in shape, and the substrate material is metal.
5. The adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design as described in claim 2, characterized in that: The vibration control channel has two or more channels; each vibration control channel contains one local resonant unit.
6. The adaptive vibration suppression method for piezoelectric metamaterials based on deep learning reverse design as described in claim 5, characterized in that: By using a trained reverse design network to reverse design the shunt circuit parameters of a nonlinear digital piezoelectric metamaterial, customized vibration suppression requirements can be met, forming an intelligent vibration suppression scheme. The specific process includes: firstly, applying a frequency sweep excitation to the structure to select the frequency band to be suppressed, and then dividing this frequency band into... The components are divided into equal parts. Then, using MATLAB software, the parts that need to be suppressed are set to 0, and the parts that do not need to be suppressed are set to 1. This label vector is input into the trained reverse engineering network to obtain the predicted circuit parameters. Then, these circuit parameters are written into the microcontroller through online programming to realize the nonlinear digital piezoelectric metamaterial of the corresponding structure. Finally, a frequency sweep excitation is applied to the structure again to see the customized vibration suppression effect.