Physical information neural network dynamic hysteresis model construction method of intelligent driving material

Abstract

The invention belongs to the technical field of intelligent manufacturing and driving control, and particularly discloses a physical information neural network dynamic hysteresis model construction method of an intelligent driving material, which comprises the following steps: constructing a magnetic saturation equation according to the magnetic saturation effect of an intelligent material driving device, and discretizing the magnetic saturation equation to obtain a discrete magnetic saturation equation; constructing a magnetic saturation neural network according to the discrete magnetic saturation equation; and embedding the magnetic saturation neural network into the NARMAX neural network, constructing a physical information neural network hysteresis model embedded with a magnetic saturation equation, and completing modeling. The built physical information neural network hysteresis model embedded with the magnetic saturation equation can reflect magnetic saturation physical information, and the problems that in the prior art, when the dynamic hysteresis nonlinear characteristic of an intelligent material device is described, modeling and prediction precision is low, calculation time is long, and interpretability is poor are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing and drive control, and specifically relates to a method for constructing a physical information neural network dynamic hysteresis model of an intelligent drive material. Background Art

[0002] Common intelligent driving materials include piezoelectric materials, shape memory alloys, magnetostrictive materials, electrorheological fluids, magnetorheological fluids, dielectric elastomers, conductive polymers and ferroelectric materials. These intelligent driving materials can achieve efficient, precise and flexible control and operation, and have unique physical properties and engineering applications in the field of intelligent manufacturing and drive control technology. However, their response mechanism to external excitation (such as electric field, magnetic field, temperature or stress) is relatively complex, and often exhibits hysteresis and nonlinear characteristics. For example, the mechanical strain or charge output generated by piezoelectric materials under the action of electric field is affected by historical excitation; shape memory alloys show hysteresis effects of phase change and shape recovery when temperature or stress changes; magnetostrictive materials show nonlinear strain response under the action of magnetic field; the viscosity of electrorheological fluids and magnetorheological fluids changes with the electric field or magnetic field; the deformation of dielectric elastomers under the action of electric field also shows hysteresis and nonlinear characteristics; the change of conductivity of conductive polymers under the action of electric field usually also shows nonlinear characteristics; the polarization state of ferroelectric materials shows obvious hysteresis and nonlinear response under the action of electric field.

[0003] However, the hysteresis effect means that its response depends not only on the current external excitation, but also on the historical excitation. In other words, the output of the material depends not only on the current input signal, but also on the memory of the past input signal. This hysteresis effect causes the material to exhibit different response characteristics at different operating points and loading histories. Nonlinear input-output relationships usually mean that the output of the material is not a simple proportional relationship to the input signal. This means that more complex mathematical models and algorithms need to be considered to describe and predict the nonlinear behavior of the material.

[0004] Common mathematical modeling to describe the hysteresis and nonlinear characteristics of smart materials can be classified into differential equation models, integral operator models, and artificial intelligence models. Differential equation models have good physical interpretability and are suitable for describing the dynamic evolution and physical mechanisms of the system, but the modeling complexity is high and the dependence on parameters is strong, which leads to the low accuracy of the established model. The integral operator model can capture the historical dependence and nonlinear characteristics of the system and is suitable for processing systems with obvious hysteresis and memory effects, but the modeling and parameter estimation are relatively complex. Artificial intelligence models have strong nonlinear modeling capabilities and adaptability, and can handle big data and complex systems, but they have poor interpretability and high requirements for data volume and computing resources. In particular, when the amount of data is small, the model performance is poor.

[0005] Given the inherent nonlinear characteristics of smart materials such as dynamic hysteresis, creep, and temperature drift, the repeatability and accuracy of their output in practical applications are significantly reduced, which brings many challenges to the application. Therefore, many scientific research institutions and researchers focus on modeling these dynamic hysteresis nonlinear characteristics in order to more accurately describe the behavioral characteristics of these materials. Neural networks and data-driven methods have attracted much attention due to their excellent performance, but they still need to be improved in terms of interpretability and transparency. Summary of the invention

[0006] The purpose of the present invention is to solve the problems of low modeling and prediction accuracy, long calculation time and poor interpretability in the prior art when describing the dynamic hysteresis nonlinear characteristics of smart material devices, and proposes a method for constructing a physical information neural network dynamic hysteresis model of smart driving materials.

[0007] The technical solution of the present invention is: a method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material, comprising the following steps:

[0008] According to the magnetic saturation effect of the smart material driving device, the magnetic saturation equation is constructed, and the magnetic saturation equation is discretized to obtain a discrete magnetic saturation equation;

[0009] According to the structural relationship of the discrete magnetic saturation equation, the signal transmission relationship between the input signal and the output signal of the discrete magnetic saturation equation and the parameters of the discrete magnetic saturation equation, a magnetic saturation neural network is constructed;

[0010] The magnetic saturation neural network is embedded into the NARMAX neural network, and a physical information neural network hysteresis model embedded with the magnetic saturation equation is constructed to complete the modeling.

[0011] The beneficial effects of the present invention are:

[0012] The present invention constructs a magnetic saturation differential equation as a magnetic saturation neural network with a special architecture, and further embeds the constructed magnetic saturation neural network into a NARMAX neural network, forming a new neural network that can reflect magnetic saturation physical information and has certain physical interpretability. At the same time, the physical information neural network hysteresis model embedded with the magnetic saturation equation constructed by the present invention has high accuracy, low calculation cost, short required time, small required data volume, and high interpretability, and can effectively solve the problems of low modeling and prediction accuracy, long calculation time, and poor interpretability in the prior art when describing the dynamic hysteresis nonlinear characteristics of smart material devices.

[0013] Finally, the use of the physical information neural network with a special network architecture designed by the present invention and embedded with the magnetic saturation equation to model the dynamic nonlinear characteristics of intelligent materials can effectively solve the key problems currently existing and has important theoretical significance and engineering application value.

[0014] Preferably, the magnetic saturation equation is:

[0015]

[0016] Wherein, t represents a continuous time variable, h(t) represents the output damping force of the magnetorheological damper, which is an intelligent material drive device based on intelligent driving material - magnetorheology, and u(t) represents the input voltage of the magnetorheological damper. represents the derivative of the input voltage u(t) of the magnetorheological damper, sgn(u(t)) represents the sign function, which takes the value -1 when u(t) is less than 0, takes the value 0 when u(t) is equal to 0, and takes the value 1 when u(t) is greater than 0, tanh(·) represents the hyperbolic tangent function, C c Represents the damping coefficient, K k Represents the stiffness coefficient, F y represents the damping force related to the magnetic field, β and λ represent the shape factors of the hysteresis curve, and F0 represents the initial force related to the initial position.

[0017] Preferably, the discrete magnetic saturation equation is:

[0018]

[0019] Where T represents the sampling interval, kT represents the discrete sampling time, h(kT) represents the output damping force of the discrete magnetorheological damper, u(kT) represents the input voltage of the discrete magnetorheological damper, represents the derivative of the input voltage u(kT) of the discrete magnetorheological damper, sgn(u(kT)) represents the sign function, which takes the value -1 when u(kT) is less than 0, takes the value 0 when u(kT) is equal to 0, and takes the value 1 when u(kT) is greater than 0, tanh(·) represents the hyperbolic tangent function, C c Represents the damping coefficient, K k represents the stiffness coefficient, Fy represents the damping force related to the magnetic field, β and λ represent the shape factors of the hysteresis curve, and F0 represents the initial force related to the initial position.

[0020] Preferably, the magnetic saturation neural network includes 1 input layer, 15 hidden layers, and 1 output layer;

[0021] The input layer includes 8 input units, namely u(kT), C1, C2, C3, C4, C5 and C6; among them, C1, C2, C3, C4, C5 and C6 are all constant vectors;

[0022] The 15 hidden layers are hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5, hidden layer 6, hidden layer 7, hidden layer 8, hidden layer 9, hidden layer 10, hidden layer 11, hidden layer 12, hidden layer 13, hidden layer 14 and hidden layer 15 respectively;

[0023] The 15 hidden layers and the output layer each have only one neuron, the connection weights between the layers of the magnetic saturation neural network are all set to 1, and there is no delay between the layers.

[0024] Preferably, the specific design process of the magnetic saturation neural network is:

[0025] Input the output data of input unit C1 to hidden layer 1 for calculation; input the output data of input unit C2 to hidden layer 3 for calculation; input the output data of input unit C3 to hidden layer 7 for calculation; input the output data of input unit C4 to hidden layer 9 for calculation; input the output data of input unit C5 to hidden layer 12 for calculation; input the output data of input unit C6 to hidden layer 14 for calculation; input the output data of input unit u(kT) to hidden layer 6 for calculation;

[0026] The output data of hidden layer 1 and the input unit The output data of the hidden layer 2 is input to the hidden layer 2 for calculation; the output data of the hidden layer 3 and the output data of the hidden layer 6 are input to the hidden layer 4 for calculation; the output data of the input unit u(kT) and the output data of the hidden layer 7 are input to the hidden layer 8 for calculation;

[0027] Input unit The output data of hidden layer 5 and the output data of hidden layer 4 are input to hidden layer 5 for calculation; the output data of hidden layer 5 and the output data of hidden layer 9 are input to hidden layer 10 for calculation; the output data of hidden layer 10 is input to hidden layer 11 for calculation; the output data of hidden layer 11 and the output data of hidden layer 12 are input to hidden layer 13 for calculation; the output data of hidden layer 13 and the output data of hidden layer 14 are input to hidden layer 15 for calculation; the output data of hidden layer 2, the output data of hidden layer 8 and the output data of hidden layer 15 are input to the output layer for processing, and output h(kT).

[0028] Preferably, the output data of the hidden layer 1, hidden layer 3, hidden layer 7, hidden layer 9, hidden layer 12 and hidden layer 14 are respectively: the damping coefficient C c , stiffness coefficient K k , the damping force F related to the magnetic field y , about the hysteresis curve shape factor β, about the hysteresis curve shape factor λ and the initial force F0 related to the initial position;

[0029] The hidden layer 11 performs normalization processing on the input data;

[0030] The hidden layer 6 performs binarization processing on the input data;

[0031] The hidden layer 2, the hidden layer 4, the hidden layer 8, the hidden layer 10 and the hidden layer 13 respectively perform element-by-element multiplication processing on their input data;

[0032] The hidden layer 5 and the hidden layer 15 perform element-by-element addition processing on their input data respectively.

[0033] Preferably, the activation functions of the hidden layer 1, hidden layer 3, hidden layer 7, hidden layer 9, hidden layer 12 and hidden layer 14 are user-defined activation functions, which are respectively expressed as: S a , S b , S c , S d , S e and S f ,have:

[0034]

[0035] Among them, e represents the natural base, x represents the input function of each layer, and d a ,d b ,d c ,d d ,d e and d f Both represent the search step parameters;

[0036] The activation function of the hidden layer 6 is the symbol activation function S sgn , and its expression formula is:

[0037] S sgn =sign(x);

[0038] The activation function of the hidden layer 11 is the hyperbolic tangent activation function S tanh , and its expression formula is:

[0039] S tanh =tanh(x);

[0040] The activation functions of the input layer, hidden layer 4, hidden layer 10, hidden layer 13, hidden layer 15 and output layer are all linear activation functions.

[0041] Preferably, the physical information neural network hysteresis model embedded in the magnetic saturation equation includes a magnetic saturation enhanced ARMAX antecedent and a deep multilayer perceptron post-component connected in sequence; the magnetic saturation enhanced ARMAX antecedent is embedded with a magnetic saturation neural network for receiving and processing the input voltage u(kT) of the discrete magnetorheological damper and the derivative of the input voltage u(kT) of the discrete magnetorheological damper The output is h(kT);

[0042] The input end of the magnetic saturation enhanced ARMAX antecedent is the input end of the physical information neural network hysteresis model embedded in the magnetic saturation equation; the input of the magnetic saturation enhanced ARMAX antecedent includes the input voltage u(kT) of the discrete magnetorheological damper, the derivative of the input voltage u(kT) of the discrete magnetorheological damper and the output displacement y of the smart material driving device; the output of the magnetic saturation enhanced ARMAX antecedent includes the output displacement y of the smart material driving device, the first-order delay of the output displacement y of the smart material driving device, the second-order delay of the output displacement y of the smart material driving device, the difference ε between the output displacement y of the smart material driving device and the output damping force h(kT) of the discrete magnetorheological damper, the first-order delay of ε, the second-order delay of ε, the input voltage u(kT) of the discrete magnetorheological damper, the first-order delay of u(kT), the second-order delay of u(kT), and the derivative of the input voltage u(kT) of the discrete magnetorheological damper The first-order delay and The second-order delay of

[0043] The input of the deep multilayer perceptron afterpart is the output of the magnetic saturation enhanced ARMAX beforepart, and the output end of the deep multilayer perceptron afterpart is the output end of the physical information neural network hysteresis model embedded in the magnetic saturation equation.

[0044] Preferably, the deep multi-layer perceptron backend includes two hidden layers and an output layer connected in sequence; the two hidden layers are fully connected layers, containing multiple neurons, and the two hidden layers use a sigmoid activation function; the output layer has only one neuron and uses a linear activation function.

[0045] Preferably, when the physical information neural network hysteresis model embedded with the magnetic saturation equation is trained, the input voltage sequence data of the magnetorheological damper is used as the input of the physical information neural network hysteresis model embedded with the magnetic saturation equation, and the output displacement sequence data is used as a label, which specifically includes the following steps:

[0046] Sending an excitation input signal of a certain frequency and a certain amplitude through a main control computer, and amplifying the power of the excitation input signal;

[0047] Apply the power-amplified excitation input signal to the smart material drive device to generate a corresponding displacement output signal, use the sensor to measure the specific displacement output, and transmit it to the main control computer for display to obtain a training data set;

[0048] The training data set is input into the physical information neural network hysteresis model embedded with the magnetic saturation equation, and the weight of the magnetic saturation neural network is trained by using the Levenberg-Marquardt algorithm to obtain the weight of the magnetic saturation neural network;

[0049] Keeping the weights of the magnetic saturation neural network unchanged, the back propagation BP algorithm is used to train the weights of the deep multilayer perceptron posterior until the mean square error (MSE) is less than 1.0×10 -5 Or the number of training steps reaches the preset value.

[0050] The beneficial effects of the above preferred solution are:

[0051] 1. The physical information neural network hysteresis model embedded in the magnetic saturation equation proposed in the present invention includes two parts: a magnetic saturation enhanced ARMAX antecedent and a deep multi-layer perceptron after-part. The enhanced ARMAX antecedent is not a traditional black box neural network, but a transparent network with clear physical characteristics and mathematical interpretability derived from the magnetic saturation equation; the deep multi-layer perceptron after-part ensures modeling accuracy and generalization ability based on data-driven, so the physical information neural network hysteresis model embedded in the magnetic saturation equation has both data-driven and physically interpretable characteristics.

[0052] 2. When training the network weights, only the weights of the six neurons corresponding to the six parameters of the magnetic saturation equation can be adaptively adjusted, and the other weights remain unchanged during the training process. Therefore, the network size and complexity are small, which is very conducive to network training.

[0053] 3. The proposed physical information neural network hysteresis model embedded with magnetic saturation equation can be used to accurately and quickly model smart material driver devices; it can be used to characterize the dynamic rate-dependent and asymmetric hysteresis response exhibited by smart materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The figure shows a flow chart of a method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material proposed in Example 1 of the present invention.

[0055] Figure 2 The figure shows the topological structure diagram of the magnetic saturation neural network proposed in Example 1 of the present invention.

[0056] Figure 3Shown is a structural diagram of a physical information neural network hysteresis model embedded with a magnetic saturation equation proposed in Example 1 of the present invention.

[0057] Figure 4 Shown is a diagram of the experimental equipment proposed in Example 2 of the present invention.

[0058] Figure 5 FIG. 2 is a schematic diagram of an input excitation signal proposed in Embodiment 2 of the present invention.

[0059] Figure 6 Shown is a schematic diagram of the output displacement predicted by the physical information neural network hysteresis model of the experimental test and embedded magnetic saturation equation proposed in Example 2 of the present invention.

[0060] Figure 7 Shown is a graph of the error results between the output of the experimental test proposed in Example 2 of the present invention and the output predicted by the physical information neural network hysteresis model embedded in the magnetic saturation equation.

[0061] Figure 8 Shown are input-output hysteresis curves predicted by the experimental test and the physical information neural network hysteresis model embedded with the magnetic saturation equation proposed in Example 2 of the present invention. DETAILED DESCRIPTION

[0062] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0063] Embodiment 1:

[0064] like Figure 1 As shown, a method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material comprises the following steps:

[0065] According to the magnetic saturation effect of the smart material driving device, the magnetic saturation equation is constructed, and the magnetic saturation equation is discretized to obtain a discrete magnetic saturation equation;

[0066] According to the structural relationship of the discrete magnetic saturation equation, the signal transmission relationship between the input signal and the output signal of the discrete magnetic saturation equation and the parameters of the discrete magnetic saturation equation, a magnetic saturation neural network is constructed;

[0067] The magnetic saturation neural network is embedded into the NARMAX neural network, and a physical information neural network hysteresis model embedded with the magnetic saturation equation is constructed to complete the modeling.

[0068] In this embodiment, when the smart material driving device is used in a shock absorption occasion, the magnetorheological device is simplified to a series connection of a viscous damping element, a stiffness element, a spring element, and a mass element. At the same time, based on the magnetic saturation effect, the magnetic saturation equation is obtained as follows:

[0069]

[0070] Wherein, t represents a continuous time variable, h(t) represents the output damping force of the magnetorheological damper. The magnetorheological damper is a vibration damping device based on the intelligent driving material - magnetorheology. It is an intelligent material driving device. The hysteresis response characteristics of the magnetorheological damper are derived from the magnetorheological material characteristics. u(t) represents the input voltage of the magnetorheological damper. represents the derivative of the input voltage u(t) of the magnetorheological damper, sgn(u(t)) represents the sign function, which takes the value -1 when u(t) is less than 0, takes the value 0 when u(t) is equal to 0, and takes the value 1 when u(t) is greater than 0, tanh(·) represents the hyperbolic tangent function, C c Represents the damping coefficient, K k represents the stiffness coefficient, Fy represents the damping force related to the magnetic field, β and λ represent the shape factors of the hysteresis curve, and F0 represents the initial force related to the initial position.

[0071] The discrete magnetic saturation equation is:

[0072]

[0073] Where T represents the sampling interval, kT represents the discrete sampling time, h(kT) represents the output damping force of the discrete magnetorheological damper, u(kT) represents the input voltage of the discrete magnetorheological damper, Represents the derivative of the input voltage u(kT) of the discrete magnetorheological damper.

[0074] In this embodiment, if Figure 2 As shown, the magnetic saturation neural network includes 1 input layer, 15 hidden layers, and 1 output layer; wherein, and They represent the node addition operator and the node multiplication operator respectively;

[0075] The input layer includes 8 input units, namely u(kT), C1, C2, C3, C4, C5 and C6; among them, C1, C2, C3, C4, C5 and C6 are all constant vectors, the values ​​of these constant input vectors are all 1, and the length of each constant input vector is the same as the length of the input signal u(kT).

[0076] The 15 hidden layers are hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5, hidden layer 6, hidden layer 7, hidden layer 8, hidden layer 9, hidden layer 10, hidden layer 11, hidden layer 12, hidden layer 13, hidden layer 14 and hidden layer 15 respectively;

[0077] The 15 hidden layers and the output layer each have only one neuron, the connection weights between the layers of the magnetic saturation neural network are all set to 1, and there is no delay between the layers.

[0078] In this embodiment, the specific design process of the magnetic saturation neural network is:

[0079] Input the output data of input unit C1 to hidden layer 1 for calculation; input the output data of input unit C2 to hidden layer 3 for calculation; input the output data of input unit C3 to hidden layer 7 for calculation; input the output data of input unit C4 to hidden layer 9 for calculation; input the output data of input unit C5 to hidden layer 12 for calculation; input the output data of input unit C6 to hidden layer 14 for calculation; input the output data of input unit u(kT) to hidden layer 6 for calculation;

[0080] The output data of hidden layer 1 and the input unit The output data of the hidden layer 2 is input to the hidden layer 2 for calculation; the output data of the hidden layer 3 and the output data of the hidden layer 6 are input to the hidden layer 4 for calculation; the output data of the input unit u(kT) and the output data of the hidden layer 7 are input to the hidden layer 8 for calculation;

[0081] Input unit The output data of hidden layer 5 and the output data of hidden layer 4 are input to hidden layer 5 for calculation; the output data of hidden layer 5 and the output data of hidden layer 9 are input to hidden layer 10 for calculation; the output data of hidden layer 10 is input to hidden layer 11 for calculation; the output data of hidden layer 11 and the output data of hidden layer 12 are input to hidden layer 13 for calculation; the output data of hidden layer 13 and the output data of hidden layer 14 are input to hidden layer 15 for calculation; the output data of hidden layer 2, the output data of hidden layer 8 and the output data of hidden layer 15 are input to the output layer for processing, and output h(kT).

[0082] In this embodiment, the output data of the hidden layer 1, hidden layer 3, hidden layer 7, hidden layer 9, hidden layer 12 and hidden layer 14 are respectively: damping coefficient C c , stiffness coefficient K k , the damping force F related to the magnetic field y , about the hysteresis curve shape factor β, about the hysteresis curve shape factor λ and the initial force F0 related to the initial position;

[0083] The hidden layer 11 performs normalization processing on the input data;

[0084] The hidden layer 6 performs binarization processing on the input data;

[0085] The hidden layer 2, the hidden layer 4, the hidden layer 8, the hidden layer 10 and the hidden layer 13 respectively perform element-by-element multiplication processing on their input data;

[0086] The hidden layer 5 and the hidden layer 15 perform element-by-element addition processing on their input data respectively.

[0087] In this embodiment, the activation functions of the hidden layer 1, hidden layer 3, hidden layer 7, hidden layer 9, hidden layer 12 and hidden layer 14 are user-defined activation functions, which are respectively expressed as: S a , S b , S c , S d , S e and S f ,have:

[0088]

[0089] Among them, e represents the natural base, x represents the input function of each layer, and d a ,d b ,d c ,d d ,d e and d f Both represent the search step parameters;

[0090] The activation function of the hidden layer 6 is the symbol activation function S sgn , and its expression formula is:

[0091] S sgn =sign(x);

[0092] The activation function of the hidden layer 11 is the hyperbolic tangent activation function S tanh , and its expression formula is:

[0093] S tanh =tanh(x);

[0094] The activation functions of the input layer, hidden layer 4, hidden layer 10, hidden layer 13, hidden layer 15 and output layer are all linear activation functions.

[0095] In this embodiment, if Figure 3As shown, the physical information neural network hysteresis model embedded in the magnetic saturation equation includes a magnetic saturation enhanced ARMAX antecedent and a deep multilayer perceptron post-component connected in sequence; the magnetic saturation enhanced ARMAX antecedent is embedded with a magnetic saturation neural network for receiving and processing the input voltage u(kT) of the discrete magnetorheological damper and the derivative of the input voltage u(kT) of the discrete magnetorheological damper The output is h(kT);

[0096] The input end of the magnetic saturation enhanced ARMAX antecedent is the input end of the physical information neural network hysteresis model embedded in the magnetic saturation equation; the input of the magnetic saturation enhanced ARMAX antecedent includes the input voltage u(kT) of the discrete magnetorheological damper, the derivative of the input voltage u(kT) of the discrete magnetorheological damper and the output displacement y of the smart material drive device; the output of the magnetic saturation enhanced ARMAX front device includes the output displacement y of the smart material drive device, the first-order delay of the output displacement y of the smart material drive device, the second-order delay of the output displacement y of the smart material drive device, and the output displacement y of the smart material drive device y The difference between the output damping force h(kT) of the discrete magnetorheological damper ε, the first-order delay of ε, the second-order delay of ε, the input voltage u(kT) of the discrete magnetorheological damper, the first-order delay of u(kT), the second-order delay of u(kT), the derivative of the input voltage u(kT) of the discrete magnetorheological damper The first-order delay and The second-order delay of

[0097] The input of the deep multilayer perceptron afterpart is the output of the magnetic saturation enhanced ARMAX beforepart, and the output end of the deep multilayer perceptron afterpart is the output end of the physical information neural network hysteresis model embedded in the magnetic saturation equation.

[0098] In this embodiment, the deep multi-layer perceptron post-processor includes two hidden layers and an output layer connected in sequence; the two hidden layers are fully connected layers, containing multiple neurons, and the two hidden layers use a sigmoid activation function; the output layer has only one neuron and uses a linear activation function.

[0099] In this embodiment, when the physical information neural network hysteresis model embedded in the magnetic saturation equation is trained, the input voltage sequence data of the magnetorheological damper is used as the input of the physical information neural network hysteresis model embedded in the magnetic saturation equation, and the output displacement sequence data is used as a label, which specifically includes the following steps:

[0100] The main control computer sends an excitation input signal of a certain frequency and a certain amplitude, and the excitation input signal is power-amplified; the power-amplified excitation input signal is applied to the smart material drive device to generate a corresponding displacement output signal, and the specific displacement output is measured by a sensor and transmitted to the main control computer for display to obtain a training data set; the excitation signal generally uses a variable amplitude swept frequency signal to excite the smart material drive device;

[0101] The training data set is input into the physical information neural network hysteresis model embedded with the magnetic saturation equation, and the weight of the magnetic saturation neural network is trained using the Levenberg-Marquardt algorithm to obtain the weight of the magnetic saturation neural network;

[0102] Keeping the weights of the magnetic saturation neural network unchanged, the back propagation BP algorithm is used to train the weights of the deep multilayer perceptron posterior until the mean square error (MSE) is less than 1.0×10 -5 Or the number of training steps reaches the preset value; the number of training steps is set to 500.

[0103] The modeling method proposed by the present invention has strong universality and is applicable to piezoelectric materials, shape memory alloys, magnetostrictive materials, electrorheological fluids, magnetorheological fluids, dielectric elastomers, conductive polymers, ferroelectric materials, etc., and is easy to implement in engineering.

[0104] Embodiment 2:

[0105] Based on Example 1, this example designs a physical information neural network embedded with a magnetic saturation equation for the smart material piezoelectric actuator to model its input and output hysteresis characteristics. Figure 4 The experimental equipment shown is used to collect the input and output data of the piezoelectric actuator.

[0106] Use Figure 5 The input voltage signal shown in the figure is used to excite the piezoelectric driver and obtain the following Figure 6 The measured output displacement data is shown. The physical information neural network hysteresis model embedded in the magnetic saturation equation is trained according to the input and output data obtained from the test to obtain the weight of the physical information neural network hysteresis model embedded in the magnetic saturation equation. Then the physical information neural network hysteresis model embedded in the magnetic saturation equation is used for prediction, and the following is obtained: Figure 6 The predicted displacement output data is shown in Figure 2. The error between the two is shown in Figure 2. Figure 7 The maximum error is less than 1.0% of full scale. Figure 8The input-output hysteresis curves of experimental test and neural network prediction are given. From the experimental results, it can be seen that the results of the experimental test are basically consistent with the results predicted by the physical information neural network hysteresis model embedded in the magnetic saturation equation, which further illustrates the effectiveness and feasibility of the parameter identification method. At the same time, for the input and output data of this case, Table 1 shows the comparative results of the traditional modeling method and the physical information neural network hysteresis model construction method embedded in the magnetic saturation equation proposed in the present invention in terms of modeling performance indicators. It can be seen that the method proposed in the present invention has higher accuracy and efficiency. After adopting the modeling method of the physical information neural network hysteresis model embedded in the magnetic saturation equation proposed in the present invention, its maximum modeling error is much lower than other common operator models and differential equation models.

[0107] Table 1 Comparison of modeling performance indicators of common models

[0108]

[0109] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for constructing a physical information neural network dynamic hysteresis model of intelligent driving materials, characterized in that: The following steps are involved: According to the magnetic saturation effect of the smart material driving device, the magnetic saturation equation is constructed, and the magnetic saturation equation is discretized to obtain the discrete magnetic saturation equation; According to the structural relationship between the terms of the discrete magnetic saturation equation, the signal transmission relationship between the input signal and the output signal of the discrete magnetic saturation equation, and the parameters of the discrete magnetic saturation equation, a magnetic saturation neural network is constructed; The magnetic saturation neural network is embedded into the NARMAX neural network, and a physical information neural network hysteresis model embedded with the magnetic saturation equation is constructed to complete the modeling.

2. The method for constructing a physical information neural network dynamic hysteresis model of intelligent driving materials according to claim 1, characterized in that: The magnetic saturation equation is: Wherein, t represents a continuous time variable, h(t) represents the output damping force of the magnetorheological damper, which is an intelligent material drive device based on intelligent driving material - magnetorheology, and u(t) represents the input voltage of the magnetorheological damper. represents the derivative of the input voltage u(t) of the magnetorheological damper, sgn(u(t)) represents the sign function, which takes the value -1 when u(t) is less than 0, takes the value 0 when u(t) is equal to 0, and takes the value 1 when u(t) is greater than 0, tanh(·) represents the hyperbolic tangent function, C c Represents the damping coefficient, K k Represents the stiffness coefficient, F y represents the damping force related to the magnetic field, β and λ represent the shape factors of the hysteresis curve, and F0 represents the initial force related to the initial position.

3. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 1, characterized in that: The discrete magnetic saturation equation is: Where T represents the sampling interval, kT represents the discrete sampling time, h(kT) represents the output damping force of the discrete magnetorheological damper, u(kT) represents the input voltage of the discrete magnetorheological damper, represents the derivative of the input voltage u(kT) of the discrete magnetorheological damper, sgn(u(kT)) represents the sign function, which takes the value -1 when u(kT) is less than 0, takes the value 0 when u(kT) is equal to 0, and takes the value 1 when u(kT) is greater than 0, tanh(·) represents the hyperbolic tangent function, C c Represents the damping coefficient, K k represents the stiffness coefficient, Fy represents the damping force related to the magnetic field, β and λ represent the shape factors of the hysteresis curve, and F0 represents the initial force related to the initial position.

4. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 3, characterized in that: The magnetic saturation neural network includes 1 input layer, 15 hidden layers, and 1 output layer; The input layer includes 8 input units, namely u(kT), C1, C2, C3, C4, C5 and C6; Among them, C1, C2, C3, C4, C5 and C6 are all constant vectors; The 15 hidden layers are hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5, hidden layer 6, hidden layer 7, hidden layer 8, hidden layer 9, hidden layer 10, hidden layer 11, hidden layer 12, hidden layer 13, hidden layer 14 and hidden layer 15 respectively; The 15 hidden layers and the output layer each have only one neuron, the connection weights between the layers of the magnetic saturation neural network are all set to 1, and there is no delay between the layers.

5. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 4, characterized in that: The specific design process of the magnetic saturation neural network is as follows: Input the output data of input unit C1 to hidden layer 1 for calculation; input the output data of input unit C2 to hidden layer 3 for calculation; input the output data of input unit C3 to hidden layer 7 for calculation; input the output data of input unit C4 to hidden layer 9 for calculation; input the output data of input unit C5 to hidden layer 12 for calculation; input the output data of input unit C6 to hidden layer 14 for calculation; input the output data of input unit u(kT) to hidden layer 6 for calculation; The output data of hidden layer 1 and the input unit The output data of the hidden layer 2 is input to the hidden layer 2 for calculation; the output data of the hidden layer 3 and the output data of the hidden layer 6 are input to the hidden layer 4 for calculation; the output data of the input unit u(kT) and the output data of the hidden layer 7 are input to the hidden layer 8 for calculation; Input unit The output data of hidden layer 5 and the output data of hidden layer 4 are input to hidden layer 5 for calculation; the output data of hidden layer 5 and the output data of hidden layer 9 are input to hidden layer 10 for calculation; the output data of hidden layer 10 is input to hidden layer 11 for calculation; the output data of hidden layer 11 and the output data of hidden layer 12 are input to hidden layer 13 for calculation; the output data of hidden layer 13 and the output data of hidden layer 14 are input to hidden layer 15 for calculation; the output data of hidden layer 2, the output data of hidden layer 8 and the output data of hidden layer 15 are input to the output layer for processing, and output h(kT).

6. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 4, characterized in that: The output data of the hidden layer 1, hidden layer 3, hidden layer 7, hidden layer 9, hidden layer 12 and hidden layer 14 are respectively: damping coefficient C c , stiffness coefficient K k , the damping force F related to the magnetic field y , about the hysteresis curve shape factor β, about the hysteresis curve shape factor λ and the initial force F0 related to the initial position; The hidden layer 11 performs normalization processing on the input data; The hidden layer 6 performs binarization processing on the input data; The hidden layer 2, the hidden layer 4, the hidden layer 8, the hidden layer 10 and the hidden layer 13 respectively perform element-by-element multiplication processing on their input data; The hidden layer 5 and the hidden layer 15 perform element-by-element addition processing on their input data respectively.

7. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 4, characterized in that: The activation functions of the hidden layer 1, hidden layer 3, hidden layer 7, hidden layer 9, hidden layer 12 and hidden layer 14 are self-defined activation functions, which are respectively expressed as: S a , S b , S c , S d , S e and S f ,have: Among them, e represents the natural base, x represents the input function of each layer, and d a ,d b ,d c ,d d ,d e and d f Both represent the search step parameters; The activation function of the hidden layer 6 is the symbol activation function S sgn , and its expression formula is: S sgn =sign(x); The activation function of the hidden layer 11 is the hyperbolic tangent activation function S tanh , and its expression formula is: S tanh =tanh(x); The activation functions of the input layer, hidden layer 2, hidden layer 4, hidden layer 5, hidden layer 8, hidden layer 10, hidden layer 13, hidden layer 15 and output layer are all linear activation functions.

8. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 5, characterized in that: The physical information neural network hysteresis model embedded with the magnetic saturation equation includes a magnetic saturation enhanced ARMAX antecedent and a deep multilayer perceptron antecedent connected in sequence; a magnetic saturation neural network is embedded in the magnetic saturation enhanced ARMAX antecedent for receiving and processing the input voltage u(kT) of the discrete magnetorheological damper and the derivative of the input voltage u(kT) of the discrete magnetorheological damper The output is h(kT); The input end of the magnetic saturation enhanced ARMAX antecedent is the input end of the physical information neural network hysteresis model embedded in the magnetic saturation equation; the magnetic saturation enhanced ARMAX antecedent includes three input units, which are the input voltage u(kT) of the discrete magnetorheological damper, the derivative of the input voltage u(kT) of the discrete magnetorheological damper, and the derivative of the input voltage u(kT) of the discrete magnetorheological damper. and the output displacement y of the smart material driving device; the magnetic saturation enhanced ARMAX antecedent includes 12 output units, namely the output displacement y of the smart material driving device, the first-order delay of the output displacement y of the smart material driving device, the second-order delay of the output displacement y of the smart material driving device, the difference ε between the output displacement y of the smart material driving device and the output damping force h(kT) of the discrete magnetorheological damper, the first-order delay of ε, the second-order delay of ε, the input voltage u(kT) of the discrete magnetorheological damper, the first-order delay of u(kT), the second-order delay of u(kT), the derivative of the input voltage u(kT) of the discrete magnetorheological damper The first-order delay and The second-order delay of The input of the deep multilayer perceptron afterpart is the output of the magnetic saturation enhanced ARMAX beforepart, and the output end of the deep multilayer perceptron afterpart is the output end of the physical information neural network hysteresis model embedded in the magnetic saturation equation.

9. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 8, characterized in that: The deep multi-layer perceptron backend includes two hidden layers and an output layer connected in sequence; the two hidden layers are fully connected layers, containing multiple neurons, and the two hidden layers use a sigmoid activation function; the output layer has only one neuron and uses a linear activation function.

10. The method for constructing a physical information neural network dynamic hysteresis model of an intelligent driving material according to claim 4, characterized in that: When the physical information neural network hysteresis model embedded with the magnetic saturation equation is trained, the input voltage sequence data of the magnetorheological damper is used as the input of the physical information neural network hysteresis model embedded with the magnetic saturation equation, and the output displacement sequence data is used as a label, which specifically includes the following steps: Sending an excitation input signal of a certain frequency and a certain amplitude through a main control computer, and amplifying the power of the excitation input signal; Apply the power-amplified excitation input signal to the smart material drive device to generate a corresponding displacement output signal, use the sensor to measure the specific displacement output, and transmit it to the main control computer for display to obtain a training data set; The training data set is input into the physical information neural network hysteresis model embedded with the magnetic saturation equation, and the weight ω1 of hidden layer 1, the weight ω2 of hidden layer 3, the weight ω3 of hidden layer 7, the weight ω4 of hidden layer 9, the weight ω5 of hidden layer 12 and the weight ω6 of hidden layer 14 in the magnetic saturation neural network are trained by the Levenberg-Marquardt algorithm to obtain the weight of the magnetic saturation neural network; Keeping the weights of the magnetic saturation neural network unchanged, the back propagation BP algorithm is used to train the weights of the deep multilayer perceptron posterior until the mean square error (MSE) is less than 1.0×10 -5 Or the number of training steps reaches the preset value.

Images