Shape memory alloy soft actuator control method based on data-driven modeling

Abstract

The application relates to a shape memory alloy soft driver control method based on data-driven modeling, which comprises the following steps: measuring and collecting input and output data of a shape memory alloy soft driver as sample data; designing an asymmetric activation function containing a hysteresis characteristic, constructing a neural network structure and selecting corresponding parameters, combining sample data and an improved gradient algorithm for iterative training, simultaneously introducing L1 regularization to generate a sparse model, and obtaining a soft driver theoretical model; inputting current soft driver corresponding input data into the soft driver theoretical model, outputting corresponding simulation output data, and controlling the soft driver according to the simulation output data. Compared with the prior art, the application adopts a neural network model, combines an improved nonlinear activation function and a gradient descent algorithm, improves the modeling accuracy and robustness of a hysteresis system, and effectively improves the subsequent control effect.

Description

Technical Field

[0001] The present invention relates to the technical field of soft robots, and in particular to a shape memory alloy soft driver control method based on data-driven modeling. Background Art

[0002] Soft robots are a new branch of robotics. Typically made of soft materials, they can arbitrarily change their shape and size over a wide range of applications, offering broad application prospects in fields such as reconnaissance, detection, rescue, and medical care. Compared to rigid robots, soft robots offer advantages such as greater environmental adaptability, flexibility, and safety. With the continuous development of smart materials, the use of smart materials with sensory capabilities similar to those of animal muscles to form robot drive structures will offer advantages such as simple structure, good environmental adaptability, low noise, and the ability to actively generate complex motions. Therefore, they are becoming a future trend in soft robotics. Shape memory alloys, as smart materials, are currently widely used in the design of drive structures for soft robots due to their high power-to-weight ratio, low drive voltage, compact size, light weight, pollution-free operation, and noise-free operation.

[0003] However, in practical applications, the mutual transformation between martensite and austenite leads to the shape memory effect of shape memory alloys, which manifests as hysteresis and nonlinear characteristics. Furthermore, shape memory alloy soft actuators have other system characteristics such as dynamic uncertainty and time-varying properties. Therefore, how to improve the modeling accuracy and robustness of shape memory alloy soft actuators, thereby improving control performance, is a hot topic in academia and industry.

[0004] Chinese patent CN105353610A discloses a modeling method for a magnetically controlled shape memory alloy actuator based on the KP (Krasnosel'skii-Pokrovskii) model. This method uses the KP model to model the hysteresis of the magnetically controlled shape memory alloy actuator, enabling the prediction of the hysteresis phenomenon of the magnetically controlled shape memory alloy actuator. However, the mathematical description and solution of the KP model are relatively complex, making it difficult to accurately determine the model parameters, which limits the further application of this technology.

[0005] Chinese patent CN111898235A discloses a parameter identification method for magnetically controlled shape memory alloy actuators based on the Duhem model. This method establishes a Duhem model that accurately describes the hysteresis nonlinearity of magnetically controlled shape memory alloy actuators, achieving high-precision modeling. However, the Duhem model involves nonlinear differential and integral equations, making its mathematical description and solution complex. Furthermore, the selection of appropriate observation and model functions for different control objects precludes widespread application of this technology in various application scenarios.

[0006] Chinese patent CN115600505A discloses a hysteresis modeling method for magnetically controlled shape memory alloy actuators based on the ARPI model. This method proposes an asymmetric Play operator and introduces nonlinear polynomials to establish an asymmetric PI model. Compared with the traditional PI model, the ARPI model improves the accuracy of describing asymmetric hysteresis. However, the modeling process requires manual extraction of hysteresis loop characteristics at different frequencies and complex model parameter identification.

[0007] Furthermore, Chinese patent CN211401593U discloses a device for testing the temperature-controlled drive characteristics of a shape memory alloy spring. This device can conveniently, efficiently, and accurately measure the relationship between temperature, deformation, and elastic force during the heating and cooling process of a shape memory alloy spring in a water bath. However, it does not consider the effects of temperature sensor delay and input voltage variations on the spring's elastic force output.

[0008] In summary, existing research has proposed numerous models to describe the hysteresis phenomenon of shape memory alloys, such as the Preisach model, the PI (Prandtl-Ishlinskii) model, the KP (Krasnosel'skii-Pokrovskii) model, and the Duhem model. However, the Preisach model suffers from a complex modeling process, low accuracy, and poor generalization. The PI model improves upon the Preisach model by having an analytical inverse, making it easier to find the hysteresis inverse model. However, the PI model can only describe symmetric hysteresis phenomena, limiting its application. The KP model can more accurately characterize the nonlinear dynamic behavior of hysteretic systems, but it typically involves nonlinear differential equations, making its mathematical description and solution complex. The Duhem model is a nonlinear hysteresis differential equation model that requires the selection of appropriate observation and model functions for different applications. This makes it difficult to select the appropriate function when modeling a new model. These factors contribute to poor accuracy, adaptability, and generalization in soft robot modeling, making it difficult to obtain accurate simulation output data, resulting in an inability to reliably and effectively improve subsequent control performance. Summary of the Invention

[0009] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a shape memory alloy soft driver control method based on data-driven modeling. It adopts a neural network model, combined with an improved nonlinear activation function and a gradient descent algorithm to improve the modeling accuracy and robustness of the hysteresis system, thereby effectively improving the control effect.

[0010] The object of the present invention can be achieved by the following technical solution: A shape memory alloy soft drive control method based on data-driven modeling, comprising the following steps:

[0011] S1. measuring and collecting input and output data of the shape memory alloy soft actuator as sample data;

[0012] S2. Design an asymmetric activation function with hysteresis characteristics, construct a neural network structure and select corresponding parameters, combine sample data and an improved gradient algorithm for iterative training, and introduce L1 regularization to produce a sparse model to obtain a theoretical model of the soft actuator;

[0013] S3. Inputting the input data corresponding to the current software driver into the software driver theoretical model, outputting corresponding simulation output data, and performing corresponding control on the software driver according to the simulation output data.

[0014] Furthermore, step S1 specifically involves building a data acquisition system to measure and collect input and output data of the shape memory alloy soft driver. The data acquisition system includes an NF bipolar power supply, which is respectively connected to an NI device and a shape memory alloy soft driver. The NI device is connected to a thermal imager and a displacement sensor. The thermal imager is used to measure the surface temperature of the soft driver in real time and transmit the temperature to the NI device.

[0015] The displacement sensor is used to measure the telescopic length of the shape memory alloy soft actuator and transmit the measurement to the NI device;

[0016] The NF bipolar power supply is used to apply voltage to the shape memory alloy soft actuator;

[0017] The NI device is used to control the voltage parameters of the NF bipolar power supply and implement data acquisition in LabVIEW to obtain input and output data of the shape memory alloy soft driver under different test conditions.

[0018] Furthermore, the input data of the shape memory alloy soft actuator includes wire diameter, load and applied voltage;

[0019] The output data of the shape memory alloy soft actuator include temperature and telescopic length.

[0020] Furthermore, the NI device specifically controls the voltage type, voltage amplitude and voltage frequency of the NF bipolar power supply.

[0021] Furthermore, the different test conditions include static test and dynamic test, and under the static test condition, the NF bipolar power supply applies a square wave to the shape memory alloy soft actuator;

[0022] Under the dynamic test conditions, the NF bipolar power supply applies a sine wave to the shape memory alloy soft actuator.

[0023] Furthermore, the step S2 specifically includes the following steps:

[0024] S21. Based on the delayed Relay operator, design a nonlinear activation function for the hysteresis system;

[0025] S22. Construct a fully connected feedforward neural network with two hidden layers, take the mean square error as the loss function, improve the structure of the neural network and select corresponding parameters;

[0026] S23. Design and improve the gradient descent algorithm, introduce L1 regularization to generate a sparse model, and iteratively train the fully connected feedforward neural network based on sample data to obtain the theoretical model of the soft drive.

[0027] Furthermore, the specific process of step S21 is as follows:

[0028] The expression of the delayed Relay operator is:

[0029]

[0030] Where y(t) is the output of the delay relay operator, u(t) is the input of the delay relay operator, ξ is the state of the delay relay operator, defined as ξ = y(t-1), and the threshold is (β, α). When the input is greater than α, the output of the delay relay operator is +1, and when the input is less than β, the output of the delay relay operator is -1.

[0031] Based on this, a nonlinear activation function f1 is designed for the hysteresis system. The expression of the designed activation function is:

[0032]

[0033] Among them, y is the output of the activation function, x is the input of the activation function, w1 and w2 are the input weights, and b1 and b2 are the input biases.

[0034] Furthermore, the specific process of step S22 is: selecting the temperature of the shape memory alloy soft driver before time t as the neural network input X t , the displacement change is used as the output Y of the neural network t The hysteresis characteristics of the shape memory alloy soft actuator are actually composed of the heating process and the cooling process. The neural network structure is modified, the input and output are preprocessed, and the output nodes are divided into the heating output Y tU And cooling output Y tD , after normalization, we get the normalized input X and output Y U 、Y D ;

[0035] The activation function of the first hidden layer selects the nonlinear activation function f1, and the second hidden layer selects the linear function f2. Let the hidden layer node be Z1=[Z 11 Z 12 Z 13 … Z 1n ],Z2=[Z 21 Z 22 Z 23 … Z 2n ], the hidden layer nodes Z1 and Z2 are expressed as:

[0036]

[0037]

[0038] Among them, w U1 、w D1 、b U1 、b D1 is the weight and bias between the input layer and the hidden layer, w U2 、w D2 、b U2 、b D2 are the weights and biases between the two hidden layers.

[0039] Furthermore, the specific process of designing the improved gradient descent algorithm in step S23 is as follows:

[0040] Determine the search direction for:

[0041]

[0042]

[0043] in, is the gradient, is the search direction of the previous step, β k Obtained by the Fletcher-Reeves formula;

[0044] Determine the optimization step size α through linear search min , that is, find α>0 so that Expand the function into a Taylor series and get the iterative formula:

[0045]

[0046] The difference expressions used for first-order and second-order differentials are:

[0047]

[0048]

[0049] Among them, δ→0;

[0050] To optimize the gradient algorithm optimization process, first give the gradient iteration initial value x0, the threshold ε>0, and calculate the gradient value k=0, repeat the following steps:

[0051] ①If ‖g k ‖<ε, stop iteration;

[0052] ②Calculate the step size and use the linear search algorithm to find

[0053] ③Update iteration point

[0054] ④Calculate the new gradient

[0055] ⑤Calculate the combination coefficient

[0056] ⑥Calculate the new gradient direction

[0057] ⑦ Let k = k + 1, and repeat step ② until step ① is satisfied.

[0058] Furthermore, the specific process of introducing L1 regularization to generate a sparse model in step S23 is as follows:

[0059] After regularization, the original loss function is converted from L(W) to λ is the regularization parameter. After denormalization, the simulation output y is:

[0060]

[0061]

[0062]

[0063]

[0064] f2()=x

[0065] Among them, w U3 、w D3 is the weight between the hidden layer and the output layer, b U3 、b D3 is the bias between the hidden layer and the output layer.

[0066] Compared with the prior art, the present invention has the following advantages:

[0067] 1. This invention targets the hysteresis characteristics of shape memory alloy soft actuators and designs a new nonlinear activation function based on a neural network model. By improving the node connection method between the hidden layer and the output layer and combining it with an improved gradient descent algorithm, a new neural network output method is realized. This can effectively improve the modeling accuracy and robustness of the hysteresis system, ensure the accuracy of the modeling, and have higher precision, adaptability, and generalization ability, thereby effectively improving the subsequent control effect.

[0068] 2. The present invention builds a data acquisition system to measure and collect the input and output data of the shape memory alloy soft driver, fully considering the possible influencing factors in the data acquisition process. By measuring the influence of wire diameter, voltage type, voltage amplitude, voltage frequency, and load on the temperature and contraction length of the shape memory alloy soft driver, the hysteresis characteristics of the shape memory alloy soft driver can be more accurately characterized, which is conducive to ensuring the accuracy of subsequent modeling.

[0069] 3. The present invention designs a nonlinear activation function based on the delay Relay operator. The activation function based on the hysteresis phenomenological characteristics can better describe the hysteresis nonlinear phenomenon of the shape memory alloy soft actuator.

[0070] 4. In the present invention, the neural network is a fully connected feedforward neural network, and the mean square error is taken as the loss function. The neural network structure is improved, the hysteresis phenomenon is divided into a heating process and a cooling process, the number of output nodes is determined by a one-to-many mapping relationship, and the improved gradient descent algorithm is used to iterate the weight bias, which can more accurately describe the hysteresis characteristics of the shape memory alloy soft drive and improve the accuracy and generalization ability of the hysteresis system modeling.

[0071] 5. The present invention introduces L1 regularization to generate a sparse model. The sparse model retains the features that contribute to the model and removes the features that do not contribute or contribute little to the model. It can effectively prevent the model from overfitting and further improve the model generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Schematic diagram of the method flow of the present invention;

[0073] Figure 2 Schematic diagram of the application process of the embodiment;

[0074] Figure 3 This is a structural diagram of the shape memory alloy soft driver in the embodiment;

[0075] Figure 4 This is a structural diagram of the hysteresis characteristic data acquisition system of the shape memory alloy soft driver in the embodiment;

[0076] Figure 5This is a schematic diagram of the delay Relay operator in the present invention;

[0077] Figure 6 This is a schematic diagram of the activation function principle of the delayed Relay operator improved in the present invention;

[0078] Figure 7 This is a diagram of the improved neural network structure in the present invention;

[0079] Explanation of the marks in the figure: 1. Driving module, 2. Silicone layer, 3. Gel wrapping layer, 4. Positive lead, 5. Negative lead, 11. Shape memory alloy wire, 12. Polyvinyl chloride elastic substrate, 13. Printed circuit fixing plate, 131. Positioning hole one, 132. Positioning hole two, 133. Positioning hole three, 134. Positioning hole four. DETAILED DESCRIPTION

[0080] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] Example

[0082] like Figure 1 As shown, a shape memory alloy soft actuator control method based on data-driven modeling includes the following steps:

[0083] S1. measuring and collecting input and output data of the shape memory alloy soft actuator as sample data;

[0084] S2. Design an asymmetric activation function with hysteresis characteristics, construct a neural network structure and select corresponding parameters, combine sample data and an improved gradient algorithm for iterative training, and introduce L1 regularization to produce a sparse model to obtain a theoretical model of the soft actuator;

[0085] S3. Inputting the input data corresponding to the current software driver into the software driver theoretical model, outputting corresponding simulation output data, and performing corresponding control on the software driver according to the simulation output data.

[0086] This embodiment applies the above technical solution, and the specific process is as follows Figure 2 As shown in FIG, it is divided into three parts: data acquisition, model building and model identification. In this embodiment, the shape memory alloy soft drive structure is as follows Figure 3As shown, it includes a driving module 1, a gel wrapping layer 2, a silicone layer 3, a positive lead 4, and a negative lead 5; the driving module 1 includes a shape memory alloy wire 11, a polyvinyl chloride elastic substrate 12, and a printed circuit fixing plate 13; the driving module 1 is encapsulated inside the gel wrapping layer 2; the gel wrapping layer 2 is encapsulated inside the silicone layer 3; the elastic substrate 12 is arranged on the middle plane of the driving module 1; there are two fixing plates 13, which are respectively arranged at both ends of the elastic substrate 12 and fixedly connected by bonding, and each fixing plate 13 is provided with two positioning holes; the positioning hole 131 is located on the middle plane of the fixing plate 13; the shape The shape memory alloy wire 11 begins at a positioning hole 131 in the fixing plate 13, passes through positioning hole 133 along the length of the elastic substrate 12, returns to positioning hole 134 in a U-shape, and then passes through positioning hole 132 along the length of the elastic substrate 12. The positive lead 4 is connected to the shape memory alloy wire 11 through positioning hole 131; the negative lead 5 is connected to the shape memory alloy wire 11 through positioning hole 132. The gel coating 2 is made of poly (N-isopropylacrylamide) hydrogel with a high heat dissipation ratio, a soft texture, and good thermal response performance; the silicone layer 3 is made of silicone with a soft texture and good mechanical properties. External electrical energy is introduced through the positive lead 4, and the negative lead 5 removes the external electrical energy. When energized, the shape memory alloy wire 11 heats up and changes in displacement.

[0087] Considering that the hysteresis curve of the shape memory alloy soft actuator is related to the wire diameter, input and external factors during data acquisition, in order to fully characterize the various parameters of the shape memory alloy soft actuator, it is necessary to comprehensively consider the influence of wire diameter, voltage type, voltage amplitude, voltage frequency and load on the temperature and contraction length of the shape memory alloy soft actuator. Therefore, this embodiment builds a data acquisition system, such as Figure 4 As shown, the system consists of a shape memory alloy soft actuator, an experimental stand, a laser displacement sensor, a thermal imager, a high-power bipolar power supply, and a National Instruments acquisition device. A Keyence LK-G150 laser displacement sensor is used to measure contraction length, and a Flytech thermal imager is used to measure temperature. Voltage is applied to the shape memory alloy soft actuator via an NF high-power bipolar power supply, and the National Instruments acquisition device is used for data acquisition and power supply parameter control. Data acquisition is implemented in LabVIEW. The shape memory alloy wire's energized length is 20 mm. During static testing, a square wave with a 50% duty cycle is applied to the shape memory alloy soft actuator. During dynamic testing, a sine wave is applied to the shape memory alloy soft actuator. The effects of wire diameter, voltage amplitude, voltage frequency, and load on the temperature and contraction length of the shape memory alloy soft actuator are studied.

[0088] To avoid the time delay of thermistors, this solution uses a thermal imager to measure the surface temperature of the soft actuator in real time. A laser displacement sensor is used to measure the telescopic length of the shape memory alloy soft actuator. A bipolar power supply is used to apply voltage to the shape memory alloy material. Parameters such as the power supply voltage and frequency are controlled using an NI acquisition device, and data acquisition is implemented in LabVIEW. Experiments were conducted with different fixed loads (20g and 50g weights), wire diameters (0.135mm and 0.2mm), voltage types (square wave and sine wave), voltage amplitudes (4V, 5V, and 6V), and voltage frequencies (0.025Hz, 0.05Hz, and 0.1Hz). Data from the displacement sensor and thermal imager were recorded for 40 seconds. This approach captured the input and output data of the shape memory alloy soft actuator under different test conditions.

[0089] After building the model, Figure 5 The figure shows the principle diagram of the delay Relay operator. Figure 6 The figure shows the principle diagram of the activation function improved by the delayed Relay operator in this scheme. Based on the delayed Relay operator, this scheme designs an activation function that is more suitable for hysteresis systems.

[0090] The expression of the delayed Relay operator is:

[0091]

[0092] y(t) is the output of the delayed relay operator, u(t) is the input of the delayed relay operator, ξ is the state of the delayed relay operator, which can be defined as ξ = y(t-1), and the threshold is (β, α). When the input is greater than α, the output of the delayed relay operator is +1, and when the input is less than β, the output of the delayed relay operator is -1.

[0093] Based on this, a nonlinear activation function f1() is designed for the hysteresis system. The expression of the designed activation function is:

[0094]

[0095] Where y is the output of the activation function, x is the input of the activation function, w1 and w2 are the input weights, and b1 and b2 are the input biases.

[0096] Figure 7 The figure shows the improved neural network structure in this scheme. A fully connected feedforward neural network with two hidden layers is defined, and the mean square error is used as the loss function. The structure of the neural network is improved and appropriate neural network parameters are selected. The temperature of the shape memory alloy soft actuator before time t is selected as the neural network input X. t , the displacement change is used as the output Y of the neural networkt The hysteresis characteristics of the shape memory alloy soft actuator are actually composed of the heating process and the cooling process. The neural network structure is modified, the input and output are preprocessed, and the output nodes are divided into the heating output Y tU And cooling output Y tD , after normalization, we get the normalized input X and output Y U 、Y D The activation function of the first hidden layer is the nonlinear activation function f1(), and the second hidden layer is the linear function f2(). Let the hidden layer node be Z1=[Z 11 Z 12 Z 13 … Z 1n ],Z2=[Z 21 Z 22 Z 23 … Z 2n ], the hidden layer nodes Z1 and Z2 can be expressed as:

[0097]

[0098]

[0099] Among them, w U1 、w D1 、b U1 、b D1 is the weight and bias between the input layer and the hidden layer, w U2 、w D2 、b U2 、b D2 are the weights and biases between the two hidden layers.

[0100] Then an improved gradient descent algorithm is used for iteration. The principle of the improved gradient descent algorithm is:

[0101] Determination of search direction, search direction It can be expressed as:

[0102]

[0103] in, is the gradient, is the search direction of the previous step, β k It can be obtained by the Fletcher-Reeves formula,

[0104] After determining the search direction, it is necessary to use linear search technology to determine the optimized step size α min , that is, find α>0 so that Expanding the function into a Taylor series, we can get the iterative formula:

[0105]

[0106] The difference expressions used for first-order and second-order differentials are:

[0107]

[0108]

[0109] Among them, δ→0.

[0110] To optimize the gradient algorithm optimization process, first give the gradient iteration initial value x0, the threshold ε>0, and calculate the gradient value k=0, repeat the following steps:

[0111] ①If ‖g k ‖<ε, stop iteration;

[0112] ②Calculate the step size and use the linear search algorithm to find

[0113] ③Update iteration point

[0114] ④Calculate the new gradient

[0115] ⑤Calculate the combination coefficient

[0116] ⑥Calculate the new gradient direction

[0117] ⑦ Let k = k + 1, and repeat step ② until step ① is satisfied.

[0118] In addition, to prevent the model from overfitting and improve the generalization ability of the model, L1 regularization is introduced to produce a sparse model; the sparse model retains the features that contribute to the model and removes the features that do not contribute or contribute little to the model. After regularization, the original loss function is converted from L(W) to λ is the regularization parameter. After denormalization, the simulation output y can be expressed as:

[0119]

[0120]

[0121]

[0122] where w U3 、w D3 is the weight between the hidden layer and the output layer, b U3 、b D3is the bias between the hidden layer and the output layer, f2()=x.

[0123] Therefore, the simulation output y can also be expressed as:

[0124]

[0125] In summary, this solution improves the structure and parameters of a common fully connected feedforward neural network model. By optimizing the node connections between the hidden and output layers, a new neural network output method is designed. An asymmetric nonlinear activation function, based on the delayed Relay operator, is designed to be more suitable for hysteresis models. An improved gradient descent algorithm is used to enhance the accuracy and generalization of hysteresis system modeling. This allows accurate simulation output data to be generated based on the resulting theoretical model of the soft actuator, effectively improving subsequent control effectiveness.

Claims

1. A shape memory alloy soft drive control method based on data driven modeling, characterized in that: The following steps are involved: S1. measuring and collecting input and output data of the shape memory alloy soft actuator as sample data; S2. Design an asymmetric activation function with hysteresis characteristics, construct a neural network structure and select corresponding parameters, combine sample data and an improved gradient algorithm for iterative training, and introduce L1 regularization to produce a sparse model to obtain a theoretical model of the soft actuator; S3, inputting the input data corresponding to the current software driver into the software driver theoretical model, outputting corresponding simulation output data, and controlling the software driver accordingly according to the simulation output data; Step S2 specifically includes the following steps: S21. Based on the delayed Relay operator, design a nonlinear activation function for the hysteresis system; S22. Construct a fully connected feedforward neural network with two hidden layers, take the mean square error as the loss function, improve the structure of the neural network and select corresponding parameters; S23. Design and improve the gradient descent algorithm, introduce L1 regularization to generate a sparse model, and iteratively train the fully connected feedforward neural network based on sample data to obtain the theoretical model of the soft drive.

2. The shape memory alloy soft driver control method based on data driven modeling according to claim 1, characterized in that: The step S1 specifically involves building a data acquisition system to measure and collect input and output data of the shape memory alloy soft driver. The data acquisition system includes an NF bipolar power supply, which is respectively connected to an NI device and the shape memory alloy soft driver. The NI device is connected to a thermal imager and a displacement sensor. The thermal imager is used to measure the surface temperature of the soft driver in real time and transmit the temperature to the NI device. The displacement sensor is used to measure the telescopic length of the shape memory alloy soft actuator and transmit the measurement to the NI device; The NF bipolar power supply is used to apply voltage to the shape memory alloy soft actuator; The NI device is used to control the voltage parameters of the NF bipolar power supply and implement data acquisition in LabVIEW to obtain input and output data of the shape memory alloy soft driver under different test conditions.

3. The shape memory alloy soft driver control method based on data driven modeling according to claim 2, characterized in that: The input data of the shape memory alloy soft actuator include wire diameter, load and applied voltage; The output data of the shape memory alloy soft actuator include temperature and telescopic length.

4. The shape memory alloy soft driver control method based on data driven modeling according to claim 2, characterized in that: The NI device specifically controls the voltage type, voltage amplitude and voltage frequency of the NF bipolar power supply.

5. The shape memory alloy soft driver control method based on data driven modeling according to claim 4 is characterized in that: The different test conditions include static test and dynamic test. Under the static test condition, the NF bipolar power supply applies a square wave to the shape memory alloy soft actuator; Under the dynamic test conditions, the NF bipolar power supply applies a sine wave to the shape memory alloy soft actuator.

6. The shape memory alloy soft driver control method based on data driven modeling according to claim 1, characterized in that: The specific process of step S21 is as follows: The expression of the delayed Relay operator is: Where y(t) is the output of the delay relay operator, u(t) is the input of the delay relay operator, ξ is the state of the delay relay operator, defined as ξ = y(t-1), and the threshold is (β, α). When the input is greater than α, the output of the delay relay operator is +1, and when the input is less than β, the output of the delay relay operator is -1. Based on this, a nonlinear activation function f1 is designed for the hysteresis system. The expression of the designed activation function is: Among them, y is the output of the activation function, x is the input of the activation function, w1 and w2 are the input weights, and b1 and b2 are the input biases.

7. The shape memory alloy soft driver control method based on data driven modeling according to claim 6, characterized in that: The specific process of step S22 is: selecting the temperature of the shape memory alloy soft driver before time t as the neural network input X t , the displacement change is used as the output Y of the neural network t The hysteresis characteristics of the shape memory alloy soft actuator are actually composed of the heating process and the cooling process. The neural network structure is modified, the input and output are preprocessed, and the output nodes are divided into the heating output Y tU and cooling output Y tD , after normalization, we get the normalized input X and output Y U 、Y D ; The activation function of the first hidden layer selects the nonlinear activation function f1, and the second hidden layer selects the linear function f2. Let the hidden layer node be Z1=[Z 11 Z 12 Z 13 …Z 1n ],Z2=[Z 21 Z 22 Z 23 …Z 2n ], the hidden layer nodes Z1 and Z2 are expressed as: Among them, w U1 、w D1 、b U1 、b D1 is the weight and bias between the input layer and the hidden layer, w U2 、w D2 、b U2 、b D2 are the weights and biases between the two hidden layers.

8. The shape memory alloy soft driver control method based on data driven modeling according to claim 7, characterized in that: The specific process of designing the improved gradient descent algorithm in step S23 is: Determine the search direction for: in, is the gradient, is the search direction of the previous step, β k Obtained by the Fletcher-Reeves formula; Determine the optimization step size α through linear search min , that is, find α>0 so that Expand the function into a Taylor series and get the iterative formula: The difference expressions used for first-order and second-order differentials are: Among them, δ→0; To optimize the gradient algorithm optimization process, first give the gradient iteration initial value x0, the threshold ε>0, and calculate the gradient value k=0, repeat the following steps: ①If ‖g k ‖<ε, stop iteration; ②Calculate the step size and use the linear search algorithm to find ③Update iteration point ④Calculate the new gradient ⑤Calculate the combination coefficient ⑥Calculate the new gradient direction ⑦ Let k = k + 1, and repeat step ② until step ① is satisfied.

9. The shape memory alloy soft driver control method based on data driven modeling according to claim 8, characterized in that: The specific process of introducing L1 regularization to generate a sparse model in step S23 is as follows: After regularization, the original loss function is converted from L(W) to λ is the regularization parameter. After denormalization, the simulation output y is: f2()=x Among them, w U3 、w D3 is the weight between the hidden layer and the output layer, b U3 、b D3 is the bias between the hidden layer and the output layer.

Images