Motion control method for shape memory alloy driving joint

Abstract

The invention discloses a motion control method for a shape memory alloy driving joint, which comprises the following steps: performing data acquisition on the shape memory alloy driving joint, establishing a feedforward neural network model as a basic skeleton, learning and approaching a complex nonlinear mapping relationship, and aiming at the core problem of the hysteresis characteristic of the shape memory alloy, establishing a model for the shape memory alloy driving joint. According to the method, an asymmetric nonlinear activation function is designed, the activation function is formed by combining two Sigmoid functions with different parameters, one function is specially used for simulating a positive process, the other function is used for simulating a negative process, and asymmetric and complex hysteresis characteristics are flexibly learned and represented by respectively setting bias parameters with independent weights; further, through a self-adaptive control method, inverse compensation is carried out on a nonlinear static link through a hysteresis compensator, compensation is carried out on a linear dynamic link through a dynamic controller, and a weight coefficient of the dynamic controller is dynamically updated based on an NLMS (Non-Least Mean Square) algorithm; and the positioning precision, the dynamic response speed and the system robustness of the shape memory alloy driving joint are improved.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a motion control method for a shape memory alloy driven joint. Background Technology

[0002] Joints based on shape memory alloys (SMAs) offer advantages over traditional motor joints, including higher power-to-weight ratio, lower driving voltage, no pollution, and no noise, enabling lightweight and dexterous robots. However, SMAs exhibit nonlinear characteristics such as hysteresis and creep, leading to lower joint drive control accuracy and slower response. Therefore, improving the model accuracy and robustness of SMA-driven joints, designing appropriate control methods, and enhancing drive performance are of significant value.

[0003] Existing control methods for robot joints driven by SMA have the following key drawbacks, making them unsuitable for requirements of high tracking accuracy, strong robustness, and low hysteresis: (1) Poor modeling accuracy: The theoretical model is complex, the parameters are difficult to identify and are prone to drift, resulting in a disconnect between simulation and actual application.

[0004] (2) Weak anti-interference ability: lacks state perception and feedback, cannot cope with load changes, and cannot compensate for SMA performance fluctuations.

[0005] (3) Hysteresis problem: The hysteresis model of SMA is prone to mismatch, which leads to a decrease in the inverse compensation effect. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a motion control method for shape memory alloy driven joints, which solves the problems of poor modeling accuracy, weak anti-interference ability, and mismatch of hysteresis characteristics in the existing SMA driven joint control methods.

[0007] The technical solution of this invention is: a motion control method for a shape memory alloy driven joint, comprising: S1. Apply voltage excitation signals of different waveforms, amplitudes and frequencies to the shape memory alloy used to drive the joint, and simultaneously collect displacement and temperature response data of the shape memory alloy under the voltage excitation signals; S2. Based on the collected response data, a feedforward neural network model is constructed to characterize the hysteresis properties of the shape memory alloy; the activation function of the hidden layer of the feedforward neural network model adopts an asymmetric hysteresis activation function; the asymmetric hysteresis activation function is composed of two independently set sigmoid functions, wherein the first sigmoid function corresponds to the positive process of applying voltage to the shape memory alloy to cause it to shrink, and the second sigmoid function corresponds to the negative process of turning off the voltage to cause the shape memory alloy to recover its original length; S3. Construct a Hammerstein model to describe the input-output characteristics of the shape memory alloy driven joint; the Hammerstein model includes a series of nonlinear static elements and a linear dynamic element; the feedforward neural network model established in step S2 to characterize the hysteresis characteristics of the shape memory alloy is used as the nonlinear static element, and its model parameters are identified using the Levenberg-Marquardt algorithm; the linear dynamic element is characterized using the FIR model, and its model parameters are identified using the NLMS algorithm; S4. For the Hammerstein model constructed in step S3, an adaptive inverse controller is constructed. The adaptive inverse controller includes a hysteresis compensator for inverse compensation of the nonlinear static element and a dynamic controller for compensation of the linear dynamic element. The hysteresis compensator is the inverse model of the feedforward neural network model, and the dynamic controller is an FIR filter. Based on the NLMS algorithm, the weight coefficients of the dynamic controller are dynamically updated according to the error between the actual output and the expected output of the shape memory alloy driven joint, so as to generate a control voltage signal applied to the shape memory alloy to complete the motion control.

[0008] Furthermore, in step S2, the feedforward neural network model is a network structure with a single hidden layer and m neurons, and its input-output relationship is expressed as follows:

[0009]

[0010] In the formula, This is the current input. This is the current output. It is the weighted output of the hidden layer neurons. and These represent the weights and biases of the hidden layer, respectively. and These are the weights and biases of the output layer, respectively. The activation function for the hidden layer is the asymmetric hysteresis activation function. It is a linear function.

[0011] Furthermore, the asymmetric hysteresis activation function Specifically, it is expressed as:

[0012] In the formula, This is the first Sigmoid function. This is the second Sigmoid function; specifically, it is expressed as:

[0013] In the formula, For the independent variable of the asymmetric hysteresis activation function, , For the weights and biases of the positive process, , Here, e represents the weights and biases of the negative process, and e is the natural constant.

[0014] Furthermore, in step S2, the input to the feedforward neural network model is a multi-scale moving average derived feature constructed from the response data; the multi-scale moving average derived feature includes a moving average feature, a standard deviation feature, and a dominant frequency feature.

[0015] Furthermore, in step S3, the linear dynamic element is represented using an FIR model. Specifically, it is expressed as:

[0016] In the formula, Here are the parameters to be identified in the FIR model, and L is the model order.

[0017] Furthermore, in step S4, a dynamic controller is used to compensate for the linear dynamic element. Specifically, it is expressed as:

[0018] In the formula, These are the weighting coefficients of the dynamic controller.

[0019] Furthermore, in step S4, the dynamic update process of the weight coefficients of the dynamic controller is specifically as follows:

[0020] In the formula, This is the weight vector at time n+1. Let be the weight vector at time n. To adjust the step size of the adaptive process speed, The actual output at time k=n With expected output The tracking error, p, is a small positive number to prevent the weight vector from becoming infinitely large. Given a double-filtered signal sequence, its elements To achieve the desired output The signal is obtained after processing by the hysteresis compensator. Then utilize the aforementioned linear dynamic link estimation model For signal Obtained by filtering.

[0021] Furthermore, in step S1, the amplitude range of the voltage excitation signal is 3-8V, and the frequency range is 0.025-0.1Hz.

[0022] The present invention also relates to a computer program product that, when executed by a processor, implements the steps of the method.

[0023] The advantages of this invention compared to the prior art are: (1) This invention innovatively proposes a control method for shape memory alloy driven joints. By acquiring data from the shape memory alloy driven joints, a feedforward neural network model is established as the basic skeleton to learn and approximate complex nonlinear mapping relationships. Addressing the core issue of the hysteresis characteristics of shape memory alloys, an asymmetric nonlinear activation function is specifically designed, consisting of two sigmoid functions with different parameters. One function is specifically used to simulate the positive process, and the other is used to simulate the negative process. Independent bias parameters are set for each function. , and , This makes it more suitable for hysteresis models, allowing for flexible learning and characterization of asymmetric and complex hysteresis characteristics. Furthermore, an adaptive control method is used to inversely compensate for the nonlinear static element through a hysteresis compensator and compensate for the linear dynamic element through a dynamic controller. Based on the NLMS algorithm, the weight coefficients of the dynamic controller are dynamically updated, which improves the positioning accuracy, dynamic response speed, and system robustness of the shape memory alloy driven joint.

[0024] (2) The method of the present invention effectively solves the limitations of the traditional hysteresis modeling method and has strong universality in practical applications. Especially under dynamic changes and complex excitation conditions, it can more accurately predict the hysteresis behavior of shape memory alloy wire. The adaptive control method adopted shows significant advantages in positioning accuracy, dynamic response speed and system robustness. Attached Figure Description

[0025] Figure 1 This is a structural diagram of the shape memory alloy driven joint of the present invention; Figure 2 This is a flowchart of the control method of the present invention; Figure 3 This is a diagram showing the composition of the shape memory alloy driven joint hysteresis characteristic testing device of the present invention; Figure 4 This is a schematic diagram of the activation function principle of the improved delayed Relay operator in this invention; Figure 5 This is a diagram of the feedforward neural network structure of the present invention; Figure 6 This is a diagram of the feedforward neural network structure with a single hidden layer in this invention; Figure 7 This is a schematic diagram of the adaptive inverse control principle of the present invention. Detailed Implementation

[0026] To better understand the technical solution of the present invention, the specific embodiments of the present invention are described below.

[0027] To better understand the method of this invention, a necessary introduction to the shape memory alloy driven joint structure is first given here, such as... Figure 1 As shown, the shape memory alloy driven joint structure provided in this embodiment mainly includes a shape memory alloy 1, a joint pivot 2, a lower leg link 3, a tension spring 4, and a thigh link 5. The lower leg link 3 and the thigh link 5 are connected via the joint pivot 2. One end of the shape memory alloy 1 is connected to the upper part of the lower leg link 3, and the other end is connected to the upper part of the thigh link 5. One end of the tension spring is connected to the lower part of the lower leg link 3, and the other end is connected to the lower part of the thigh link 5. In the initial state, the shape memory alloy 1 and the tension spring 4 are in their original lengths, and the shape memory alloy driven joint is in a contracted state. When the shape memory alloy driven joint extends, energizing the shape memory alloy 1 causes it to contract from its original length to its shortest length, driving the lower leg link 3 to rotate around the joint pivot 2. The tension spring 4 is passively stretched from its original length to its maximum length, generating tension. When the shape memory alloy drives the joint to contract, the power to the shape memory alloy 1 is cut off, and the shape memory alloy 1 gradually recovers from its shortest length to its original length; the tension spring 4 releases the tension, and the joint contracts from its maximum length to its original length, which drives the lower leg connecting rod 3 to rotate around the joint axis 2, and the shape memory alloy drives the joint to return to the contracted state.

[0028] Reference Figure 2 The diagram shown is a flowchart of the control method proposed in this invention, and the specific steps are as follows: (1) Data collection The hysteresis curve of shape memory alloy (SMI) driven joints is related to wire diameter, input, and external factors. To fully characterize the various parameters of SMI soft actuators, it is necessary to comprehensively consider the effects of voltage type, voltage amplitude, and voltage frequency on the temperature and contraction length of the SMI driven joint. To investigate the influence mechanism of different input signal excitation parameters on the hysteresis characteristics of SMI wires, an experimental platform for SMI driven joints was constructed, such as... Figure 3 As shown, the experimental platform mainly consists of an experimental support, an SMA software driver, a laser displacement sensor, a thermal imager, an NF bipolar power supply, and an NI data acquisition device. The dynamic response characteristics of shape memory alloy wires under different waveform excitation signals within the parameter range of amplitude 3-8V and frequency 0.025-0.1Hz were tested, and the voltage-displacement-temperature response data of the shape memory alloy wires were simultaneously acquired and recorded.

[0029] (2) Establishing a model In terms of model building, this invention constructs a neural network model based on an asymmetric hysteresis activation function. By combining the physical properties of the Relay operator with the nonlinear mapping capability of the neural network, an asymmetric hysteresis activation function is introduced to overcome the inherent defects of traditional activation functions in one-to-many mapping. At the same time, its asymmetric structure allows it to better describe the hysteresis characteristics in real-world systems.

[0030] As attached Figure 4 The diagram shown illustrates the neural network structure of this invention. The feedforward neural network structure with a single hidden layer is shown in the attached figure. Figure 5 A feedforward neural network with a single hidden layer and m neurons can be represented in the following form: (1) (2) in, This is the current input. This is the current output. It is the weighted output of the hidden layer neurons. and These represent the weights and biases of the hidden layer, respectively. and These represent the weights and biases of the output layer, respectively. is the activation function of the hidden layer, which in this example is an asymmetric hysteresis function. Typically, it is a linear function. Neural networks use different algorithms and cost functions to find the optimal weights and biases through multiple iterations, enabling the simulated output to fit the input.

[0031] Because the displacement-voltage relationship of shape memory alloy wire exhibits hysteresis, the input-output relationship of a neural network is a one-to-one mapping; for a given input, its output is also determined. This invention improves the activation function based on the Relay operator, such as... Figure 6 As shown, the hysteresis activation function is divided into positive and negative parts, each composed of two different sigmoid functions. The improved activation function produces a certain degree of hysteresis. Furthermore, for asymmetric hysteresis loops in practical systems, the activation function adopts an asymmetric structure, which can be expressed in the following form.

[0032] (3) in, , These correspond to the positive process (applied voltage, shape memory alloy shrinks) and negative process (disconnected voltage, shape memory alloy returns to its original length) of the asymmetric hysteresis activation function, respectively, and can be expressed in the following form: (4) in, , For the weights and biases of the positive process, , The weights and biases for the negative process.

[0033] Furthermore, this invention abandons the overly complex recurrent neural network structure and employs derived features calculated from multi-scale moving averages to overcome the limitations of feedforward neural networks in processing time series data. By introducing derived features based on multi-scale moving averages, the invention effectively captures and utilizes key features such as frequency and amplitude in time series data. The principle is as follows: 1) The original time series data is represented as follows: (5) in, for The raw data at that moment This represents the length of the time series.

[0034] 2) Multiple scales are defined as follows: (6) in, For the first The window length corresponding to each scale The number of scales.

[0035] 3) For each scale Calculate the following three derived features.

[0036] Based on multi-scale moving average characteristics: (7) Based on multi-scale standard deviation features: At the point of time Above, the window length is The moving standard deviation is expressed as: (8) in, It is a moving average within the window.

[0037] Based on multi-scale dominant frequency characteristics: Let the discrete Fourier transform within the window be: (9) The corresponding window length is At the point of time The main frequency is: (10) Where arg max represents the frequency corresponding to the maximum Fourier transform amplitude. The sampling frequency.

[0038] (3) Identification Model To improve the control performance of shape memory alloy wire actuators, the design of a feedforward controller involves two main steps: First, an accurate hysteresis model H(•) is established to describe the hysteresis characteristics of the controlled object, i.e., the feedforward neural network model established in step 2); second, this model is inverted to obtain the corresponding inverse hysteresis model H. 1 (•) performs inverse compensation on the hysteresis model. However, since feedforward control is an open-loop control, the control accuracy may decrease if there are external disturbances or instability of the controlled object in the system.

[0039] Therefore, the Hammerstein model is introduced, such as Figure 7 As shown, the model consists of a nonlinear static element and a linear dynamic element connected in series.

[0040] The input signal u of the nonlinear static link is nonlinearly transformed to generate an intermediate variable v, which is then passed to the linear dynamic link. The linear dynamic link generates the system output y based on v. The proposed AHFNN (Asymmetric Hysteresis Feedforward Neural Network) model models the nonlinear static link of the shape memory alloy wire hysteresis system. The parameters of the AHFNN model are identified using the Levenberg-Marquardt algorithm.

[0041] The linear dynamic element is characterized by the finite-length unit impulse response (FIR) model, expressed as: (11) Where v(k) is the input of the dynamic element, y(k) is the output of the dynamic element, and W(z) is the input of the dynamic element. -1 Let w0, ..., wL be the polynomials composed of delay elements, w0, ..., wL be the parameters to be identified in the model, called weight coefficients, and L be the order of the polynomial. The Normalized Least Mean Square (NLMS) algorithm is used to identify the parameters of the FIR linear dynamic model.

[0042] (4) Compensation control Design an adaptive inverse controller based on the FxNLMS algorithm. In the FxNLMS controller, the static hysteresis H(·) is obtained through the AHFNN inverse hysteresis model H. 1 (·) performs compensation, based on the controller C(z) of the FIR digital filter. 1 Compensation linear dynamic G(z) 1 ).like Figure 7As shown, x(k) is the reference displacement signal, i.e., the desired displacement waveform. y(k) is the actual displacement output waveform. n(k) is the noise interference. x(k) passes through the hysteresis compensator H 1 (·) After filtering, x is obtained h (k). x h (k) passes through C(z) 1 After compensation, the input voltage signal u(k) is obtained. This is a dynamic estimation model for a linear dynamic element, constructed based on an FIR filter. h (k) was The estimated dynamic filtering produces a filtered signal f(k). m (k) represents the difference between the reference signal x(k) and the actual output y(k). The adaptive controller effectively improves the convergence speed of the control algorithm by normalizing the filtered signal f(k) to suppress the influence of large-range amplitude fluctuations of the reference signal on the system.

[0043] Assumption: (12) The filtered reference signal x can be estimated. h (k).

[0044] If an FIR-based feedforward controller can accurately capture inverse dynamics (13) Based on FIR controller The coefficient iterative update algorithm aims to synchronously minimize the tracking error. The expected value is the objective function. Ideally, the actual displacement can effectively track the reference displacement, i.e.: (14) Dynamic estimation of linear dynamic elements For and dynamic compensator Each is composed of an FIR filter. (15) in, and These are the weight coefficients of the FIR filters, corresponding to the linear dynamic compensation and estimated dynamic model of the controlled object. L is the length of the two FIR filters.

[0045] These are fixed weighting coefficients, determined offline based on the controlled object model.

[0046] Adaptive weight coefficients Then, by minimizing the mean square value of the tracking error online... It is dynamically updated.

[0047] (16) Weight coefficients were implemented based on the NLMS algorithm. Update: The weight self-tuning process is as follows: (17) Where μ is the constant step size for adjusting the speed of the adaptive process, and p is a small positive number to prevent the weight vector from becoming infinitely large. It is a weight vector. Given a double-filtered signal sequence, its element f(k) can be represented as: (18) in, It is an estimation dynamic based on FIR. The elements of the weight vector.

[0048] It is understood that this invention has been described through embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of this invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific circumstances without departing from the spirit and scope of this invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are protected by this invention.

[0049] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A motion control method for a shape memory alloy driven joint, characterized in that, include: S1. Apply voltage excitation signals of different waveforms, amplitudes and frequencies to the shape memory alloy used to drive the joint, and simultaneously collect displacement and temperature response data of the shape memory alloy under the voltage excitation signals; S2. Based on the collected response data, a feedforward neural network model is constructed to characterize the hysteresis properties of the shape memory alloy; the activation function of the hidden layer of the feedforward neural network model adopts an asymmetric hysteresis activation function; the asymmetric hysteresis activation function is composed of two independently set sigmoid functions, wherein the first sigmoid function corresponds to the positive process of applying voltage to the shape memory alloy to cause it to shrink, and the second sigmoid function corresponds to the negative process of turning off the voltage to cause the shape memory alloy to recover its original length; S3. Construct a Hammerstein model to describe the input-output characteristics of the shape memory alloy driven joint; the Hammerstein model includes a series of nonlinear static elements and a linear dynamic element; the feedforward neural network model established in step S2 to characterize the hysteresis characteristics of the shape memory alloy is used as the nonlinear static element, and its model parameters are identified using the Levenberg-Marquardt algorithm; the linear dynamic element is characterized using the FIR model, and its model parameters are identified using the NLMS algorithm; S4. For the Hammerstein model constructed in step S3, an adaptive inverse controller is constructed. The adaptive inverse controller includes a hysteresis compensator for inverse compensation of the nonlinear static element and a dynamic controller for compensation of the linear dynamic element. The hysteresis compensator is the inverse model of the feedforward neural network model, and the dynamic controller is an FIR filter. Based on the NLMS algorithm, the weight coefficients of the dynamic controller are dynamically updated according to the error between the actual output and the expected output of the shape memory alloy driven joint, so as to generate a control voltage signal applied to the shape memory alloy to complete the motion control.

2. The motion control method for a shape memory alloy driven joint according to claim 1, characterized in that: In step S2, the feedforward neural network model is a network structure with a single hidden layer and m neurons, and its input-output relationship is expressed as follows: In the formula, This is the current input. This is the current output. It is the weighted output of the hidden layer neurons. and These represent the weights and biases of the hidden layer, respectively. and These are the weights and biases of the output layer, respectively. The activation function for the hidden layer is the asymmetric hysteresis activation function. It is a linear function.

3. The motion control method for a shape memory alloy driven joint according to claim 1 or 2, characterized in that: The asymmetric hysteresis activation function Specifically, it is expressed as: In the formula, This is the first Sigmoid function. This is the second Sigmoid function; specifically, it is expressed as: In the formula, For the independent variable of the asymmetric hysteresis activation function, , For the weights and biases of the positive process, , Here, e represents the weights and biases of the negative process, and e is the natural constant.

4. The motion control method for a shape memory alloy driven joint according to claim 1, characterized in that: In step S2, the input to the feedforward neural network model is a multi-scale moving average derived feature constructed from the response data; the multi-scale moving average derived feature includes a moving average feature, a standard deviation feature, and a dominant frequency feature.

5. The motion control method for a shape memory alloy driven joint according to claim 1, characterized in that: In step S3, the linear dynamic element is represented using the FIR model. Specifically, it is expressed as: In the formula, Here are the parameters to be identified in the FIR model, and L is the model order.

6. The motion control method for a shape memory alloy driven joint according to claim 5, characterized in that: In step S4, a dynamic controller is used to compensate for the linear dynamic element. Specifically, it is expressed as: In the formula, These are the weighting coefficients of the dynamic controller.

7. The motion control method for a shape memory alloy driven joint according to claim 6, characterized in that: In step S4, the dynamic update process of the weight coefficients of the dynamic controller is specifically as follows: In the formula, This is the weight vector at time n+1. Let be the weight vector at time n. To adjust the step size of the adaptive process speed, The actual output at time k=n With expected output The tracking error, p, is a small positive number to prevent the weight vector from becoming infinitely large. Given a double-filtered signal sequence, its elements To achieve the desired output The signal is obtained after processing by the hysteresis compensator. Then utilize the aforementioned linear dynamic link estimation model For signals Obtained by filtering.

8. The motion control method for a shape memory alloy driven joint according to claim 1, characterized in that: In step S1, the amplitude range of the voltage excitation signal is 3-8V and the frequency range is 0.025-0.1Hz.

9. A computer program product, characterized in that: When the computer program product is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images