Miniature DES magnetostrictive position control unit self-adaptive control method based on improved sliding mode algorithm

Abstract

The invention relates to a self-adaptive control method for a miniature DES (Data Encryption Standard) magnetostrictive position control unit based on an improved sliding mode algorithm, which solves the technical problem of how to improve the response speed and precision of the displacement of an electromagnetic driving flexible joint, and comprises a data feedback and processing part and an improved sliding mode control part. The method comprises the steps that a capacitive dielectric elastomer sensor is detected and filtered, current displacement is obtained after mapping, the displacement and expectation are input into a controller again for circulation, and improved synovial membrane control uses an RBF neural network to correct the tensile force of the dielectric elastomer sensor, a composite weight updating law composed of a pose state and a sliding mode surface is adopted to realize weight online optimization, and at the same time, delay interference compensation is used to further improve robustness and control precision. The method is suitable for executing the dynamic task in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic drive flexible joints, and more specifically, to an adaptive control method for a micro DES magnetostrictive position control unit based on an improved sliding mode algorithm. Background Art

[0002] In the field of robot joint technology, flexible joints have advantages such as good deformation ability, high safety, multiple degrees of freedom, and light weight, and can support robots to complete more difficult complex tasks.

[0003] Among many driving methods of flexible joints, electromagnetic drive has advantages such as high positioning accuracy, fast response speed, and compact structure. The flexible joint driven by electromagnetic can refer to the invention patent application with the application publication number of CN116787486A. However, the response speed and accuracy of the displacement of the existing electromagnetic drive flexible joint need to be improved. Summary of the Invention

[0004] The present invention is to solve the technical problem of how to improve the response speed and accuracy of the displacement of the electromagnetic drive flexible joint, and provides an adaptive control method for a micro DES magnetostrictive position control unit based on an improved sliding mode algorithm.

[0005] The improved sliding mode algorithm is a sliding mode control algorithm that uses delay interference compensation and RBF neural network for deep optimization.

[0006] In the data processing and feedback process, the controller provides an input quantity to the power supply unit, and the power supply unit outputs a specified voltage or current to the micro DES magnetostrictive position control unit. The micro DES magnetostrictive position control unit generates a driving force under the influence of electromagnetic force, drives the capacitive dielectric elastomer sensor to stretch, resulting in a change in the capacitance value of the capacitive dielectric elastomer sensor. The capacitance of the sensor is detected and filtered by a capacitance detection device to remove interference fluctuations, and then compared with the initial capacitance obtained in advance. After the data is filtered and mapped, the current displacement is obtained, and the displacement and the desired value are re-input to the controller for circulation.

[0007] The present invention discloses an adaptive control method for a micro DES magnetostrictive position control unit based on an improved sliding mode algorithm. The micro DES magnetostrictive position control unit includes a dielectric elastomer sensor and a magnetostrictive actuator. The dielectric elastomer sensor is provided with an upper part and a lower part. The upper part is coated on the magnetic ring fixing sleeve of the magnetostrictive actuator, and the lower part is coated on the support base of the magnetostrictive actuator. The closed-loop control method includes the following steps:

[0008] Step 1, use an LCR bridge to measure the capacitance value of the dielectric elastomer sensor multiple times, and take the average of the multiple measurement results to calculate the initial capacitance value C0;

[0009] Step 2, construct an RBF neural network and an improved sliding mode controller; the improved sliding mode controller is constructed using a nonlinear sliding surface as the basis and a nonlinear discrete terminal sliding mode surface:

[0010] s(k) = c 1 e 1 (k) + c 2 e 2 (k) + c 3 sigα(e 1 (k))

[0011] In the formula, s(k) represents the constructed sliding surface, e 1 (k), e 2 (k) represent the displacement error and velocity error respectively, c 1 , c 2 , c 3 are all constants;

[0012] Step 3, set the desired displacement;

[0013] Step 4, input the desired displacement into the improved sliding mode controller;

[0014] Step 5, the improved sliding mode controller outputs a signal to the power supply unit;

[0015] Step 6, the power supply unit supplies power to the coil of the magnetostrictive actuator;

[0016] Step 7, the displacement of the mover of the magnetostrictive actuator drives the dielectric elastomer sensor to stretch;

[0017] Step 8, the LCR bridge detects the capacitance value Cp1 of the dielectric elastomer sensor;

[0018] Step 9, filter the capacitance value Cp1 through a Kalman filter to obtain the capacitance value Cp2;

[0019] Step 10, subtract the initial capacitance value C0 from the capacitance value Cp2 to obtain the capacitance change value ΔCp;

[0020] Step 11, filter the capacitance change value ΔCp through a Kalman filter, and then obtain the current displacement x 1 ; The mapping process is carried out through the function y = kx. In the function, x represents the capacitance change value, y represents the displacement, and k represents a constant coefficient;

[0021] Step 12, the difference between the desired displacement and the current displacement x 1 is the displacement error e 1 (k). Input the displacement error e 1 (k) into the improved sliding mode controller;

[0022] Step 13, input the current displacement x 1 into the RBF neural network:

[0023]

[0024] wherein, represents the estimated value of the weight between the hidden layer and the output layer, which is a column vector; H j represents the output of the j-th hidden layer neuron, c j and b j respectively represent the center and width of the j-th neuron, x 1 represents the current displacement of the input;

[0025] obtain the corrected value of the sensor tension

[0026] Step 14, delay interference compensation. At time k, substitute the measurement data back to time k - 1 through the state equation of the system to estimate the interference d(k - 1) at the previous moment, and assume that the interference changes slowly in a short time, and regard d(k - 1) as the estimated value of the interference d(k) at the current moment; subsequently, add a compensation term to the control input;

[0027] Step 15, input the corrected value of the sensor tension and the interference estimated value obtained by delay interference compensation into the improved sliding mode controller together to obtain the output u(k). After the output u(k) is input into the power supply unit, then data feedback and processing are performed to obtain a new displacement value, completing the closed-loop control.

[0028] Preferably, the RBF neural network includes an input layer, a hidden layer, and an output layer. The centers of the hidden layer neurons are determined by the K-means clustering algorithm. The specific process is as follows:

[0029]

[0030] wherein, U j represents the j-th cluster, |U j | represents the number of samples in U j , u ji represents the sample in U j , D m represents the maximum distance between the clustering centers pairwise, and n represents the number of hidden layer neurons; first, preprocess the neural network using the dielectric elastomer sensor force-displacement data obtained from the tensile test. The samples for clustering are the sensor force-displacement data. Perform the K-means clustering algorithm on the sensor force-displacement data to obtain the key center points. Subsequently, use the obtained neuron center points and widths, and use the RBF neural network to fit the sensor force-displacement data to obtain the initial weights;

[0031] Compound weight update law:

[0032] ΔW j (k) = -η * c 2 T * s(k) * H j (x 1 (k))

[0033] In the formula, ΔW j (k) represents the weight change value at time k, η represents the learning rate, and c 2 is a constant, and T represents the sampling time of the discrete system;

[0034] While using the RBF neural network to fit the tensile force of the dielectric elastomer sensor, a compound weight update law composed of the pose state and the sliding mode surface is adopted to realize online weight optimization.

[0035] The beneficial effects of the present invention are quick response, precise positioning, strong robustness, and adaptation to the micro DES magneto-induced position control unit to perform dynamic tasks in complex environments.

[0036] An improved sliding mode control with interference compensation and an RBF neural network is adopted to design a nonlinear sliding mode surface to adapt to the nonlinear characteristics of the electromagnetic system. While using the RBF neural network to correct the tensile force of the dielectric elastomer sensor, a compound weight update law composed of the pose state and the sliding mode surface is adopted to realize online weight optimization, effectively suppressing the chattering phenomenon of the sliding mode algorithm. The coupled interference compensation technology effectively deals with the interference situations that cannot be confirmed in the model, further improving the control accuracy and robustness of the system.

[0037] It can not only achieve high-precision displacement control, but also has strong adaptive ability, can quickly and effectively respond to system dynamic changes and external interferences, thus significantly improving the control accuracy, response speed and overall stability.

[0038] Further features and aspects of the present invention will be clearly recorded in the following description of the specific embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a schematic structural diagram of the magneto-induced actuator;

[0040] Figure 2 is a schematic structural diagram of the micro DES magneto-induced position control unit;

[0041] Figure 3 is a schematic structural diagram of the dielectric elastomer sensor;

[0042] Figure 4 is the front view of the magneto-induced actuator;

[0043] Figure 5 isFigure 4 Cross-sectional view in the A-A direction in

[0044] Figure 6 It is a schematic structural diagram of the composition of the magnetostrictive actuator, the preparation of the dielectric elastomer sensor, and the connection of the dielectric elastomer sensor to the magnetostrictive actuator;

[0045] Figure 7 It is the general flowchart of the control method;

[0046] Figure 8 It is the data feedback and processing flowchart of the control method;

[0047] Figure 9 It is the flowchart of the adaptive control process based on the sliding mode algorithm;

[0048] Figure 10 It is the response curves of each algorithm for the step test;

[0049] Figure 11 It is the response curves of each algorithm for the instantaneous fault test;

[0050] Figure 12 It is the response curves of each algorithm for the 1.5g loaded sine test;

[0051] Figure 13 It is the response curves of each algorithm for the sine test with different loads.

[0052] Explanation of symbols in the figure:

[0053] 1. Magnetostrictive actuator, 2. Dielectric elastomer sensor, 2-1. Upper part, 2-2. Lower part, 3. Micro DES magnetostrictive position control unit, 4. Spherical connecting pair, 5. Magnetic ring fixing sleeve, 6. Coil, 7. Sleeve, 8. Support base, 9. Magnetic ring, 10. Magnetic column, 11. Magnetic block, 12. Buffer spring; 13. Is the protective layer, 14. Copper foil, 15. Electrode layer, 16. Dielectric layer. Specific implementation mode

[0054] The following refers to the accompanying drawings and further elaborates on the present invention with specific embodiments.

[0055] As Figure 1 shown, the magnetostrictive actuator 1 has a conventional structure. As Figure 2 shown, the dielectric elastomer sensor 2 is connected to the magnetostrictive actuator 1 to form the micro DES magnetostrictive position control unit 3. Figure 3 It is the schematic structural diagram of the dielectric elastomer sensor 2.

[0056] As Figure 4 and 5As shown in the figure, the magnetostrictive actuator 1 is composed of a stator, a mover, a spring housing, etc. The coil 6 is fixed above the support base 8 and is fixedly opposed to the sleeve 7 to form the stator. The inner diameter of the magnetic ring 9 is larger than the diameter of the magnetic column 10 to facilitate the assembly of the two. The magnetic ring 9, the magnetic column 10, and the magnetic block 11 are arranged in series in the same magnetic pole order and form the mover together with the magnetic ring fixing sleeve 5. The spherical connecting pair 4 serves as the upper fixing point, and preferably a ball pair connection is used, and other connection methods can also be selected. The diameter of the lower circular end face of the spherical connecting pair 4 and the inner diameter of the upper part of the magnetic ring fixing sleeve 5 form an interference fit. The inner diameter of the upper part of the magnetic ring fixing sleeve 5 is slightly larger than the outer diameter of the magnetic ring 9, and the diameter of the through hole in the lower part of the magnetic ring fixing sleeve 5 is slightly smaller than the inner diameter of the magnetic ring 9 and forms an interference fit with the magnetic column 10 at the same time. The magnetic column 10 sequentially passes through the through hole in the lower part of the magnetic ring fixing sleeve 5 and the inner hole of the magnetic ring 9 according to the magnetic pole assembly method and contacts the lower end face of the spherical connecting pair 4 together with the upper end face of the magnetic ring 9. Selecting the cooperation between the magnetic ring 9 and the magnetic column 10 can significantly improve the electromagnetic driving force. There is a blocking ring at the top of the sleeve 7, and its diameter is larger than the diameter of the magnetic column 10. At the same time, the diameter of the magnetic block 11 is larger than the diameter of the blocking ring at the top of the sleeve 7, which plays a blocking role to prevent the mover from separating from the stator. There is a certain gap between the magnetic block 11 and the inner wall of the sleeve 7 to ensure that the mover composed of the spherical connecting pair 4, the magnetic ring fixing sleeve 5, the magnetic ring 9, the magnetic column 10, and the magnetic block 11 can move freely in the sleeve 7. Heat dissipation grooves are provided on the wall of the sleeve 7 to prevent heat accumulation. At the same time, a buffer spring 12 with a relatively large stiffness coefficient is fixed at the bottom of the sleeve 7 to reduce the starting current and prevent the impact of falling during a fault. The outer diameter of the sleeve 7 is slightly larger than the outer diameter of the coil 6, and the inner diameter is slightly smaller than the inner diameter of the coil 6, forming a clearance fit with the coil. At the same time, a part of the coil 6 is led out and connected to the power supply unit. At the same time, the sleeve 7 and the support base 8 form an interference fit. The support base 8 can form a fixed structure with bolts and nuts, or different connection methods can be selected. There is no rigid constraint similar to a bearing between the stator and the mover, but it is driven by the interaction of the magnetic fields between the coil and the magnetic block, which can not only ensure the connection between the stator and the mover but also facilitate maintenance and disassembly.

[0057] The materials used for the magnetic ring 9, the magnetic column 10, and the magnetic block 11 can be permanent magnetic materials, including rare earth permanent magnetic materials (neodymium iron boron Nd2Fe14B), samarium cobalt (SmCo), alnico (AlNiCo), ferrite permanent magnetic materials, etc. The spherical connecting pair 4, the magnetic ring fixing sleeve 5, the sleeve 7, and the support base 8 can be manufactured by using stereolithography 3D printing technology or machining of non-magnetic materials.

[0058] The sensor part adopts a capacitive dielectric elastomer sensor (DES), such as Figure 6The dielectric elastomer sensor 2 includes a dielectric layer 16, an electrode layer 15 on both sides of the dielectric layer, a copper foil 14 located on the electrode layer for connection and output, and a protective layer 13 arranged on the side of each electrode layer away from the dielectric layer. Specifically, the preparation process is: first, a protective layer 13 is laid, and then a copper foil 14 is added on the first protective layer 13, and then an electrode layer 15 is laid, and the electrode layer adopts a duplex shape, in which the bottom of the I-shaped at one end contacts the copper foil added for the first time; the dielectric layer is laid; the copper foil 14 is added for the second time, and the copper foil overlaps with the bottom of the I-shaped at the other end; the electrode layer 15 is laid for the second time, and the copper foil 14 contacts with the bottom of the I-shaped at the other end; the protective layer is laid; the sensor can be packaged and integrated with the driver after being cut along the mold. The copper foil 14 can be extended by welding other wires. The protective layer 13 is based on platinum silicone, and different materials can be added according to different requirements to achieve different protection effects. The electrode layer 15 can be made of carbon black dispersion, carbon nanotube dispersion, conductive silver paste and other materials, and different materials can be selected according to the tensile and conductive properties. The material of the dielectric layer 16 can be different polymer materials, which can be flexibly adjusted according to different capacitance change ranges.

[0059] The dielectric elastomer sensor 2 is coated on the magnetostrictive actuator 1, referring to Figure 2 and 3 The upper part 2-1 of the dielectric elastomer sensor 2 is coated on the magnetic ring fixing sleeve 5, and the lower part 2-2 is coated on the supporting base 8. Thus, a micro DES magnetically induced position control unit 3 is formed.

[0060] The control algorithm disclosed in this application mainly includes two core modules: a data feedback and processing part and an improved synovial control part. In the data processing and feedback link, the controller first provides an accurate input to the power supply unit. The power supply unit can be a programmable DC power supply or a PCB power supply system. The power supply unit outputs a specified voltage or current to the coil 6 according to the input. The coil 6 generates an electromagnetic force that interacts with the permanent magnet to generate a corresponding driving force. The mover moves, and the magnetic ring fixing sleeve 5 moves to drive the dielectric elastomer sensor 2 to stretch, causing the area of ​​the dielectric elastomer sensor to change, thereby causing the capacitance value of the sensor to change. In order to accurately obtain the capacitance value of the sensor, a high-precision capacitance detection device, such as an LCR bridge meter, is used to detect and transmit the capacitance of the sensor in real time. At the same time, software filtering is used to ensure the reliability of the data for environmental noise and interference fluctuations. The software filtering uses the Kalman filter of a one-dimensional static system:

[0061]

[0062] In formula (1), represents the prior state estimate at time k, represents the posterior state estimate at time k-1, that is, the measured value after filtering at time k-1, represents the prior error covariance matrix at time k, P k-1 represents the posterior error covariance matrix at time k-1, Q represents the covariance of the process noise, K k represents the Kalman gain at time k, which is used to balance prediction and observation. R represents the observation noise, and this parameter can be adjusted to control the filtering effect. z k represents the measurement value of the sensor.

[0063] Reference Figure 7 、 8 、9, the adaptive control method of the micro DES magnetostrictive position control unit based on the improved sliding mode algorithm disclosed by the present invention mainly includes the following steps:

[0064] Step 1, use an LCR bridge to measure the capacitance value of the dielectric elastomer sensor 2 multiple times, and take the average of the multiple measurement results to calculate the initial capacitance value C0.

[0065] Step 2, construct an RBF neural network and an improved sliding mode controller.

[0066] During the stretching process of the dielectric elastomer sensor 2 driven by the mover displacement of the magnetostrictive actuator 1, the dielectric elastomer sensor will exert an additional pulling force on the magnetostrictive actuator. Therefore, an RBF neural network is used to correct the sensor pulling force. The RBF neural network adopted is a three-layer structure, including an input layer, a hidden layer, and an output layer. The centers of the hidden layer neurons are determined by the K-means clustering algorithm. Initialize the clusters and centers, and use the Euclidean distance between the samples and the center points to repeatedly adjust the clusters and centers until the clusters and centers no longer change, then the determination of each center can be completed. The specific process is as follows:

[0067]

[0068] In the formula, U j represents the jth cluster, |U j | represents the number of samples in U j , u ji represents the sample in U j , D m represents the maximum distance between the clustering centers pairwise, and n represents the number of neurons in the hidden layer. First, use the force-position data of the dielectric elastomer sensor obtained through the stretching test to preprocess the neural network. The samples for clustering are the force-position data of the sensor. Perform the K-means clustering algorithm on the force-position data of the sensor to obtain the key center points. Subsequently, use the obtained neuron center points and widths, and use the RBF neural network to fit the force-position data of the sensor to obtain the initial weights.

[0069] Compared with the weight iteration method based on the gradient descent method in the traditional RBF neural network, the present invention designs a composite weight update law based on the pose state and the sliding mode surface. By combining the real-time pose state of the system with the sliding mode surface information, this update law dynamically adjusts the network weights, achieving online optimization of the RBF neural network during the control process. This method not only avoids the gradient problem that the traditional gradient descent method may fall into, but also significantly reduces the sliding mode chattering phenomenon, better adapting to the control requirements of complex nonlinear systems. The specific update law is as follows:

[0070] ΔW j (k)= -η*c 2 T*s(k)*H j (x 1 (k))(3)

[0071] In the formula, ΔW j (k) represents the weight change value at time k, η represents the learning rate, and c 2 is a constant, and T represents the sampling time of the discrete system.

[0072] An improved sliding mode controller is constructed. Using the nonlinear sliding mode surface as the basis, while using the RBF neural network to fit the tensile force of the dielectric elastomer sensor, a composite weight update law composed of the pose state and the sliding mode surface is adopted to realize online weight optimization, effectively suppressing the chattering phenomenon of the sliding mode algorithm. The coupled disturbance compensation technology effectively deals with the unconfirmed disturbance situation during the control process, further improving the control accuracy and robustness of the system. It is constructed using a more suitable nonlinear discrete terminal sliding mode surface:

[0073] s(k)= c 1 e 1 (k)+ c 2 e 2 (k)+ c 3 sigα(e 1 (k)) (4)

[0074] In the formula, s(k) represents the constructed sliding mode surface, and e 1 (k), e 2 (k) represent the displacement error and the velocity error respectively, and c 1 , c 2 , c 3 are all constants.

[0075] Step 3, set the desired displacement;

[0076] Step 4, input the desired displacement into the improved sliding mode controller,

[0077] Step 5, the improved sliding mode controller outputs a signal to the power supply unit,

[0078] Step 6, the power supply unit supplies power to the coil of the magnetostrictive actuator 1;

[0079] Step 7, the displacement of the mover of the magnetostrictive actuator 1 drives the dielectric elastomer sensor 2 to stretch.

[0080] Step 8, the LCR bridge detects the capacitance value Cp1 of the dielectric elastomer sensor,

[0081] Step 9, the capacitance value Cp1 is filtered through a Kalman filter to obtain the capacitance value Cp2.

[0082] Step 10, the capacitance change value ΔCp is obtained by subtracting the initial capacitance value C0 from the capacitance value Cp2

[0083] Step 11, the capacitance change value ΔCp is filtered through a Kalman filter and then mapped to obtain the current displacement x 1 . The mapping process is carried out through the function y = kx. In the function, x represents the capacitance change value, y represents the displacement, and k represents a constant coefficient; let x = ΔCp, and the y value can be calculated, and the y value is the current displacement x 1 .

[0084] Step 12, the difference between the desired displacement and the current displacement x 1 is the displacement error e 1 (k), and the displacement error e 1 (k) is input into the improved sliding mode controller.

[0085] Step 13, the RBF neural network adopted by the present invention takes the current displacement of the controlled object as the input and outputs the correction value of the sensor tension

[0086]

[0087] In the formula, represents the estimated value of the weight between the hidden layer and the output layer, which is a column vector; H j represents the output of the j-th hidden layer neuron, c j and b j respectively represent the center and width of the j-th neuron, and x 1 represents the input current displacement.

[0088] Step 14, the present invention uses the delayed disturbance compensation technology to further improve the control effect. The delayed disturbance compensation technology is a technology used in a control system to cancel unknown disturbances, and it uses historical data to estimate and compensate for the disturbances at the current moment. The delayed disturbance compensation has the advantages of simple implementation and small computational complexity, and is suitable for the application scenarios with high real-time requirements of the present invention. The specific implementation method is:

[0089] At time k, the measured data is substituted back to time k-1 through the state equation of the system to estimate the disturbance d(k-1) at the previous moment. Assuming that the disturbance changes slowly in a short period of time, d(k-1) is regarded as the estimated value of the disturbance d(k) at the current moment (time k); subsequently, a compensation term is added to the control input to offset the influence of the disturbance on the system. The delayed disturbance compensation technology can effectively reduce the influence of unknown disturbances, such as external perturbations and model uncertainties, on the system performance by real-time estimating and compensating the disturbance, and improve the control accuracy and robustness of the system.

[0090] Step 15, the corrected value of the sensor tension and the disturbance estimated value obtained by delayed disturbance compensation are input into the improved sliding mode controller together to obtain the output u(k). After the output u(k) is input into the power supply unit, then data feedback and processing are performed to obtain a new displacement value, completing one closed-loop control.

[0091] To ensure the performance of the closed-loop system, the closed-loop and steady-state performance of the control algorithm are analyzed in this application by using the Lyapunov method and the scalar dynamic system. The Lyapunov method is an important mathematical tool for analyzing the stability of a system. Its core idea is to judge the stability of the system by constructing an energy function (called the Lyapunov function). The Lyapunov function V(k) is a scalar function that satisfies V(k)≥0. By analyzing the time derivative V'(k) of this function, the stability of the system can be judged: if V'(k)<0, the system is stable; the V(k) selected in this application is as follows:

[0092]

[0093] This invention adopts a discrete system. By discarding the high-order terms, ΔV(k) can be obtained as follows:

[0094]

[0095] In the formula, δ * represents the upper bound of the absolute value of the first derivative of the disturbance.

[0096] Substitute the composite weight update law of the RBF neural network, and take the robust term σ>δ * to ensure the stability of the system. Consider the scalar dynamic system:

[0097] z(k + 1) = z(k) - l 1 sigαz(k) - l 2 z(k) + g(k) (8)

[0098] If l 1 >0, 0<l 2If \(0 < \alpha < 1\) and there exists \(g(k)<\gamma\) where \(\gamma>0\), then the state \(z(k)\) is always bounded and there exists a finite number \(K>0\) such that: * \[

[0099]

[0100] where the function \(\Psi(\alpha)\) is defined as:

[0101]

[0102] Through analysis, it can be obtained that:

[0103]

[0104] Furthermore, substituting into the scalar dynamic system gives:

[0105]

[0106] To obtain the highest precision, \(\alpha = 2 / 3\) should be chosen, and the precision of the tracking error \(e(k)\) is \(O(T)\). Compared with the nonlinear sliding mode \(O(T)\), the improved sliding mode algorithm using RBF neural network and delay interference compensation has higher steady-state precision. 1 \[ 3 \] 2 The above control method can achieve rapid response, accurate positioning, and strong robustness, and is suitable for the micro DES magnetostrictive position control unit to perform dynamic tasks in complex environments.

[0107] To further verify the performance of the proposed algorithm, an experimental platform is built using an LCR bridge, a signal generator, a power amplifier, a laser displacement sensor, and a PC to conduct experimental verification.

[0108] The experimental steps are as follows:

[0109] Step 1: Connect the power supply system. Use a BNS to USB cable to connect the PC and the signal generator, use a BNS cable to connect the signal generator and the power amplifier, and confirm that the power amplifier uses an external signal. Connect the power amplifier to the coil of the micro magnetostrictive position control unit.

[0110] Step 2: Connect the detection system. Use a USB cable to connect the PC and the bridge, and ensure that the bridge is connected to the sensor of the micro magnetostrictive position control unit; at the same time, use a USB cable to connect the PC and the laser displacement sensor, and then use a DC regulated power supply to supply power to the laser displacement sensor.

[0111] \[

[0112] Step 3: Confirm the magnification ratio. Adjust the signal generator to generate a signal, and adjust the magnification ratio of the power amplifier. Fix a suitable magnification ratio for subsequent experiments.

[0113] Step 4: Run the test program. Test the communication between the PC, the signal generator, the bridge instrument, and the laser displacement sensor to ensure that the PC can control the output voltage and receive the feedback data from the bridge instrument and the laser displacement sensor.

[0114] Step 5: Conduct the formal experiment. Adjust the sampling time to 0.05 s, run the experimental program, start the experiment, and record the results.

[0115] The experiment adopted step signal testing, instantaneous fault simulation, and sine signal testing under different loads, and compared the specific performance differences of four algorithms: the fast terminal sliding mode algorithm optimized by RBF neural network and delay interference compensation (D-RBF-FTSM), the fast terminal sliding mode algorithm optimized only by RBF neural network (RBF-FTSM), the fast terminal sliding mode algorithm without optimization (FTSM), and the PID algorithm.

[0116] The step test adopted an 8 mm step signal, and the response curves of each algorithm are as Figure 10 shown. In addition to the rise time and settling time, the time period from 15 s to 18 s was selected to evaluate the maximum error, average error, and root mean square error. The criteria are as follows:

[0117] Max Error=max(e 1 (k))

[0118]

[0119] The index data is shown in Table 1. Among them, D-RBF-FTSM has the fastest settling time of 2.65 s, and at the same time has the lowest maximum error of 0.0419 mm and the lowest average error of 0.0150 mm, showing the best comprehensive performance, followed by RBF-FTSM.

[0120]

[0121] Table 1 Performance comparison of each algorithm in step test

[0122] To further compare the fault tolerance of each algorithm, a simulated fault experiment was conducted. Based on the 8 mm step signal, an instantaneous open circuit fault was simulated at 17 s, and the response of each algorithm was observed. The response of each algorithm is as Figure 11 shown.

[0123] In the fault response experiment, the recovery time, overshoot, average error from 18 s to 25 s, root mean square error, and average error from 21 s to 25 s were selected as the evaluation criteria.

[0124] Its index data is shown in Table 2. Combining Figure 11 with the average error in the two time periods in Table 2, it can be seen that after the power supply is restored, D-RBF-FTSM quickly drops from the overshoot state to the steady state. Its average error during the entire 18 - 25 s is 0.1463 mm, while the average error from 21 - 25 s drops to 0.0359 mm, dropping to the above-mentioned 8 mm step test magnitude. It demonstrates excellent fault tolerance.

[0125]

[0126] Table 2 Comparison of the performance of each algorithm in the instantaneous fault test

[0127] To test the dynamic performance of each algorithm, the experiment set x r = sin(0.05 * T)+5 signal for testing. At the same time, in order to deeply evaluate the robustness of each algorithm, no-load, 0.5 g load, 1 g load, and 1.5 g load were set for testing. The time period from 15 s to 140 s was selected for evaluation. In addition to the maximum error, average error, and root mean square error, phase difference and delay were additionally set to evaluate the tracking ability of the algorithm. At the same time, the rolling standard deviation was adopted to measure the chattering degree of the sliding mode algorithm.

[0128]

[0129] Δθ = arg(X[it]) - arg(Y[ie])

[0130]

[0131] The standards for phase difference, delay, and rolling standard deviation are as shown above. Among them, x, y, it, and ie represent the expected signal, the feedback signal of the sensor, and their corresponding frequency components respectively. avg represents taking the average, K represents different nodes, one node is taken every 30 in the data sample, and μ represents the average value of the samples between two nodes. Selecting the 1.5 g load for specific analysis, the response curves of each algorithm are as Figure 12 shown, and its evaluation index data is shown in Table 3.

[0132]

[0133] Table 3 Comparison of the performance of each algorithm in the 1.5 g load sine test

[0134] Combining Figure 12From the data in Table 3, it can be seen that although the FTSM algorithm has strong robustness, its Astd error index of 0.1054 mm far exceeds that of the other three algorithms, revealing a severe jitter phenomenon. On the contrary, although the RBF-FTSM significantly suppresses jitter, it loses its robustness, and its phase difference of -0.2926 rad and delay of -2.75 s are much higher than those of the FTSM algorithm and the D-RBF-FTSM algorithm. For the D-RBF-FTSM algorithm, all indicators are better than the other three algorithms, especially the average error of 0.1642 mm, Astderror of 0.0548 mm, and phase difference of -0.0747 rad, demonstrating excellent robustness.

[0135] The responses of the other three groups of on-load experiments are as Figure 13 shown, and the evaluation data are shown in Table 4.

[0136]

[0137] Table 4 Comparison of the performance of each algorithm under different loads

[0138] Experiments have proved that the RBF neural network used in the present invention effectively suppresses the jitter phenomenon of the sliding mode by using the composite weight update rate, and at the same time uses delay interference compensation to ensure the robustness of the control system, improving the response speed and control accuracy of the system.

[0139] The above description is only for the preferred embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications.

Claims

1. An adaptive control method for a micro DES magnetic position control unit based on an improved synovial algorithm, characterized in that: The micro DES magnetostrictive position control unit comprises a dielectric elastomer sensor and a magnetostrictive actuator. The dielectric elastomer sensor is provided with an upper part and a lower part. The upper part is coated on a magnetic ring fixing sleeve of the magnetostrictive actuator, and the lower part is coated on a supporting base of the magnetostrictive actuator. The closed-loop control method comprises the following steps: Step 1, using an LCR bridge meter to measure the capacitance value of the dielectric elastomer sensor multiple times, taking an average value of the multiple measurement results to calculate an initial capacitance value C0; Step 2, construct the RBF neural network and the improved sliding film controller; the improved sliding film controller is constructed by using the nonlinear sliding film surface as the basis and the nonlinear discrete terminal sliding mode surface: s(k)=c1e1(k)+c2e2(k)+c3sigα(e1(k)) In the formula, s(k) represents the constructed sliding surface, e1(k) and e2(k) represent the displacement error and velocity error respectively, and c1, c2, and c3 are all constants; Step 3, setting the expected displacement; Step 4, inputting the desired displacement to the improved synovial controller; Step 5, the improved synovial controller outputs a signal to the power supply unit; Step 6, the power supply unit supplies power to the coil of the magnetostrictive actuator; Step 7, the displacement of the mover of the magnetostrictive actuator drives the dielectric elastomer sensor to stretch; Step 8, using an LCR bridge meter to detect the capacitance value Cp1 of the dielectric elastomer sensor; Step 9, filtering the capacitance value Cp1 through a Kalman filter to obtain a capacitance value Cp2; Step 10, subtracting the initial capacitance value C0 from the capacitance value Cp2 to obtain a capacitance change value ΔCp; Step 11, filtering the capacitance change value ΔCp through a Kalman filter, and then mapping to obtain the current displacement x1; the mapping process is performed through the function y=kx, in which x represents the capacitance change value, y represents the displacement, and k represents that the coefficient is a constant value; Step 12, the difference between the expected displacement and the current displacement x1 is the displacement error e1(k), and the displacement error e1(k) is input into the improved sliding film controller; Step 13, input the current displacement x1 into the RBF neural network: In the formula, Represents the estimated value of the weight between the hidden layer and the output layer, which is a column vector; H j represents the output of the jth hidden layer neuron, c j With b j They represent the center and width of the jth neuron respectively, and x1 represents the current displacement of the input; Get the correction value of the sensor tension Step 14, delayed disturbance compensation, at time k, the measured data is substituted back to time k-1 through the system state equation to estimate the disturbance d(k-1) at the previous moment, and assuming that the disturbance changes slowly in a short time, d(k-1) is regarded as the estimated value of the disturbance d(k) at the current moment; then, the compensation term is added to the control input; Step 15: Correct the sensor tension The interference estimation value obtained by delayed interference compensation is input into the improved sliding membrane controller together to obtain the output u(k). After the output u(k) is input into the power supply unit, the data is fed back and processed to obtain a new displacement value to complete the closed-loop control.

2. The adaptive control method of a micro DES magnetic position control unit based on an improved synovial algorithm according to claim 1 is characterized in that: The RBF neural network includes an input layer, a hidden layer, and an output layer. The center of the hidden layer neurons is determined by the K-means clustering algorithm. The specific process is: Where U j represents the jth cluster, |U j | indicates U j The number of internal samples, u ji Indicates U j Samples within, D m represents the maximum distance between cluster centers, and n represents the number of neurons in the hidden layer. First, the force data of the dielectric elastomer sensor obtained through the tensile test is used to preprocess the neural network. The sample to be clustered is the sensor force data. The sensor force data is clustered using the K-means clustering algorithm to obtain the key center point. Then, the obtained neuron center point and width are used to fit the sensor force data using the RBF neural network to obtain the initial weight. Perform the composite weight update law: ΔW j (k)=-η*c2T*s(k)*H j (x1(k)) In the formula, ΔW j (k) represents the weight change value at time k, η represents the learning rate, c2 is a constant, and T represents the sampling time of the discrete system; While the RBF neural network is used to fit the tension of the dielectric elastomer sensor, the composite weight update law composed of posture state and sliding surface is used to realize online optimization of weight.

Images