Bionic eyeball based on co-sited integrated des magnetic parallel joint and control method

Abstract

The application relates to a bionic eyeball based on a co-sited integrated DES magnetic parallel joint and a control method, which solves the technical problem of how to miniaturize the bionic eyeball and enable high-precision multi-degree-of-freedom movement, and comprises an eyeball body, a lower base plate, an upper cover plate, a first DES magnetic driving device, a second DES magnetic driving device and a third DES magnetic driving device, the bottom of the second DES magnetic driving device is connected with the lower base plate through a rotary pair, the output end of the second DES magnetic driving device is connected with the upper cover plate through a spherical pair, the bottom of the first DES magnetic driving device is connected with the lower base plate through a rotary pair, the output end of the first DES magnetic driving device is connected with the upper cover plate through a spherical pair, the bottom of the third DES magnetic driving device is connected with the lower base plate through a rotary pair, the output end of the third DES magnetic driving device is connected with the upper cover plate through a spherical pair, and the eyeball body is connected with the upper cover plate.

Description

Technical Field

[0001] This invention relates to the field of bionic eye technology, and more specifically, to a bionic eye based on a co-located integrated DES magnetostrictive parallel joint and a control method thereof. Background Technology

[0002] As the core execution and perception carrier of active vision systems, bionic eyes hold significant application prospects in fields such as intelligent manufacturing, medical assistance, and human-computer interaction. Active vision emphasizes rapidly redirecting gaze and stably acquiring critical information in dynamic environments, which places comprehensive demands on the miniaturization, multi-degree-of-freedom motion, high dynamic response, and high-precision closed-loop control of bionic eyes. How to achieve efficient collaborative integration of actuation, perception, and control within limited volume and energy consumption constraints has become a key research focus for researchers in this field.

[0003] Currently, most bionic eyeballs are driven by motors, linkages, ropes, or pneumatic actuators, but their large size makes them unsuitable for the stringent space requirements of bionic eyeballs. For a detailed description of the drive structure of a bionic eyeball, please refer to the invention patent application CN122496625A. Therefore, how to miniaturize bionic eyeballs while enabling them to perform multi-degree-of-freedom movements with high precision is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention aims to solve the technical problem of how to miniaturize a bionic eyeball and enable it to perform multi-degree-of-freedom movements with high precision, and provides a bionic eyeball based on a co-located integrated DES magnetostrictive parallel joint and a control method thereon.

[0005] This invention provides a bionic eyeball based on a co-located integrated DES magnetostrictive parallel joint, comprising a DES magnetostrictive parallel joint and an eyeball body. The DES magnetostrictive parallel joint includes a lower substrate, an upper cover plate, a first DES magnetostrictive actuator, a second DES magnetostrictive actuator, and a third DES magnetostrictive actuator. The bottom of the second DES magnetostrictive actuator is connected to the lower substrate via a revolute joint, and the output end of the second DES magnetostrictive actuator is connected to the upper cover plate via a ball joint. The bottom of the first DES magnetostrictive actuator is connected to the lower substrate via a revolute joint, and the output end of the first DES magnetostrictive actuator is connected to the upper cover plate via a ball joint. The bottom of the third DES magnetostrictive actuator is connected to the lower substrate via a revolute joint, and the output end of the third DES magnetostrictive actuator is connected to the upper cover plate via a ball joint. The first, second, and third DES magnetostrictive actuators are evenly distributed in the radial circumferential direction. The eyeball body is connected to the upper cover plate.

[0006] The present invention also provides a control method for the above-mentioned bionic eyeball, comprising the following steps:

[0007] Establish a global static coordinate system {O0-X0Y0Z0}, a dynamic coordinate system {O7-X7Y7Z7}, and a local coordinate system {O i -X i Y i Z i}, i = 1-6;

[0008] Create the following formula:

[0009] (11)

[0010] In the formula, α, β, and γ are the rotation angle, pitch angle, and yaw angle of the upper cover plate, respectively;

[0011] Creating a mathematical model for the control system:

[0012]

[0013]

[0014]

[0015] In the formula, x1 represents the displacement of the actuator, x2 represents the speed of the actuator, and k s The system stiffness coefficient is represented by m, where m represents the mass of the magnet moving in the actuator, and c represents the mass of the magnet. s F represents the equivalent damping of the system. d F represents the pulling force of the sensor on the actuator. kl d represents the known load. r For external disturbances to the model, C0, C 1, C 2, C 3, C4 represents the coefficients of the fitted function, u represents the input voltage, and R represents the resistance of the coil;

[0016] Create the dynamic equation for the tracking error:

[0017]

[0018]

[0019] Euler discretization of the dynamic equations:

[0020]

[0021]

[0022] In the formula, T represents the sampling period;

[0023] Sliding membrane algorithm:

[0024] (20)

[0025] In the formula, ;

[0026] Using RBF neural network To fit the uncertainty F in the control law d The control law is further modified using delay interference compensation to obtain the control law. :

[0027] (twenty one)

[0028] Delay interference compensation for:

[0029] (twenty two)

[0030] Other Recorded as:

[0031] (twenty three)

[0032] Meanwhile, the weight iteration of the RBF neural network satisfies the following equation:

[0033] (twenty four)

[0034] Where γ represents the learning rate;

[0035] The system model is written as:

[0036] (27)

[0037] In the formula, It is a state transition function; It is a nonlinear measurement function; This is process noise; For measuring noise;

[0038] The prior state mean and prior error covariance are created as follows:

[0039] (31)

[0040] In the formula, It is a moment Prior state estimation; It is a moment The prior error covariance.

[0041] Preferably, the method further includes the following steps:

[0042] Predicting the Sigma point propagates through the measurement function:

[0043] (32)

[0044] in It is the first From the Sigma points of the measurement, the mean of the predicted measurement is further obtained as follows:

[0045] (33)

[0046] Simultaneously, the measurement prediction covariance and the state-measurement cross-covariance are obtained as follows:

[0047] (34)

[0048] Define the error vector and error covariance:

[0049] (35)

[0050] Perform a pre-calibration first:

[0051]

[0052]

[0053]

[0054] in Based on the old covariance The obtained pre-correction gain;

[0055] Define normalized squares:

[0056] (37)

[0057] Given confidence probability And define the chi-square threshold:

[0058] (38)

[0059] when At that time, it was considered that the current statistical model was basically consistent; when At this point, it is considered that at least one noise covariance matrix has become significantly mismatched, requiring an adaptive update to be triggered;

[0060] Next, the posterior update is performed, defining the posterior residual:

[0061] (39)

[0062] in, In pre-calibration state The measurement residuals below; using Regenerate posterior Sigma points And propagated to the measurement space to obtain Then define:

[0063] (40)

[0064] Therefore, the original estimate of the measurement noise covariance is taken as:

[0065] (41)

[0066] Define measurement noise update weights:

[0067]

[0068] in, It is the minimum update intensity of the measurement noise; It is the sensitivity coefficient; then:

[0069]

[0070] in, This represents the positive semidefinite projection operator;

[0071] For the process noise covariance, define the state change caused by this pre-correction:

[0072]

[0073] And further, we obtained:

[0074]

[0075] Define weights:

[0076]

[0077] in, It is the minimum update intensity of process noise; It is the sensitivity coefficient; then:

[0078]

[0079] Finally, the final estimate for the current moment is obtained:

[0080]

[0081] Finally, the pose estimate is extracted from the state vector:

[0082]

[0083] Reconstruction of pose error and branch compensation:

[0084]

[0085] in It is the compensation displacement of the outer ring on the three branches; It is a length reference; It is the final displacement reference sent to the inner ring.

[0086] The advantages of this invention are its small size, multi-degree-of-freedom eye movement, fast response speed, high movement precision, and strong flexibility.

[0087] Further features and aspects of the present invention will be clearly described in the following detailed description with reference to the accompanying drawings. Attached Figure Description

[0088] Figure 1 This is an axonometric view of a bionic eyeball based on a co-located integrated DES magnetostrictive parallel joint;

[0089] Figure 2 yes Figure 1 The front view of the bionic eyeball shown;

[0090] Figure 3 yes Figure 2 A cross-sectional view along the AA direction;

[0091] Figure 4 yes Figure 2 A magnified view of a section at point B in the middle;

[0092] Figure 5 yes Figure 2 A magnified view of a section at point C;

[0093] Figure 6 It is an axonometric drawing of a bionic eyeball from another perspective;

[0094] Figure 7 This is a schematic diagram of the structure of a DES magnetostrictive parallel joint;

[0095] Figure 8 This is a schematic diagram of the coordinate system establishment;

[0096] Figure 9 It is a block diagram of the control system principle;

[0097] Figure 10 These are experimental data on the performance of each drive unit in the UFK outer ring test;

[0098] Figure 11This is a 3D curve display of the UFK outer ring test;

[0099] Figure 12 This refers to the performance of each drive unit in the AUFK outer loop test;

[0100] Figure 13 This is a 3D curve display of the AUFK outer ring test.

[0101] Explanation of symbols in the diagram:

[0102] 1. Lower substrate, 2. First DES magnetostrictive actuator, 3. Second DES magnetostrictive actuator, 4. Third DES magnetostrictive actuator, 5. Eyeball, 6. Ball joint, 7. Rotary joint, 8. Upper cover plate. Detailed Implementation

[0103] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0104] Magnetoactors are compact, easily miniaturized, and have a fast response. For details on the specific structure of a magnetoacter, please refer to the invention patent application with publication number CN120038746A.

[0105] Dielectric elastomer sensors (DES) are characterized by flexible attachment, thinness, and customizable structure. Therefore, by wrapping the outer wall of the magnetostrictive actuator body with a dielectric elastomer sensor, the sensor is tightly attached to the actuator, providing continuous displacement feedback, thus forming a DES magnetostrictive actuator. For example... Figure 1-7 As shown, a first DES magnetostrictive actuator 2, a second DES magnetostrictive actuator 3, and a third DES magnetostrictive actuator 4 are provided. The bottom of the second DES magnetostrictive actuator 3 is connected to the lower substrate 1 via a rotating joint 7, and the output end of the second DES magnetostrictive actuator 3 is connected to the upper cover plate 8 via a ball joint 6. Similarly, the bottom of the first DES magnetostrictive actuator 2 is connected to the lower substrate 1 via a rotating joint, and the output end of the first DES magnetostrictive actuator 2 is connected to the upper cover plate 8 via a ball joint. The bottom of the third DES magnetostrictive actuator 4 is connected to the lower substrate 1 via a rotating joint, and the output end of the third DES magnetostrictive actuator 4 is connected to the upper cover plate 8 via a ball joint. The first DES magnetostrictive actuator 2, the second DES magnetostrictive actuator 3, and the third DES magnetostrictive actuator 4 are evenly distributed in the radial circumferential direction. The eyeball 5 is connected to the upper cover plate 8.

[0106] The lower substrate 1, upper cover plate 8, first DES magnetostrictive actuator 2, second DES magnetostrictive actuator 3, third DES magnetostrictive actuator 4, and corresponding ball joints and rotary joints constitute a DES magnetostrictive parallel joint. The DES magnetostrictive parallel joint combines miniaturization and multiple degrees of freedom. The three DES magnetostrictive actuators work together to enable the eyeball body 5 to move in multiple degrees of freedom.

[0107] Magnetoactors provide fast and controllable axial driving force. During the driving process, the dielectric elastomer sensor (DES) generates a near-linear capacitance change due to mechanical stretching, which can be calibrated and mapped into branch displacement feedback. At the same time, the tension-displacement characteristics of the DES will act in reverse on the actuator dynamics, thus forming a control object where driving, sensing and structure are coupled and coexist.

[0108] The magnetoactuator can be 23 mm long, 11 mm wide, and has a maximum stroke of 9 mm.

[0109] To achieve precise control, a dual-closed-loop integrated control algorithm based on adaptive state estimation and robust tracking control was created:

[0110] First, establish such Figure 8 The coordinate system shown is used to establish a global static coordinate system {O0-X0Y0Z0} at the geometric center of the end face of the lower substrate 1; the geometric centers of the rotating joint connected to the bottom of the DES magnetostrictive actuator are denoted as A1, A2, and A3, and the points where the upper cover plate 8 connects to the ball joint are denoted as B1, B2, and B3. Points A1, A2, and A3 form an equilateral triangle, and points B1, B2, and B3 also form an equilateral triangle; a local coordinate system {O0-X0Y0Z0} is established. i -X i Y i Z i}, i=1-6, specifically, establish coordinate systems {O1-X1Y1Z1}, {O3-X3Y3Z5}, and {O5-X5Y5Z5} at the geometric center A1, establish coordinate systems {O2-X2Y2Z2}, {O4-X4Y4Z4}, and {O6-X6Y6Z6} at the center of the length direction of the main body of the magnetostrictive actuator, and establish a moving coordinate system {O7-X7Y7Z7} at the geometric center of triangle B1B2B3.

[0111] Secondly, let O7B i and O0A i The distances (i = 1–3) are all r. Let the displacements of the upper cover plate 8 along the X0, Y0, and Z0 directions be x0, y0, and z0, respectively. The attitude transformation of the upper cover plate is described using XYZ Euler angles. Let the upper cover plate first rotate around the X7 axis by an angle α, then around the Y7 axis by an angle β, and finally around the Z7 axis by an angle γ, where α, β, and γ are defined as the rotation angle, pitch angle, and yaw angle, respectively. The rotation matrix of the upper cover plate's moving coordinate system {O7} relative to the global static coordinate system {O0} is given by the formula. As shown.

[0112] (1)

[0113] Define S as the geometric center of the ball joint connected to the upper cover plate 8. i (i = 1–3), assume ls l represents the axial distance between the ball joint and the center of the moving coordinate system of the upper cover plate. r The axial distance between the center of the rotating joint and the center of the lower base plate in the static coordinate system represents the distance between the three ball joint points S of the moving platform. i (i = 1–3) The coordinates of the upper cover plate in the moving coordinate system {O7-X7Y7Z7} and the three revolute joints R of the static platform i (i = 1–3) The coordinates of {O0-X0Y0Z0} in the static coordinate system of the lower substrate are as shown in the formula. As shown:

[0114] (2)

[0115] Let l0 represent the initial state of S i R i The length of the ball joint of the moving platform in the static coordinate system of the lower base plate is shown in the following equation. As shown. Among them. 0 P7 represents the position transformation vector of the upper cover plate's moving coordinate system {O7} relative to the lower substrate's static coordinate system {O0}.

[0116] (3)

[0117] From this, we can obtain the connection S. i R i The position vector L of (i = 1, 2, 3) in the static coordinate system of the lower substrate i It can be expressed as the formula As shown:

[0118] (4)

[0119] The position vector L1 of link S1R1 in the static coordinate system is as follows: As shown:

[0120] (5)

[0121] Given that the secondary center point S1 of the sphere can only move within the XZ plane, its structure is as shown in the equation. The constraints are shown.

[0122] (6)

[0123] The position vector L2 of link S2R2 in the static coordinate system is as follows: As shown.

[0124] (7)

[0125] The secondary center point S2 of the sphere can only be in the plane. Internal motion, its composition is as follows: The constraints are shown.

[0126] (8)

[0127] The position vector L3 of link S3R3 in the static coordinate system is as follows: As shown.

[0128] (9)

[0129] The secondary center point S3 of the sphere can only be in the plane. Internal motion, its composition is as follows: The constraints are shown.

[0130] (10)

[0131] For the three constraint equations ,Mode ,Mode Solving the equations simultaneously yields the following expression: .

[0132] (11)

[0133] Of the six motion parameters describing the upper cover plate, α, β, and z0 are independent, while x0, y0, and γ can be represented by independent poses. By performing inverse kinematics analysis on the 3-RPS structure, the mapping relationship between the moving coordinate system of the upper cover plate and the static coordinate system of the lower base plate can be obtained. From the equation... , , The length of the linkage in a certain position of the upper cover plate can be determined, and based on this, the actuator can be controlled to extend or retract to the corresponding length according to the position requirements. Linkage S i R i Length l i As shown in the formula As shown

[0134] (12)

[0135] J is defined as shown in equation (13), where J represents the velocity Jacobian matrix from the end pose of the center of the joint cover plate to each branch drive unit.

[0136] (13)

[0137] Then, a dual-closed-loop integrated control framework, including outer-loop state estimation compensation and inner-loop robust drive control, is established, such as... Figure 9As shown, three DES magnetostrictive units constitute the inner loop execution objects of the DES magnetostrictive parallel joint control system. After the ideal poses α, β, and z0 are input, the nominal reference displacements of each branch are first obtained through inverse kinematics calculation; subsequently, the DES feeds back the lengths l of the three branches at the current moment. s1 l s2 l s3 The gyroscope is mounted on the upper cover plate, and its feedback mechanism provides attitude angles α and β at the end point, which together serve as the outer loop observations. The outer loop constructs state equations based on kinematic constraints and performs online estimation and correction of z0 and the end-point pose within the AUKF framework. The estimation results are then mapped to three-branch displacement compensation quantities via a velocity Jacobian matrix, and superimposed with the nominal reference as the inner loop control input. This forms a dual closed-loop control path of "state estimation—error compensation—precise drive".

[0138] (14)

[0139] By filtering the two z0 solutions in each equation and discarding values ​​outside the interval, three z0 values ​​obtained from different actuator length values ​​are obtained. Further, in AUKF, the covariance of process noise and observation noise is adaptively adjusted based on the residuals, and the three z0 candidate values ​​and attitude angle measurements are fused to obtain an end-state estimate that better reflects actual operating conditions. Subsequently, by combining the error between the target pose and the estimated pose, and mapping the task space to the branch space using the velocity Jacobian matrix, the real-time compensation amounts for each of the three inner-loop actuators can be obtained.

[0140] Next, the inner loop driver is controlled; a fast terminal sliding mode algorithm (D-RBF-FTSM) combining neural networks and delay disturbance compensation is employed. The mathematical model of the control system is shown below:

[0141]

[0142]

[0143]

[0144] Where x1 represents the displacement of the actuator, x2 represents the speed of the actuator, and k s The system stiffness coefficient is represented by m, where m represents the mass of the magnet moving in the actuator, and c represents the mass of the magnet. s F represents the equivalent damping of the system. d F represents the pulling force of the sensor on the actuator. kl d represents the known load (which is 0). r We can consider external disturbances, including friction and unknown loads, as C0 and C2. 1, C 2, C 3,C4 represents the coefficients of the fitted function, u represents the input voltage, and R represents the resistance of the coil.

[0145]

[0146]

[0147] For simplicity, the definition is:

[0148] (17)

[0149] Substitute the mathematical model formula of the control system into the formula. , The dynamic equation for the tracking error can be obtained:

[0150] (17)

[0151] (18)

[0152] Euler discretization of formulas (17) and (18):

[0153]

[0154]

[0155] Where T represents the sampling period.

[0156] The discrete-time FTSM method is used, and the sliding surface is nonlinear, so it is selected as:

[0157] (20)

[0158] In the formula, .

[0159] Using RBF neural network To fit the uncertainty F in the control law d The control law is further modified using delay interference compensation to obtain the control law. :

[0160] (twenty one)

[0161] Delay interference compensation for:

[0162] (twenty two)

[0163] Other Recorded as:

[0164] (twenty three)

[0165] Meanwhile, the weight iteration of the RBF neural network satisfies the following equation:

[0166] (twenty four)

[0167] Where γ represents the learning rate.

[0168] Next, we will discuss outer loop state estimation and error compensation.

[0169] The outer loop uses three-way DES displacement feedback and gyroscope attitude angles as joint observations. First, based on the inverse kinematics constraints of the parallel mechanism, candidate values ​​for the end effector height are calculated from the lengths of the three branches. Then, the candidate height values ​​and attitude angle measurements are input into the AUKF for online fusion estimation of the end effector pose. Considering the significant time-varying characteristics of DES measurement noise, assembly deviations, and external disturbances under different operating conditions, innovative statistics are used to adaptively adjust the process noise covariance and observation noise covariance during the filtering process, thereby enhancing the adaptability of state estimation to model mismatch and measurement fluctuations. After obtaining the end effector pose estimation, the outer loop calculates the deviation between the target pose and the estimated pose, and maps the task space error to the displacement compensation of each branch through the Jacobian matrix. This compensation is then superimposed with the nominal inverse kinematics output and transmitted to the inner loop to achieve real-time correction of coupling errors, parameter drift, and system deviations.

[0170] Define the task space pose vector as

[0171] (25)

[0172] in, It is a moment The true end-effector pose; It is the desired pose; This represents two degrees of freedom in posture; This represents the height component at the end center.

[0173] Define the state vector of the outer loop AUKF. With measurement vector for

[0174] (26)

[0175] in, This corresponds to generalized velocity. For the first The length of each branch is measured by DES. ; The attitude angle of the gyroscope is measured.

[0176] The system model is written as:

[0177] (27)

[0178] In the formula, It is a state transition function; It is a nonlinear measurement function; This is process noise; For measuring noise; and Both allow for the unknown and time-varying.

[0179] Sigma point generation and time prediction: Defining parameters in unscented transformation :

[0180] (28)

[0181] in, It is the state dimension, here ; Controlling the scaling factor of the Sigma point relative to the mean; It is an additional scaling parameter.

[0182] Given the posterior estimate after the last sampling. and posterior error covariance Then the current sample number A Sigma point is defined as:

[0183] (29)

[0184] in, For the first Sigma points; This represents the square root of the covariance matrix obtained according to Cholesky's method. For its first The column. The weights corresponding to each point are:

[0185] (30)

[0186] in Used to compensate for higher-order statistics of prior distributions.

[0187] Substituting these Sigma points into the state equation, we obtain the prior state mean and prior error covariance as follows:

[0188] (31)

[0189] In the formula, It is a moment Prior state estimation; It is a moment The prior error covariance.

[0190] Next, the adaptive triggering design. The predicted Sigma point is propagated through the measurement function:

[0191] (32)

[0192] in It is the first There are Sigma points for each measurement. The predicted mean of the measurement is further obtained as follows:

[0193] (33)

[0194] Simultaneously, the measurement prediction covariance and the state-measurement cross-covariance are obtained as follows:

[0195] (34)

[0196] Define the error vector and error covariance:

[0197] (35)

[0198] In order to construct the posterior residual at the current moment, without immediately treating this update as the final output, a pre-correction is performed first:

[0199]

[0200]

[0201]

[0202] in Based on the old covariance The obtained "pre-correction gain"; Used only for subsequent assessment of the degree of mismatch and estimation of new [mismatches]. This is not the final output.

[0203] Define normalized squares:

[0204] (37)

[0205] Given confidence probability And define the chi-square threshold:

[0206] (38)

[0207] when At that time, it was considered that the current statistical model was basically consistent; when At this point, it is considered that at least one noise covariance matrix has become significantly mismatched, requiring an adaptive update to be triggered.

[0208] Next, we perform a posterior update. We define the posterior residual:

[0209] (39)

[0210] in, In pre-calibration state The measurement residuals below. To avoid absorbing all the state estimation errors incorrectly. In this case, a posterior dispersion term needs to be introduced. Using... Regenerate posterior Sigma points And propagated to the measurement space to obtain Then define:

[0211] (40)

[0212] Therefore, the original estimate of the measurement noise covariance is taken as:

[0213] (41)

[0214] Define measurement noise update weights

[0215]

[0216] in, It is the minimum update intensity of the measurement noise; It is the sensitivity coefficient; It is a small constant that avoids a denominator of zero. Therefore:

[0217]

[0218] in, This represents the positive semidefinite projection operator.

[0219] For the process noise covariance, define the state change caused by this pre-correction:

[0220]

[0221] And further, we obtained:

[0222]

[0223] here The PSD matrix is ​​naturally a matrix that can be viewed as "the energy that corrects the model prediction at the current time." If this correction energy is consistently large, it indicates that the process model does not adequately describe the current operating conditions, or... If the value is set too small, the process noise covariance should be increased.

[0224] Define weights:

[0225]

[0226] in, It is the minimum update intensity of process noise; It is the sensitivity coefficient. Therefore:

[0227]

[0228] Finally, the final estimate for the current moment is obtained:

[0229]

[0230] Finally, the pose estimate is extracted from the state vector:

[0231]

[0232] Reconstruction of pose error and branch compensation:

[0233]

[0234] in It is the compensation displacement of the outer ring on the three branches; It is a length reference obtained from nominal inverse kinematics; This is the final displacement reference sent to the inner ring.

[0235] The following comparative experiments and prototype verification will be conducted:

[0236] The experimental system employs three sets of programmable DC power supplies-bridge circuits as excitation and detection devices for each branch. A JY62 gyroscope is used at the joint end effector for tilt angle testing, and a third-party motion capture system is introduced to synchronously measure the actual pose. The experiment comprises two categories: first, a comparative experiment of the control effects of the UFK outer loop and the AUKF outer loop, used to evaluate the tracking accuracy, startup stability, and 3D synthesis error under different state estimation strategies; second, a biomimetic binocular prototype experiment, used to verify the feasibility of the proposed drive-sensor co-located integrated structure and dual-closed-loop control algorithm in binocular cooperative motion.

[0237] Figure 10 and Figure 11 The test results of a large-scale motion space under UFK outer loop control are shown. The target trajectory with three degrees of freedom is shown in the following formula:

[0238] (51)

[0239] Table 1: Three-degree-of-freedom test evaluation based on UFK outer ring

[0240]

[0241] Table 2. Three-degree-of-freedom test evaluation based on AUFK outer ring.

[0242]

[0243] Table 3 Three-dimensional indicators of three-degree-of-freedom test

[0244]

[0245] Figure 12 and 13 The test results under AUKF outer loop control are presented. The beta, alpha, and high response curves in the first row clearly show the improved trajectory tracking effect after the introduction of adaptive state estimation. Table 2 gives the error evaluation indexes for the three degrees of freedom under the AUKF outer loop, while Table 3 further compares the differences in three-dimensional comprehensive error between the UFK and AUKF outer loop strategies. The three-dimensional error is calculated using the normalized Euclidean distance between the target trajectory and the actual trajectory, and the maximum error, average error, root mean square error, and standard deviation are statistically obtained based on this distance. The normalization scale is taken as 0.15 times the range of the target trajectory. Table 3 shows that in the 75–130 s stage, the three-dimensional average error of the AUKF outer loop decreases from 1.528948 to 0.774568, a decrease of 49.34%; in the 130–180 s stage, the three-dimensional average error decreases from 2.744065 to 1.328190, a decrease of 51.60%. Meanwhile, the maximum error and root mean square error of the AUKF outer loop in both stages are lower than those of the UFK outer loop, indicating that the proposed adaptive state estimation method can more effectively suppress the adverse effects of noise fluctuations and model mismatch on tracking performance under large-scale motion and strong coupling conditions.

[0246] It is evident that the dual-loop integrated control exhibits a low average error in high-dynamic trajectory tracking, and its response during the startup phase is smoother and its resistance to coupling disturbances is stronger.

Claims

1. A bionic eyeball based on a co-located integrated DES magnetostrictive parallel joint, characterized in that, The device includes a DES magnetostrictive parallel joint and an eyeball. The DES magnetostrictive parallel joint includes a lower substrate, an upper cover plate, a first DES magnetostrictive actuator, a second DES magnetostrictive actuator, and a third DES magnetostrictive actuator. The bottom of the second DES magnetostrictive actuator is connected to the lower substrate via a revolute joint, and the output end of the second DES magnetostrictive actuator is connected to the upper cover plate via a ball joint. The bottom of the first DES magnetostrictive actuator is connected to the lower substrate via a revolute joint, and the output end of the first DES magnetostrictive actuator is connected to the upper cover plate via a ball joint. The bottom of the third DES magnetostrictive actuator is connected to the lower substrate via a revolute joint, and the output end of the third DES magnetostrictive actuator is connected to the upper cover plate via a ball joint. The first, second, and third DES magnetostrictive actuators are evenly distributed in the radial circumferential direction. The eyeball is connected to the upper cover plate.

2. A control method applied to the bionic eyeball of claim 1, characterized in that, Includes the following steps: Establish a global static coordinate system {O0-X0Y0Z0}, a dynamic coordinate system {O7-X7Y7Z7}, and a local coordinate system {O i -X i Y i Z i }, i = 1-6; Create the following formula: In the formula, α, β, and γ are the rotation angle, pitch angle, and yaw angle of the upper cover plate, respectively; Creating a mathematical model for the control system: In the formula, x1 represents the displacement of the actuator, x2 represents the speed of the actuator, and k s The system stiffness coefficient is represented by m, where m represents the mass of the magnet moving in the actuator, and c represents the mass of the magnet. s F represents the equivalent damping of the system. d F represents the pulling force of the sensor on the actuator. kl d represents the known load. r For external disturbances to the model, C0, C 1, C 2, C 3, C4 represents the coefficients of the fitted function, u represents the input voltage, and R represents the resistance of the coil; Create the dynamic equation for the tracking error: Euler discretization of the dynamic equations: In the formula, T represents the sampling period; Sliding membrane algorithm: In the formula, ; Using RBF neural network To fit the uncertainty F in the control law d The control law is further modified using delay interference compensation to obtain the following: Delay interference compensation for: Other Recorded as: Meanwhile, the weight iteration of the RBF neural network satisfies the following equation: Where γ represents the learning rate; The system model is written as: In the formula, It is a state transition function; It is a nonlinear measurement function; This is process noise; For measuring noise; The prior state mean and prior error covariance are created as follows: （31） In the formula, It is a moment Prior state estimation; It is a moment The prior error covariance.

3. The control method according to claim 2, characterized in that, It also includes the following steps: Predicting the Sigma point propagates through the measurement function: in, It is the first From the Sigma points of the measurement, the mean of the predicted measurement is further obtained as follows: Simultaneously, the measurement prediction covariance and the state-measurement cross-covariance are obtained as follows: Define the error vector and error covariance: Perform a pre-calibration first: in Based on the old covariance The obtained pre-correction gain; Define normalized squares: Given confidence probability And define the chi-square threshold: when At that time, it was considered that the current statistical model was basically consistent; when At this point, it is considered that at least one noise covariance matrix has become significantly mismatched, requiring an adaptive update to be triggered; Next, the posterior update is performed, defining the posterior residual: in, In pre-calibration state The measurement residuals below; using Regenerate posterior Sigma points And propagated to the measurement space to obtain Then define: Therefore, the original estimate of the measurement noise covariance is taken as: Define measurement noise update weights: in, It is the minimum update intensity of the measurement noise; It is the sensitivity coefficient; then: in, This represents the positive semidefinite projection operator; For the process noise covariance, define the state change caused by this pre-correction: And further, we obtained: Define weights: in, It is the minimum update intensity of process noise; It is the sensitivity coefficient; then: Finally, the final estimate for the current moment is obtained: Finally, the pose estimate is extracted from the state vector: Reconstruction of pose error and branch compensation: in It is the compensation displacement of the outer ring on the three branches; It is a length reference; It is the final displacement reference sent to the inner ring.

Images