CN113927596B - Width neural learning-based teleoperation limited time control method for time-varying output constraint robot - Google Patents

Abstract

The invention discloses a time-varying output constraint robot finite time control method of width neural learning. Based on integral barrier Lyapunov function, width neural learning algorithm and finite time theory, a time-varying output constrained finite-time controller based on width neural learning is innovatively proposed. The direct integral obstacle Lyapunov function ensures that the output of the system is within the time-varying boundary; the width neural learning algorithm combines the advantages of traditional neural network and width learning, and based on the inverse dynamics observer, the width learning algorithm solves the problem of external force perception. At the same time, the model uncertainty in the design process of the control rate u _xj is eliminated; the finite time theory ensures the robot's high-precision and rapid tracking of the reference signal. In summary, the algorithm ensures a stable, safe and efficient interaction between the robot and the environment, and improves the reliability and efficiency of the teleoperation system.

Description

A time-varying output-constrained robot teleoperation based on width neural learning time control method

technical field

The invention belongs to the technical field of robot control, and in particular relates to a time-varying output-constrained robot teleoperation finite-time control method based on width neural learning.

Background technique

Teleoperation technology makes full use of human intelligence and robot operation ability. It largely extends human perception and behavior capabilities in remote, unstructured and dangerous environments, and is an indispensable key technology in deep space, deep sea and deep ground exploration. Compared with intelligent robots, teleoperation technology fully considers the deficiencies of current intelligent technologies, such as decision-making issues involved in emergencies, safety and constraint issues in operations. This technology combines human perception and decision-making capabilities with robot operation capabilities, and overall enhances the teleoperation system's ability to deal with emergencies in an unstructured environment. It is currently the most realistic human-machine hybrid intelligence strategy.

The composition of the teleoperation system is complex, and the generation mechanism of operation instructions and the long-distance transmission of instructions inevitably cause uncertain delays to the system. Uncertain delay seriously affects the stability of the system and degrades the performance of the system; model uncertainty also poses a threat to the stability of the system. At the same time, limited by the operation time and work space, the robot end effector needs to complete the operation task within the expected time while satisfying the physical constraints. For example, in deep space exploration, due to the limitation of on-orbit operation time, when the robot performs inspection tasks (such as passing through a small space), it needs to use limited time to complete the on-orbit operation task under the condition of ensuring safety.

Contents of the invention

The technical problem solved by the present invention is: Based on the above problems, this patent proposes a time-varying output-limited robot teleoperation limited-time control method, which provides a feasible solution for the actual work of the teleoperation system.

The technical solution of the present invention is: a time-varying output-constrained robot teleoperation finite-time control method based on width neural learning, which is characterized in that it includes the following steps:

Step 1: Dynamically model the robotic arm;

Step 2: Realize the estimation of operator force and environment force by width neural learning algorithm and inverse dynamics observer;

Step 3: Design a finite-time control law with time-varying output constraints, eliminate the influence of model uncertainty, and realize high-precision tracking and rapid convergence of the teleoperated robot.

A further technical solution of the present invention is: the system is a pair of n-degree-of-freedom mechanical arms, and the model expression is:

Among them, j ∈ {m, s}, respectively identify the master and slave robots. q _j ∈R ^n×1 are the acceleration, velocity and position of the joint space respectively; /> x _j ∈ R ^n×1 are the acceleration, velocity and position of the operating space; M _xj is the inertia matrix, C _xj is the centrifugal force and Coriolis force matrix, g _xj is the gravity term matrix, u _xj is the control input, F _mν =F _h represents the force exerted by the operator, and F _sν =F _e represents the contact force between the slave robot and the environment.

The further technical solution of the present invention is: said step 2 includes the following sub-steps:

Step 2.1: Linearize the model as:

In the formula is the linear regression matrix, η is the parameter vector of the manipulator, and Θ∈R ^n×1 is the product of the two;

Step 2.2: External force evaluation using deep neural learning algorithm:

where q _j , is the input of the neural network;

Step 2.3: Based on the inverse dynamics observer, the dynamic model (15) of the manipulator is further linearized as

In the formula is the linear regression matrix, η is the parameter vector of the manipulator, and Θ∈R ^n×1 is the product of the two;

Step 2.4: Design a deep neural learning algorithm to estimate the external force:

The input of the neural network is q _j , Estimation of the interaction force is achieved by a wide neural learning algorithm.

The further technical scheme of the present invention is: in described step 3, comprise following sub-step:

Step 3.1: By designing the controller u _xj , realize the reference trajectory Limited time tracking, while ensuring that the output x _η1 is within the constraint area, ie />

The control law of the master-slave robot is

In formula (22): for /> The generalized inverse of , satisfying the following conditions

The update law in formula (22) is selected as follows:

In formula (22), the error variables e _j1 , e _j2 , j∈{m,s} are

In the formula: α _j1 is the virtual control quantity to be designed.

Equation (10) can generate the reference trajectory of the master end based on the operator's behavior,

Equation (27) where represent the estimated values of operator force and environmental force, respectively, where: /> are the acceleration, velocity and position of the host reference trajectory respectively; /> are the proportional coefficients of operator force estimation and environmental force estimation; M _r , C _r , g _r are target impedance parameters of operator behavior;

Step 3.2: The reference trajectory of the slave is: T _f (t) is the network transmission delay from the host to the slave; design the virtual control quantity α _j1 :

where 0<ξ _j <1, and γ _j1 are respectively:

In formula (22): are the operator force and the contact force between the robot and the environment, k _j1 , k _j2 , χ _j , b _j are positive constants, /> is a positive diagonal matrix. Here the weight vector of the neural network is W _j obtained by the width neural learning algorithm in step 2, />

Invention effect

The technical effects of the present invention are: the present invention focuses on solving problems such as system instability caused by time delay and uncertainty in the teleoperating system, difficulty in sensing external forces when the robot interacts with an unknown environment, and limited operating space and operating time of the robot. , a finite-time control method for robot teleoperation with time-varying output constraints based on width neural learning is proposed. Among them, the width neural learning algorithm effectively combines incremental learning and RBF neural network to realize the estimation of operator force and environmental force, and at the same time eliminate the negative impact caused by the uncertainty of the system model; the time-varying output constraint algorithm ensures The end position of the robot will not exceed the time-varying boundary, which improves the safety of the operation; the limited time control method ensures the fast tracking ability of the robot trajectory. This method does not require a force sensor, and the model of the system does not need to be accurately known. On the premise of ensuring the safe operation of the system, the high-precision tracking and rapid control of the teleoperated robot are realized.

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

(1) The present invention designs a width neural learning algorithm to realize the estimation of operator force and environment force; meanwhile, it eliminates the influence of model uncertainty in controller design. Compared with the traditional neural network, the network can guarantee a higher estimation accuracy, while saving a lot of computing pressure and not requiring sufficient learning conditions;

(2) A robust finite-time controller is designed to realize the finite-time control of the robot when the operating time is limited. The invention has a faster convergence speed and achieves the goal of efficient interaction between the robot and the environment.

(3) The present invention considers the safety problem of the operation process, and constrains the output of the robot, and the constraint boundary is a time-varying boundary, and the constant value boundary is a special form of the time-varying boundary, so the method has stronger practicability and practical significance.

Description of drawings

Fig. 1 teleoperation system control frame diagram;

Figure 2 Width Neural Learning Algorithm Framework;

Figure 3. Width neural learning algorithm based on inverse dynamics;

Figure 4 Simulation effect diagram (x, y, z axis as an example); (a) The picture shows the tracking effect on the x-axis and y-axis

(b) The picture shows the tracking effect on the z-axis and the operating space

Detailed ways

Referring to Fig. 1-Fig. 4, the concrete steps of this method and are provided as follows:

Step 1: Dynamic modeling of the system

Step 2: Design a width neural learning algorithm, and then realize the estimation of operator force and environmental force based on the inverse dynamics observer;

Step 3: Design a finite-time control law with time-varying output constraints, eliminate the influence of model uncertainty, and realize high-precision tracking and rapid convergence of the teleoperated robot (the control law of the master side is similar to that of the slave side, so the variable j∈{m,s }unified expression)

Combining the above steps, the stable, safe and efficient interaction between the teleoperation system and the environment can be realized.

The teleoperation system is complex, so it is necessary to summarize the general framework in order to develop the subsequent discussion more clearly.

Step 1: The system consists of a pair of n-degree-of-freedom manipulators. The dynamics in the operating space are as follows. For simplicity, the dynamics of the master and slave ends are uniformly written as:

In the formula: j ∈ {m, s}, respectively, the identification of master and slave robots. q _j ∈R ^n×1 are the acceleration, velocity and position of the joint space respectively. /> x _j ∈R ^n×1 are the acceleration, velocity and position of the operation space, respectively. M _xj is the inertia matrix, C _xj is the centrifugal force and Coriolis force matrix, g _xj is the gravity term matrix, u _xj represents the control input, F _mν = F _h represents the force applied by the operator, F _sν = F _e represents the slave robot contact with the environment.

Step 2: Based on the inverse dynamics observer, a width neural learning algorithm is designed to estimate the interaction force (the contact force between the master operator and the master robot, and the contact force between the slave robot and the environment). The dynamic model (15) of the manipulator can be linearized as

In the formula is the linear regression matrix, η is the parameter vector of the manipulator, and Θ∈R ^n×1 is the product of the two. Due to the inability to accurately model the manipulator and the model itself has certain deviations, the traditional inverse dynamics observer cannot accurately estimate the external force. Therefore, a deep neural learning algorithm is designed to estimate the external force:

The input of the neural network is q _j , Estimation of the interaction force is achieved by a wide neural learning algorithm. Due to the change of the external environment and the uncertainty of the system, the traditional algorithm needs to adjust the neural network parameters and retrain it, but the width neural learning algorithm designed by the present invention combines RBF and incremental learning, and by setting the threshold deviation, Adaptively increase nodes to get a better training network, see Figure 2. Based on the inverse dynamics observer, this width neural network algorithm can be used to estimate the interaction force; at the same time, the width neural learning algorithm can compensate the uncertainty of the system model. Pseudo-code of the width-based neural learning algorithm is shown in Table 1.

Table 1 Learning Algorithm Based on Width Neural

In the table: X is followed by q _j , The relevant input vectors, W and β are the weight vector and the radial basis vector of the RBF neural network respectively, and A ^m and A ^m+1 are defined pattern matrices.

Where: D=(A ^m ) ⁺ H _m+1 ,

In the formula, C=H _m+1 -A ^m D, so the weight can be obtained:

W _ei =(λI+(A ^m ) ^T A ^m ) ^-1 (A ^m ) ^T Y (20)

Step 3: In order to ensure the safety of the operation, it is necessary to limit the position output of the robot, that is, the end position of the robot is within the time-varying boundary. This patent adopts the direct barrier Lyapunov function (IBLF), width neural learning algorithm and finite time theory, and proposes a time-varying output constraint finite time control method based on width neural learning for the first time.

Control objective: by designing the controller u _xj , to realize the control of the reference trajectory Limited time tracking, while ensuring that the output x _η1 is within the constraint area, ie />

Design the control law of the master-slave robot as

In formula (22): for /> The generalized inverse of , satisfying the following conditions

The update law in formula (22) is selected as follows:

In formula (22), the error variables e _j1 , e _j2 , j∈{m,s} are

In the formula: α _j1 is the virtual control quantity to be designed.

Equation (26) can generate the reference trajectory of the master end based on the operator's behavior,

Equation (27) where Represent the estimated values of operator force and environmental force respectively (no force measurement device is required, and the width neural learning algorithm is used to estimate in step 1, see step 2), where: /> are the acceleration, velocity and position of the host reference trajectory respectively; /> are the proportional coefficients of operator force estimation and environmental force estimation; M _r , C _r , and g _r are the target impedance parameters of operator behavior, respectively.

The reference trajectory of the slave end is: T _f (t) is the network transmission delay from the master to the slave. Equation (22), the designed virtual control quantity α _j1

where 0<ξ _j <1, and γ _j1 are respectively:

In formula (22): are the operator force and the contact force between the robot and the environment, k _j1 , k _j2 , χ _j , b _j are positive constants, /> is a positive diagonal matrix. Here the weight vector of the neural network is W _j obtained by the width neural learning algorithm in step 2, />

For the teleoperation system (15), the virtual control quantity (28), control quantity (22) and update law (24) are selected to ensure the closed-loop stability of the teleoperation system and realize safe and efficient interaction between the robot and the environment.

The overall process framework of the system: the control framework of the remote control system based on time-varying output constraints is shown in Figure 1. The position signal x _m (t) of the master end is transmitted to the slave end through the communication link, and the reference signal of the slave end is obtained Then design a time-varying output-constrained finite-time controller u _xs based on width neural learning (see step 3), which can realize the slave robot to the reference signal /> High precision and fast tracking. At the same time, the environmental force of the slave is estimated by using the width neural learning algorithm (see step 2), and the virtual environment parameters are passed to the master. At the master end, the reconstruction of the environment force of the slave end is realized at the master end by using the motion information of the master end and the virtual environment parameters of the slave end. In order to enable the operator to have better force perception, the reference trajectory of the master is generated based on the operator's behavior. Then, a time-varying output constraint finite-time controller u _xm based on width neural learning is designed on the master end (see step 3), to realize the master-end robot’s control of the master-end reference signal /> High precision and fast tracking. (Since the master-side control law u _xm and the slave-side control law u _xs have the same form, they are expressed uniformly)

In summary, the present invention discloses a time-varying output-constrained robot finite-time control method based on width neural learning. Based on integral barrier Lyapunov function, width neural learning algorithm and finite time theory, a time-varying output constrained finite-time controller based on width neural learning is innovatively proposed. The direct integral obstacle Lyapunov function ensures that the output of the system is within the time-varying boundary; the width neural learning algorithm combines the advantages of traditional neural network and width learning, and based on the inverse dynamics observer, the width learning algorithm solves the problem of external force perception. At the same time, the model uncertainty in the design process of the control rate u _xj is eliminated; the finite time theory ensures the robot's high-precision and rapid tracking of the reference signal. In summary, the algorithm ensures a stable, safe and efficient interaction between the robot and the environment, and improves the reliability and efficiency of the teleoperation system.

Claims (1)

1. A teleoperation limited time control method of a time-varying output constraint robot based on width neural learning is characterized by comprising the following steps:

step 1: carrying out dynamic modeling on the mechanical arm; the mechanical arm is a pair of mechanical arms with n degrees of freedom, and the model expression is as follows:

wherein j is { m, s, is the master-slave robot identity respectively;acceleration, speed and position of joint space respectively; />Acceleration, speed and position of the operation space respectively; m is M _xj As an inertial mass matrix, C _xj For centrifugal and Golgi force matrices, g _xj As a gravity term matrix, u _xj Representing a control input;

step 2: estimating the force of an operator and the environmental force through a width neural learning algorithm and an inverse dynamics observer;

step 2.1: linearizing the model to:

in the method, in the process of the invention,is a linear regression matrix, eta is a parameter vector related to the mechanical arm, and theta epsilon R ^n×1 Is the product of the two;

step 2.2: external force assessment using a deep neural learning algorithm:

wherein the method comprises the steps ofIs an input to the neural network;

step 2.3: based on the inverse dynamics observer, the dynamics model (1) of the mechanical arm is further linearized into:

in the middle ofIs a linear regression matrix, eta is a parameter vector related to the mechanical arm, and theta epsilon R ^n×1 Is the product of the two;

step 2.4: the deep neural learning algorithm is designed to realize the estimation of external force:

the input of the neural network isThe interactive force can be estimated through a width neural learning algorithm;

step 3: and designing a time-varying output constraint finite time control law, resolving the uncertain influence of the model, and realizing high-precision tracking and rapid convergence of the teleoperation robot.

Step 3.1: by designing the controller u _xj Realize the reference trackLimited time tracking while guaranteeing output x _η1 Within the restricted area, i.e.)>

The control law of the master-slave robot is as follows:

in formula (6):the operator force and the contact force between the robot and the environment, respectively +.>Is->Is generalized inverse of (1), satisfies the following conditions

The update law in equation (6) is selected as follows:

error variable e in formula (6) _j1 ,e _j2 J epsilon { m, s } is

Wherein:alpha is the reference track of the master and slave ends _j1 The virtual control quantity to be designed;

a reference trajectory of the master may be generated based on operator behavior,

in (10) whereinRepresenting an estimate of the operator force and the ambient force, respectively, wherein: />Acceleration, speed and position of the main end reference track respectively; />Scaling factors for the operator force estimate and the ambient force estimate, respectively; mr, cr and gr are respectively target impedance parameters of the operator behavior;

step 3.2: the reference track of the slave end is:tf (t) is the network transmission delay from the master end to the slave end; designing virtual control quantity alpha _j1 ：

Wherein, the xi j is more than 0 and less than 1,and γj1 are respectively:

wherein: k (k) _j1 ,k _j2 ,χ _j ,b _j Is a normal number of times, and the number of times is equal to the normal number,is a right-angle array; the weight vector of the neural network is W obtained by the step 2 wide neural learning algorithm _j ，/>