A rail profiling flaw detection method based on a robotic arm

Abstract

The present invention discloses a rail profiling flaw detection method based on a manipulator, comprising the following steps: manual teaching, trajectory generation, regression curve generation, and autonomous flaw detection; the profiling flaw detection operation of the seamless rail joint is completed by the instructor and the manipulator in collaboration, the dynamic motion primitive DMP is used to learn the teaching trajectory of the manipulator, and the motion trajectory of the flaw detection probe driven by the manipulator is obtained, and then the regression curve of the stiffness data and the contact force data is obtained by using the Gaussian mixture model, and the regression curves are respectively the mapping of the end position of the manipulator to the end stiffness of the human arm and the end force of the manipulator, and the torque of the manipulator on the motion trajectory can be obtained, and the above data are used to realize the force / position hybrid closed-loop control of the manipulator through variable impedance control. The present invention can enable the manipulator to realize rail profiling flaw detection through trajectory learning and force / position hybrid control, and only simple teaching is required.

Description

Technical Field

[0001] The invention belongs to the technical field of rail detection, and in particular relates to a rail profiling flaw detection method based on a mechanical arm. Background Art

[0002] Tracks are the foundation of rolling stock operation, and their condition determines the safety and efficiency of rail transportation. As the years of operation increase, the tracks will be damaged by compression, elongation, bending, crushing, wear, and fracture, so it is necessary to conduct regular damage inspections on the tracks.

[0003] Currently, the commonly used non-destructive testing technologies include ultrasonic testing technology, radiographic testing technology, penetration testing technology, magnetic particle testing technology, eddy current testing technology, etc. Among them, ultrasonic and eddy current testing technology is the most widely used non-destructive testing technology. Since the ultrasonic wave of the flaw detection decays quickly in the air, it is necessary to ensure that the probe is close to the side and use a coupling agent. In order to ensure the smooth progress of the test, the probe needs to detect each part of the rail according to a specific trajectory, and it needs to be in contact with the rail at all times during the movement without causing a large impact.

[0004] In the prior art, the above operation relies on manual control by maintenance personnel. Firstly, the degree of automation is low and it completely relies on the operating experience of the maintenance personnel. Secondly, the flaw detection process cannot ensure that the probe is always in contact with the rail, and the contact force with the rail is difficult to control evenly. Therefore, the prior art not only has low precision, but also low efficiency. Summary of the invention

[0005] In response to one or more of the above defects or improvement needs in the prior art, the present invention provides a track profiling flaw detection method based on a robotic arm, which adopts DMP to learn the teaching trajectory of the robotic arm, uses a Gaussian mixture model to obtain the regression curve of the stiffness data and the contact force data, and uses the above-obtained data through variable impedance control to realize force / position hybrid closed-loop control of the robotic arm.

[0006] To achieve the above object, the present invention provides a rail profiling flaw detection method based on a robotic arm, comprising the following steps:

[0007] S1 manual teaching; the instructor and the robot arm work together to complete the contour flaw detection of the seamless rail joint. The end of the robot arm is equipped with a six-dimensional force / torque sensor. The instructor wears an electromyographic bracelet and drags the robot arm to complete the flaw detection operation. The robot arm records the end position, speed and acceleration when the operation is completed at a certain frequency f. The electromyographic bracelet collects the instructor's stiffness data at the same frequency f. The six-dimensional force / torque sensor also measures the contact force at the end joint at the frequency f. Repeat the teaching m times;

[0008] S2 trajectory generation; use dynamic motion primitives to learn the teaching trajectory using the terminal data set generated by m repeated teaching; determine the number of basis functions N and the system gain term according to the complexity of the teaching trajectory obtained in step S1, and calculate the target forcing function f target , the basis function weight can be obtained through the target forcing function, and then the forcing function f of the learning trajectory can be obtained. f is introduced into the robot end system to obtain the robot learning trajectory;

[0009] S3 regression curve generation: Map the stiffness data measured by the electromyographic bracelet to the actual end stiffness of the human arm, establish Gaussian mixture models of the end stiffness of the human arm and the end force of the robotic arm, perform Gaussian mixture regression, and obtain the regression curves of the end stiffness of the human arm and the end force of the robotic arm respectively;

[0010] S4 autonomous flaw detection; the robot arm drives the flaw detection device to detect the seamless rail joint according to the learning trajectory generated in step S2, and the variable impedance control strategy based on the position ensures that the contact force between the flaw detection device and the joint surface during the detection process is within the limited range of the model calculation curve in step S3.

[0011] As a further improvement of the present invention, in step S2, the terminal data set is a terminal data set with n data points collected on each trajectory during the m teaching process. where t=[t0,t1,…,t n-1 ] T is the time vector of the sampling point, y demo =[y0,y1,…,y n-1 ] T , They are the spatial coordinate vector, velocity vector and acceleration vector of the sampling point at the end of the robotic arm.

[0012] As a further improvement of the present invention, in step S2, the learning process of the dynamic motion primitive is as follows:

[0013] Determine the number of basis functions N and the system gain term, and obtain the target forcing function f through the terminal data set target :

[0014]

[0015] Among them, α y and β y is the system gain term, g is the trajectory target value, that is, y n-1 .

[0016] The nonlinear part is composed of a series of weighted superposition of Gaussian functions, and the trajectory is simulated by changing the weight value:

[0017]

[0018] Among them, ψ i =exp(-h i (xc i ) 2 ); x is used to control the phase of f, using the classical system, N is the number of basis functions. Usually, the more complex the trajectory, the more basis functions are needed; ω i is the corresponding basis function ψ i The weight of can be obtained by the following formula:

[0019]

[0020] In formula (3),

[0021] Substituting the obtained weight value into formula (2) to obtain the complete nonlinear part, the position, velocity and acceleration information of the learning trajectory can be calculated by the following formula:

[0022]

[0023] As a further improvement of the present invention, in step S3, the regression curve calculation process is as follows:

[0024] The data set ξ=[FK y d ] T , remember F =F T ,ξ K =K T , The Gaussian mixture model is used to encode the data, where a data point ξ∈R D×N The probability is:

[0025]

[0026] Where R is the real number field, π k ∈[0,1] is the prior probability, and K is the number of Gaussian distributions, D is the dimension of the data, μ k ∈R D ,Σ k ∈R D×D Represent the mean and covariance matrix of the k-th Gaussian distribution, respectively. Given the input variable ξ y , output ξ F and K The conditional probability distributions are:

[0027]

[0028] in and Respectively represent the mean and variance of the kth Gaussian distribution in the posterior probability:

[0029]

[0030] in and are the k-th Gaussian distribution means of F (or K) and y, respectively. and is the covariance matrix:

[0031]

[0032] ξ y The probability of being in the kth Gaussian distribution is:

[0033]

[0034] Gaussian mixture regression is given by ξ y Next F and K The expectation of the conditional probability distribution of :

[0035]

[0036] As a further improvement of the present invention, in step S4, the autonomous flaw detection process is as follows:

[0037] First, the stiffness expectation curve K d 、Expected end force curve F d , and the learning trajectory x d The error between the actual value of the end of the robot arm is input into the variable impedance controller and the displacement correction value x is output. c Then, combined with the original position of the flaw detector at the end of the robot arm, the target position to be adjusted by the robot arm is obtained. The target position is converted into the angle τ of each joint through the inverse kinematics of the robot arm and input into the position control loop. Then, the actual adjusted position q is output through the forward solution of the robot arm kinematics. At this time, a new contact force F will be generated between the flaw detector and the surface of the rail joint. e , thus achieving closed-loop control.

[0038] As a further improvement of the present invention, the robotic arm is fixedly installed on the detection vehicle, and a six-dimensional force / torque sensor and a detection probe are installed at the end of the robotic arm; the instructor wears the myoelectric bracelet on his arm, and the robotic arm's six-dimensional force / torque sensor, detection probe and myoelectric bracelet are all electrically connected to the host computer.

[0039] In general, the above technical solution conceived by the present invention has the following beneficial effects compared with the prior art:

[0040] (1) The rail profiling flaw detection method based on a robot arm of the present invention completes the profiling flaw detection operation of the seamless rail joint through the cooperation of the instructor and the robot arm, uses the dynamic motion primitive DMP to learn the robot arm teaching trajectory, obtains the motion trajectory of the flaw detection probe driven by the robot arm, and then uses the Gaussian mixture model to obtain the regression curve of the stiffness data and the contact force data. The regression curves are respectively the mapping of the end position of the robot arm to the end stiffness of the human arm and the end force of the robot arm, which can obtain the torque of the robot arm on the motion trajectory, and the above data are used to realize the force / position hybrid closed-loop control of the robot arm through variable impedance control. The present invention can enable the robot arm to realize rail profiling flaw detection through trajectory learning and force / position hybrid control, and only needs simple teaching.

[0041] (2) In the rail profiling flaw detection method based on a robotic arm of the present invention, in a variable impedance control strategy, the position deviation is obtained through an impedance controller through the deviation between the actual contact force and the expected force, and then the position of the flaw detection device at the end of the robotic arm is adjusted to ensure that the flaw detection device is always in contact with the seamless rail joint and the contact force is within a controllable range. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A technical roadmap of a rail profiling flaw detection method based on a robotic arm according to an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of hardware equipment involved in the rail profiling flaw detection method based on a robot arm according to an embodiment of the present invention;

[0044] Figure 3 A flow chart of a DMP learning a teaching trajectory involved in a rail profiling flaw detection method based on a robot arm according to an embodiment of the present invention;

[0045] Figure 4 The present invention discloses a variable impedance control process involved in a rail profiling flaw detection method based on a robotic arm according to an embodiment of the present invention.

[0046] In all the drawings, the same reference numerals represent the same technical features, specifically: 1-robotic arm, 2-six-dimensional force / torque sensor, 3-detection probe, 4-railway track, 5-electromyography bracelet, 6-host computer. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] Figure 1 The technical roadmap of the rail profiling flaw detection method based on a robotic arm according to an embodiment of the present invention is as follows: Figure 2 FIG. 1 is a diagram of hardware equipment involved in the rail profiling flaw detection method based on a robot arm according to an embodiment of the present invention. In conjunction with the accompanying drawings, the rail profiling flaw detection method based on a robot arm according to the present invention comprises the following steps:

[0049] S1 manual teaching; the instructor and the robot arm work together to complete the contour flaw detection of the seamless rail joint. The end of the robot arm is equipped with a six-dimensional force / torque sensor. The instructor wears an electromyographic bracelet and drags the robot arm to complete the flaw detection operation. The robot arm records the end position, speed and acceleration when the operation is completed at a certain frequency f. The electromyographic bracelet collects the instructor's stiffness data at the same frequency f. The six-dimensional force / torque sensor also measures the contact force at the end joint at the frequency f. Repeat the teaching m times;

[0050] Collect a terminal data set with n data points on each trajectory where t=[t0,t1,…,t n-1 ] T is the time vector of the sampling point, y demo =[y0,y1,…,y n-1 ] T , They are the spatial coordinate vector, velocity vector and acceleration vector of the sampling point at the end of the robotic arm.

[0051] like Figure 2 As shown, the robotic arm 1 is fixedly installed on the detection vehicle, and a six-dimensional force / torque sensor 2 and a detection probe 3 are installed at the end of the robotic arm 1; the instructor wears an electromyographic bracelet 5 on his arm, and the robotic arm 1, the six-dimensional force / torque sensor 2, the detection probe 3 and the electromyographic bracelet 5 are all electrically connected to the host computer 6.

[0052] S2 trajectory generation: The terminal data set generated by m repeated teachings is used to learn the teaching trajectory using dynamic motion primitives (DMP). The dynamic motion primitive learning process is as follows: Figure 3 As shown, the number of basis functions N and the system gain term are determined according to the complexity of the teaching trajectory obtained in step S1, and the target forcing function f is calculated. target, the basis function weight can be obtained through the target forcing function, and then the forcing function f of the learning trajectory can be obtained. The learning trajectory of the robot arm can be obtained by introducing f into the robot end system. The specific implementation process is as follows:

[0053] Determine the number of basis functions N and the system gain term, and obtain the target forcing function f through the terminal data set target :

[0054]

[0055] Among them, α y and β y is the system gain term, g is the trajectory target value, that is, y n-1 .

[0056] The nonlinear part is composed of a series of weighted superposition of Gaussian functions, and the trajectory is simulated by changing the weight value:

[0057]

[0058] Among them, ψ i =exp(-h i (xc i ) 2 ); x is used to control the phase of f, using the classical system, N is the number of basis functions. Usually, the more complex the trajectory, the more basis functions are needed; ω i is the corresponding basis function ψ i The weight of can be obtained by the following formula:

[0059]

[0060] In formula (3),

[0061] Substituting the obtained weight value into formula (2) to obtain the complete nonlinear part, the position, velocity and acceleration information of the learning trajectory can be calculated by the following formula:

[0062]

[0063] S3 regression curve generation: the stiffness data measured by the electromyographic bracelet is mapped to the actual end stiffness of the human arm, and Gaussian mixture models of the end stiffness of the human arm and the end force of the robotic arm are established respectively. Gaussian mixture regression is performed to obtain the regression curves of the end stiffness of the human arm and the end force of the robotic arm respectively.

[0064] The specific implementation process is as follows:

[0065] The data set ξ=[FK yd ] T , remember F =F T ,ξ K =K T , The Gaussian mixture model is used to encode the data, where a data point ξ∈R D×N The probability is:

[0066]

[0067] Where R is the real number field, π k ∈[0,1] is the prior probability, and K is the number of Gaussian distributions, D is the dimension of the data, μ k ∈RD,Σ k ∈R D×D Represent the mean and covariance matrix of the kth Gaussian distribution respectively. Given the input variable ξy, the output ξ F and K The conditional probability distributions are:

[0068]

[0069] in and Respectively represent the mean and variance of the kth Gaussian distribution in the posterior probability:

[0070]

[0071] in and are the k-th Gaussian distribution means of F (or K) and y, respectively. and is the covariance matrix:

[0072]

[0073] ξ y The probability of being in the kth Gaussian distribution is:

[0074]

[0075] Gaussian mixture regression is given by ξ y Next F and K The expectation of the conditional probability distribution of :

[0076]

[0077] S4 autonomous flaw detection; the robot arm drives the flaw detection device to detect the seamless rail joint according to the learning trajectory generated in step S2, and the variable impedance control strategy based on the position ensures that the contact force between the flaw detection device and the joint surface during the detection process is within the limited range of the model calculation curve in step S3.

[0078] The specific implementation process is as follows:

[0079] The variable impedance control process is as follows: Figure 4 As shown, firstly, the stiffness expectation curve K d 、Expected end force curve F d , and the learning trajectory x d The error between the actual value of the end of the robot arm is input into the variable impedance controller and the displacement correction value x is output. c Then, combined with the original position of the flaw detector at the end of the robot arm, the target position to be adjusted by the robot arm is obtained. The target position is converted into the angle τ of each joint through the inverse kinematics of the robot arm and input into the position control loop. Then, the actual adjusted position q is output through the forward solution of the robot arm kinematics. At this time, a new contact force F will be generated between the flaw detector and the surface of the rail joint. e The variable impedance control strategy of the present invention uses the deviation between the actual contact force and the expected force to obtain the position deviation through the impedance controller, and then adjusts the position of the flaw detector at the end of the robot arm to ensure that the flaw detector is always in contact with the seamless rail joint and the contact force is within a controllable range.

[0080] The rail profiling flaw detection method based on a robot arm of the present invention completes the profiling flaw detection operation of the seamless rail joint through the cooperation of the instructor and the robot arm, uses the dynamic motion primitive DMP to learn the robot arm teaching trajectory, obtains the motion trajectory of the flaw detection probe driven by the robot arm, and then uses the Gaussian mixture model to obtain the regression curve of the stiffness data and the contact force data. The regression curves are respectively the mapping of the end position of the robot arm to the end stiffness of the human arm and the end force of the robot arm, and the torque of the robot arm on the motion trajectory can be obtained, and the above data are used to realize the force / position hybrid closed-loop control of the robot arm through variable impedance control. Through trajectory learning and force / position hybrid control, the present invention can enable the robot arm to realize rail profiling flaw detection through simple teaching.

[0081] In order to better understand the rail profiling flaw detection method of the present invention, a specific embodiment is provided below:

[0082] The instructor wearing the myoelectric bracelet 5 drags the robot arm to perform a complete flaw detection operation on the rail head of the weld of the railway track 4, and repeats it 5 times. The data collected by the robot arm 1, the six-dimensional force / torque sensor 2, the detection probe 3 and the myoelectric bracelet 5 during the teaching process. The collected data include: the end position, velocity, acceleration, end force and torque of the robot arm, and the stiffness data of the human arm.

[0083] Each device starts recording data at the same time when teaching starts, and stops recording data at the same time after teaching is completed. After completing a complete data collection, the data is transmitted to the host computer 6. The teaching and data collection work of the robot arm detection work is completed.

[0084] The collected data set of the end-of-arm robot Import it into the DMP algorithm for trajectory learning to obtain the position, velocity and acceleration information of the learning trajectory.

[0085] First, the basis function number N and the system gain term α are determined according to the specific situation of the teaching trajectory. y , β y , Substitute the teaching data into formula (1) to obtain the target forcing function f target ; Then calculate the weight ω of each basis function through local weighted regression i , and calculate the learning trajectory forcing function f according to the weight; finally, the forcing function is introduced into the second-order system at the end of the manipulator to calculate the learning trajectory y d ,speed and acceleration The data set ξ=[FK y d ] T , Gaussian mixture regression is used to calculate the stiffness expectation curve K d and the desired end force curve F d .

[0086] The learning trajectory y d ,speed Acceleration Stiffness expected curve K d and the desired end force curve F d Input it into the controller, and record the actual contact force and stiffness of the end of the robot arm at the same time. Input the difference between it and the expected data into the variable impedance controller, and output the displacement correction value x c , and then combined with the original position of the flaw detection device at the end of the robot arm, the target position to be adjusted by the robot arm is obtained. The target position is converted into the angle τ of each joint through the inverse kinematics of the robot arm and input into the position control loop, thereby realizing the closed-loop control of the robot arm.

[0087] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A rail profiling flaw detection method based on a robotic arm, characterized in that: The following steps are involved: S1 manual teaching; the instructor and the robot arm work together to complete the contour flaw detection of the seamless rail joint. The end of the robot arm is equipped with a six-dimensional force / torque sensor. The instructor wears an electromyographic bracelet and drags the robot arm to complete the flaw detection operation. The robot arm records the end position, speed and acceleration when the operation is completed at a certain frequency f. The electromyographic bracelet collects the instructor's stiffness data at the same frequency f. The six-dimensional force / torque sensor also measures the contact force at the end joint at the frequency f. Repeat the teaching m times; S2 trajectory generation; Will m The terminal data set generated by repeated teaching uses dynamic motion primitives to learn the teaching trajectory; According to the complexity of the teaching trajectory obtained in step S1, the number of basis functions N and the system gain term are determined, and the target forcing function is calculated. , the basis function weight can be obtained through the target forcing function, and then the forcing function f of the learning trajectory can be obtained. f is introduced into the robot end system to obtain the robot learning trajectory; S3 regression curve generation: Map the stiffness data measured by the electromyographic bracelet to the actual end stiffness of the human arm, establish Gaussian mixture models of the end stiffness of the human arm and the end force of the robotic arm, perform Gaussian mixture regression, and obtain the regression curves of the end stiffness of the human arm and the end force of the robotic arm respectively; S4 autonomous flaw detection: the robot arm drives the flaw detection device to detect the seamless rail joint according to the learning trajectory generated in step S2, and the variable impedance control strategy based on the position ensures that the contact force between the flaw detection device and the joint surface during the detection process is within the limited range of the model calculation curve in step S3; In step S2, the terminal data set is m The terminal data set with n data points collected on each trajectory during the teaching process ,in is the time vector of the sampling point, , , are the spatial coordinate vector, velocity vector and acceleration vector of the sampling point at the end of the robot arm respectively; In step S2, the learning process of the dynamic motion primitive is as follows: Determine the number of basis functions N and the system gain term, and the target forcing function can be obtained through the terminal data set : (1) in, and is the system gain term, g is the trajectory target value, that is, ; The nonlinear part is composed of a series of weighted superposition of Gaussian functions, and the trajectory is simulated by changing the weight value: (2) in, ; x is used to control the phase of f, using the classical system, ; N is the number of basis functions. Usually, the more complex the trajectory, the more basis functions are needed; is the corresponding basis function The weight of can be obtained by the following formula: (3) In formula (3), , ; Substituting the obtained weight value into formula (2) to obtain the complete nonlinear part, the position, velocity and acceleration information of the learning trajectory can be calculated by the following formula: (4)。 2. The rail profiling flaw detection method based on a robotic arm according to claim 1, characterized in that: In step S3, the regression curve calculation process is as follows: The data set is obtained based on the stiffness data and end contact force data recorded during teaching in step S1 and the learning trajectory position information obtained in step S2 ,remember , , ; The Gaussian mixture model is used to encode the data, where a data point The probability is: (5) Where R is the field of real numbers, is the prior probability, and , K is the number of Gaussian distributions, D is the dimension of the data, Represent the mean and covariance matrix of the k-th Gaussian distribution, respectively. Given the input variable , output and The conditional probability distributions are: (6) in and Respectively represent the mean and variance of the kth Gaussian distribution in the posterior probability: (7) in and are the k-th Gaussian distribution means of F or K and y respectively, and is the covariance matrix: (8) The probability of being in the kth Gaussian distribution is: (9) Gaussian mixture regression is given by Down and The expectation of the conditional probability distribution of : (10)。 3. The rail profiling flaw detection method based on a robotic arm according to claim 1 or 2, characterized in that: In step S4, the autonomous flaw detection process is as follows: First, the stiffness expectation curve , terminal force expected curve , and learning trajectories The error between the actual value of the end of the robot arm is input into the variable impedance controller and then the displacement correction value is output. Then, combined with the original position of the flaw detection device at the end of the robot arm, the target position to be adjusted by the robot arm is obtained. The target position is converted into the angle of each joint through the inverse kinematics of the robot arm. , input into the position control loop, and then the actual adjusted position q is output through the kinematics of the robot arm. At this time, a new contact force will be generated between the flaw detection device and the surface of the rail joint. , thus achieving closed-loop control.

4. The rail profiling flaw detection method based on a robot arm according to claim 1 or 2, characterized in that: The mechanical arm is fixedly installed on the detection vehicle, and a six-dimensional force / torque sensor and a detection probe are installed at the end of the mechanical arm; the instructor wears the myoelectric bracelet on his arm, and the mechanical arm, torque sensor, detection probe and myoelectric bracelet are all electrically connected to the host computer.

Images