WO2020133881A1

WO2020133881A1 - Learning control method for mechanical apparatus, and mechanical apparatus learning control system having learning function

Info

Publication number: WO2020133881A1
Application number: PCT/CN2019/086712
Authority: WO
Inventors: 张冶; 李松洋; 王杰高
Original assignee: 南京埃斯顿机器人工程有限公司
Priority date: 2018-12-29
Filing date: 2019-05-13
Publication date: 2020-07-02
Also published as: CN109773787A

Abstract

A learning control method for a mechanical apparatus, and a mechanical apparatus learning control system having a learning function. A learning server, a mechanical control unit, a driving unit, a teaching system and a sensor are provided for a mechanical apparatus; a learning analysis unit is constructed in the mechanical control unit; during a learning process of the mechanical apparatus, segmented arrangement and analysis are carried out on process information and learning correction of an action; and process value data is integrated, and the arranged process value data is saved in the learning server. Process value information of a mechanical apparatus is recorded, a new processing process is analyzed and learning correction is acquired from a learning server according to an analysis result, so that there is no need to re-use a sensor to re-learn the processing process, thereby improving production efficiency.

Description

[Name of invention formulated by ISA according to Rule 37.2] 　Mechanical equipment control learning method and mechanical equipment control learning system with learning function

Technical field

The invention relates to the technical field of machine learning, and relates to machine equipment learning control, which is a machine equipment learning method with a learning function and a server system thereof.

Background technique

Mechanical equipment learning control is the control that optimizes the processing technology. The processing technology generally includes roughing, finishing, assembling, inspection, and packaging. The mechanical equipment learns the above processes to achieve optimal control of the processing technology. At present, scholars and engineers have applied learning control to robot control and other occasions. As described in the following literature:

The Chinese patent "Robot with Learning Control Function" (CN102189550A) discloses a robot with learning control function. A sensor is installed on a part of the position control object, and the learning control unit learns through the target trajectory or position detected and operated by the sensor. The learning correction amount is obtained, and the data collected by the external sensor includes position, speed, acceleration, inertia, etc. This method relies on the same predetermined action, and different actions need to be re-learned, so that the previous learning results cannot be reused.

The Chinese patent application "Robot Device with Learning Function" (CN106965171A) discloses a robot device that calculates a predetermined motion for improving the robot based on the physical quantity detected by the sensor during the predetermined motion of the robot during the predetermined motion The learning correction amount is calculated by the learning expansion unit through the relationship between the learning correction amount and the information of the learned predetermined action. For the information of the new action, the calculated relationship is used to calculate the learning correction amount for improving the new action. study again. In this technical solution, since the learning expansion unit is built into the control unit of the mechanical equipment, other similar mechanical equipment cannot use the learning result of the mechanical equipment, that is, there is a problem that the other mechanical equipment needs to re-learn to construct the learning expansion unit.

Summary of the invention

The problem to be solved by the present invention is that the existing mechanical equipment learning control method has a single learning object and is not universal. For new action instructions or new equipment, it is often necessary to re-learn, and the efficiency of the learning control method is not high.

The technical solution of the present invention is: a control learning method of mechanical equipment, a learning server, a mechanical control section, a driving section, a teaching system and a sensor are provided for the mechanical equipment,

The teaching system sends action commands to the machine control part, the drive part is used to drive the machine action, and the sensor is used to obtain the actual action information of the machine;

The mechanical control unit includes a motion analysis unit, a learning control unit, a storage unit, a learning analysis unit, and a motion control unit. The motion analysis unit performs motion analysis on the motion commands issued by the teaching system, and the motion analysis unit sends the analyzed motion information to the learning The control unit, storage unit, learning analysis unit, and motion control unit; the motion control unit sends control information to the drive unit, and the learning control unit learns the motion information from the motion analysis unit and the information collected by the sensor, which deviates from the two The amount of learning correction is obtained and stored in the storage unit, where the machine learns different predetermined actions, including action commands in the case of different positions, speeds, and accelerations; the stored action information and corresponding learning are cached in the storage unit The correction amount is transmitted to the learning and analysis department; the learning and analysis department sorts out the action information, that is, the learning correction amount, and uploads the sorted information to the learning server.

After the machine learns the specified action command, the learning analysis unit analyzes the new action command issued by the teaching system, and obtains the learning correction amount of the corresponding or similar action from the learning server according to the analysis result. In the new motion, the motion compensation is performed on the driving part. The motion control part performs robot motion control according to the motion information of the new motion command transmitted by the motion analysis part, and transmits the motion signal to the drive part. The drive part according to the motion information and the learning correction amount To make the machine execute according to the command action.

Further, the learning of the learning control unit and the analysis of the learning analysis unit are as follows:

Provided machinery for N groups of motion track according to a predetermined operation instruction, the learning control section obtained by learning the learning correction amount corresponding to N sets of tracks, the movement trajectory N _i is divided into M _i segment, the N sets of different trajectories that there are M sets operation information and the corresponding study Correction amount,

Record a piece of motion information as X _k , the corresponding learning correction amount is L _k , the motion information includes relative position, speed, acceleration and inertia, k=1, 2...M, correct the trajectory data of each section, calculate each trajectory Relative to the position of the starting point of the trajectory, the corrected trajectory position data is obtained, that is, the corrected action information Y _k is obtained, and W _k =[Y _k ,L _k ] is set to construct the action value data W=[W ₁ ,W ₂ ,...W _k ,...W _M ],

Solve the similarity D _pq of the motion information after pairwise correction with the time warping distance,

D _pq = CanonicalWarpingDistance[Y _p ,Y _q ]p,q∈M,p≠q

And in the same way, solve the similarity E _pq corresponding to the pairwise learning correction amount,

E _pq = CanonicalWarpingDistance[L _p ,L _q ]p,q∈M,p≠q

Then the similarity F _{pq of the} pairwise action value data,

F _pq = αD _pq + βE _pq

Among them, α, β are weight coefficients;

When F _pq <ε ₁ , the action value data W _p and W _q are considered to be similar, and the action value data is redundant; when the M sets of data are processed by similarity calculations, the redundant action value data is removed And organize and send to the learning server for use.

For new action instructions, after segmenting the trajectory of the new action and correcting the position data, the corrected action information of the segmented action information is Z _j , and the similarity solution analysis is performed with Y _k in the action value data W, Using the time warping distance, the similarity G _{jk of the} action information is obtained,

G _jk = CanonicalWarpingDistance[Z _j ,Y _k ]

When G _ij <ε ₂ , the learning correction amount of the new action corresponds to the learning correction amount L _k in the learning server, and the above process is repeatedly solved until the learning correction amount of the segmented trajectory of the new action instruction is all obtained.

As a preferred method, when calculating the similarity between the action information and the learning correction amount, the similarity calculation method used includes a time warping distance algorithm, a neural network algorithm, a fuzzy control algorithm, a genetic algorithm, and a simulated annealing algorithm.

Further, for the new action instruction, the obtained action information and learning correction amount are uploaded to the learning server.

The method of the present invention is used to optimize the action of mechanical equipment and process optimization of mechanical equipment.

The present invention also proposes a mechanical equipment server system with a learning function, including a learning server and a mechanical control unit. The mechanical equipment itself has an action instruction system, a drive unit, and a sensor. The action instruction system is used to issue an action instruction to the machine control unit, drive The part is used to drive the operation of the mechanical device, and the sensor is used to obtain the actual action information of the mechanical device; the learning server and the mechanical control part are storage media provided with a computer program that implements the method of claim 1 when the computer program runs.

The setting method of learning server includes mechanical equipment LAN server, enterprise server and cloud server.

The invention provides a mechanical equipment learning method with a learning function and a server system thereof. By recording the process value information of the mechanical equipment, the new processing technology is analyzed, and the learning correction amount is obtained from the learning server according to the analysis result, without the need to use The sensor re-learns the processing technology. Has the following beneficial effects:

1) The present invention proposes a control learning method for mechanical equipment. By learning predetermined instructions and learning the technological value information of mechanical actions, a learning server is constructed. When a new action instruction or new device is encountered, the new action instruction or For new machinery and equipment, the motion analysis unit analyzes the new motion instructions, and based on the results of the matching learning server, can obtain the learned learning correction amount without repeated learning.

2) The control learning method and server system of the mechanical equipment proposed in the present invention do not need to be bound to fixed mechanical equipment, and can be flexibly replaced and run on different mechanical equipment. The more process value information in the learning server, the more learned The more action instructions, the more devices the invention can be applied to.

3) The learning server system of the present invention can improve the working efficiency of mechanical equipment, reduce the cost of mechanical equipment and the cost of production line systems.

4) In the method of the present invention, the motion analysis unit analyzes the new motion command, and updates the learning server of the mechanical device with the new motion information and the data of the learning correction amount according to the result of the matching learning server, so that the learning server of the present invention continuously self Enrichment.

5) The method of the present invention can be used to assist in suppressing vibrations during the operation of mechanical equipment to achieve agile movements.

6) The method of the present invention can optimize the action of mechanical equipment from multiple angles, such as beat, power consumption, motion accuracy, etc., and improve production efficiency.

7) The method of the present invention can optimize the processing technology of mechanical equipment from multiple angles, casting, forging, etc., and improve the product processing quality of mechanical equipment.

8) The learning server system established by the present invention can be used by replacing equipment, and other mechanical equipment does not need to be re-learned. The motion analysis unit can directly obtain the corresponding learning correction amount from the learning server for the new motion.

BRIEF DESCRIPTION

Figure 1 is a schematic diagram of the method of the present invention.

2 is a schematic diagram of robot learning in the method of the present invention.

FIG. 3 is a flowchart of constructing a learning server in the present invention.

4 is a schematic diagram of the robot in operation in the method of the present invention.

FIG. 5 is a flowchart of using the learning server system in the present invention.

detailed description

The invention provides a mechanical equipment learning method with a learning function and a server system thereof. By recording the process value information of the mechanical equipment, the new processing technology is analyzed, and the learning correction amount is obtained from the learning server according to the analysis result, without the need to use The sensor re-learns the processing technology. In order to describe the convenience of mechanical equipment, the embodiments of the present invention take industrial robots and motion optimization control as examples, but are not limited to industrial robots, but also include other mechanical equipment such as servo drives, and are not limited to motion optimization control, but also include process optimization. control.

Hereinafter, a mechanical device learning server system according to an embodiment of the present invention will be described with reference to the drawings, in which the mechanical device takes an industrial robot as an example, that is, the following description regarding the robot is regarded as a mechanical device.

FIG. 1 shows a schematic diagram of a mechanical device learning method and a server system thereof according to an embodiment of the present invention. The robot is equipped with a learning server, robot control unit, drive unit, teaching system, and sensors. The teaching system sends motion commands to the robot control unit, hoping that the robot can perform the desired actions. The drive unit is used to drive the robot to move.

The robot control unit is composed of a motion analysis unit, a learning control unit, a storage unit, a learning analysis unit, and a motion control unit. The motion analysis unit mainly analyzes the motion commands transmitted from the teaching system. The motion analysis includes kinematics planning and dynamic analysis. The motion analysis unit sends the analyzed motion information to the learning control unit, storage unit, and learning analysis. Part, and motion control part. The learning control unit learns the motion information transmitted by the motion analysis unit and the information collected by the sensor, obtains the learning correction amount, and stores it in the storage unit. The robot learns with a predetermined motion in the work space. The predetermined motion needs to include different positions, speeds, accelerations, etc. to ensure the quality of the robot learning. The learning action information and the corresponding learning correction amount are stored in the storage unit, and when the learning is completed, it is transmitted to the learning analysis unit. When the learning is completed, the learning control unit and the storage unit may not be required. The learning analysis section sorts the action value information, removes redundant information, and uploads the action value information to the learning server. After the robot learning is completed, for the new action command issued by the teaching system, the learning analysis part analyzes the action, and obtains the learning correction amount of the corresponding or similar action from the learning server according to the analysis result, and acts on the drive part make up. The motion control unit performs robot motion control on the motion information transmitted from the motion analysis unit, and transmits the motion signal to the driving unit. The driving part is mainly composed of a servo drive and a servo motor. The servo drive drives the servo motor through position, speed or current feedback control to make the robot mechanism perform the expected action.

The sensor is installed on the target part of the robot's position control, and is used to obtain physical information such as the position, velocity, or acceleration of the target part.

The sensors involved in the above embodiments include encoders, position sensors, speed sensors, acceleration sensors, vision sensors, force sensors, angular velocity sensors, gyro sensors, inertial measurement units, and the like.

The following describes the construction and use of the learning server system.

Figure 3 shows a flow chart for building a learning server. First, the teaching system issues motion commands to the robot control unit. The motion analysis unit of the robot control unit analyzes the motion commands. The learning control unit obtains the motion information and the data collected by the external sensor to learn to obtain the learning correction of the overall motion. And save the motion information and learning correction amount to the storage unit. Next, the learning analysis unit takes out the action information and the learning correction amount from the storage unit, and processes the learning correction amount and the action information in segments to construct action value data, and removes the redundant result after analysis. Finally, the learning analysis department uploads the final analysis results to the learning server.

Remember the different trajectories of the position, speed, acceleration, inertia, etc. of N groups of robots in the working space. An acceleration sensor is used as an example of the sensor installed in the target part, and the physical quantity detected by the sensor is the acceleration of the target part controlled by the robot position. The velocity component is obtained after one integration, and the position component is obtained after two integrations.

M _i can be determined according to the length of the trajectory and the number of interpolation points. The first group of trajectories is divided into M ₁ segments, the second group of trajectories is divided into M ₂ segments, and so on.

D _pq = CanonicalWarpingDistance[Y _p ,Y _q ] p,q∈M,p≠q

E _pq = CanonicalWarpingDistance[L _p ,L _q ] p,q∈M,p≠q

Then the similarity F _{pq of the} pairwise action value data,

F _pq = αD _pq + βE _pq

Among them, α, β are weight coefficients;

When F _pq <ε ₁ , the action value data W _p and W _q are considered to be similar, and the action value data is redundant. ε ₁ is the similarity threshold, corresponding to the actual operating conditions of different models, different loads, different speeds, different inertias, etc., which can be adjusted in advance according to the experiment.

After the M sets of data are processed by similarity calculations, the redundant action value data is removed and sorted, and sent to the learning server for use by the robot and other robots.

Fig. 5 shows a flow chart of the use of the learning server system. First, the teaching system issues motion commands to the robot control unit, and the motion analysis unit in the control unit analyzes the motion commands. The learning analysis unit performs further segmentation processing on the analyzed action information, and obtains the corresponding learning correction amount from the learning server after analysis. Finally, the learning correction amount is added to the motion information output by the motion control unit and transmitted to the drive unit.

G _jk = CanonicalWarpingDistance[Z _j ,Y _k ]

When G _ij <ε ₂ , the learning correction amount of the new action corresponds to the learning correction amount L _k in the learning server, and the above process is repeatedly solved until the learning correction amount of the segmented trajectory of the new action instruction is all obtained. The threshold ε ₂ here is also obtained based on the experimental adjustment of different models, different loads, different speeds, different inertias and other working conditions, and is preset.

As explained above, according to the method and system of the present invention, on the learned robot and the robot without any learning, the new motion instruction does not need to spend time and effort to install and re-install sensors, re-learn, and the learning analysis part analyzes the motion information, Get the learning correction amount directly from the learning server.

The present invention uses the time warping distance to construct a learning server system and use it, but it is not limited to this method, and other methods of calculating similarity may also be used. , Such as error sum of squares, least squares, correlation coefficient, neural network algorithm, fuzzy control algorithm, genetic algorithm and simulated annealing algorithm.

The mechanical equipment learning control of the present invention is optimized control for the processing technology, including the motion control and technological process control of the mechanical equipment. Construct a learning and analysis section in the control section of machinery and equipment. During the learning process of machinery and equipment, process and analyze the process information and learning correction amount stored in the storage section, and use time warping distance to integrate process value data and organize The post-process value data is saved to the learning server. The process value data stored in the learning server can be used not only for the learned mechanical equipment, but also for other unlearned mechanical equipment. The learning server can be updated according to the requirements to ensure the high efficiency of the mechanical equipment performance.

Claims

A control learning method for mechanical equipment, which is characterized by setting a learning server, a mechanical control section, a driving section, a teaching system and sensors for the mechanical equipment,

The teaching system sends action commands to the machine control part, the drive part is used to drive the machine action, and the sensor is used to obtain the actual action information of the machine;

The mechanical control unit includes a motion analysis unit, a learning control unit, a storage unit, a learning analysis unit, and a motion control unit. The motion analysis unit performs motion analysis on the motion commands issued by the teaching system, and the motion analysis unit sends the analyzed motion information to the learning The control unit, storage unit, learning analysis unit, and motion control unit; the motion control unit sends control information to the drive unit, and the learning control unit learns the motion information from the motion analysis unit and the information collected by the sensor, which deviates from the two The amount of learning correction is obtained and stored in the storage unit, where the machine learns different predetermined actions, including action commands in the case of different positions, speeds, and accelerations; the stored action information and corresponding learning are cached in the storage unit The correction amount is transmitted to the learning and analysis department; the learning and analysis department sorts out the action information, that is, the learning correction amount, and uploads the sorted information to the learning server.

After the machine learns the specified action command, the learning analysis unit analyzes the new action command issued by the teaching system, and obtains the learning correction amount of the corresponding or similar action from the learning server according to the analysis result. In the new motion, the motion compensation is performed on the driving part. The motion control part performs robot motion control according to the motion information of the new motion command transmitted by the motion analysis part, and transmits the motion signal to the drive part. The drive part according to the motion information and the learning correction amount To make the machine execute according to the command action.
The control learning method for mechanical equipment according to claim 1, wherein the learning of the learning control unit and the analysis of the learning analysis unit are as follows:

Provided machinery for N groups of motion track according to a predetermined operation instruction, the learning control section obtained by learning the learning correction amount corresponding to N sets of tracks, the movement trajectory N i is divided into M i segment, the N sets of different trajectories that there are M sets operation information and the corresponding study Correction amount,

Record a piece of motion information as X k , the corresponding learning correction amount is L k , the motion information includes relative position, speed, acceleration and inertia, k=1, 2...M, correct the trajectory data of each section, calculate each trajectory Relative to the position of the starting point of the trajectory, the corrected trajectory position data is obtained, that is, the corrected action information Y k is obtained, and W k =[Y k ,L k ] is set to construct the action value data W=[W 1 ,W 2 ,...W k ,...W M ],

Solve the similarity D pq of the motion information after pairwise correction with the time warping distance,

D pq = CanonicalWarpingDistance[Y p ,Y q ] p,q∈M,p≠q

And in the same way, solve the similarity E pq corresponding to the pairwise learning correction amount,

E pq = CanonicalWarpingDistance[L p ,L q ] p,q∈M,p≠q

Then the similarity F pq of the pairwise action value data,

F pq = αD pq + βE pq

Among them, α, β are weight coefficients;

When F pq <ε 1 , ε 1 is the set similarity threshold, the information of action value data W p and W q is considered to be similar, and action value data is redundant;

After the M sets of data are processed by similarity calculations, the redundant action value data is removed and sorted, and sent to the learning server for use.
A control learning method for mechanical equipment according to claim 2, characterized in that for a new motion command, the trajectory of the new motion is segmented, and after the position data is corrected, the motion information after the segmented motion information is corrected For Z j , perform a similarity solution analysis with Y k in the action value data W, and use the time warping distance to obtain the similarity of action information G jk ,

G jk = CanonicalWarpingDistance[Z j ,Y k ]

When G ij <ε 2 , ε 2 is the set threshold, the learning correction amount of the new action corresponds to the learning correction amount L k in the learning server, and the above process is repeatedly solved until the learning correction amount of the segmented trajectory of the new action instruction is all get.
A control learning method for mechanical equipment according to claim 2 or 3, characterized in that when calculating the similarity between the action information and the learning correction amount, the similarity calculation method used further includes sum of squared errors, least squares, correlation Coefficient, neural network algorithm, fuzzy control algorithm, genetic algorithm and simulated annealing algorithm.
The control learning method of a mechanical device according to claim 1, wherein for the new motion instruction, the obtained motion information and the learning correction amount are uploaded to the learning server.
A control learning method for mechanical equipment according to claim 1, characterized in that it is used for optimal control of the movement of mechanical equipment and process optimization of the mechanical equipment.
A mechanical equipment control learning system with a learning function is characterized by including a learning server and a mechanical control part. The mechanical equipment itself has an action command system, a driving part and a sensor. The action command system is used to issue a motion command to the mechanical control part and drive The part is used to drive the operation of the mechanical device, and the sensor is used to obtain the actual action information of the mechanical device; the learning server and the mechanical control part are storage media provided with a computer program that implements the method of claim 1 when the computer program runs.
The machine equipment control learning system according to claim 6, characterized in that the learning server is arranged in a manner including a machine equipment LAN server, an enterprise server, and a cloud server.