A kind of robot neural network formula computed moment control device training platform and training method
Technical field:
The invention belongs to the Robot Control Technology field, relate to a kind of robot neural network formula computed moment control device training platform and training method, be specifically related to a kind of high speed parallel robot neural network formula computed moment control device training platform and training method based on virtual machine real electric.
Background technology:
The design feature of parallel robot self makes it need not obtain widespread use in the field of very big work space in the high rigidity of needs, high precision or big load; In addition, its drive unit places on the fixed platform mostly or near the position of fixed platform, motion parts is in light weight like this, and dynamic response is good, can realize high-speed motion.
Aspect control, at present, most of Business Machine hands all adopt linear single shaft control modes such as PD and PID control, although this control mode can reach satisfied precision under low speed, but high-speed parallel manipulator can't realize control quality satisfied under the high-speed motion by this controller because of the influence of centrifugal force, Coriolis and gravity item.
In fact, computed moment control method clear thinking, be widely used in the robot controller design, but because robot is a very complicated multiple-input and multiple-output nonlinear system, change when having, strong coupling and nonlinear dynamic characteristic, traditional factored moment model can't accurately approach actual robot control model, therefore can not satisfy the occasion that degree of precision requires.And neural network has the ability of the non-linear and uncertain object of any complexity of study, utilizes the uncertainty factor of neural network learning high speed parallel robot system, and the factored moment model is compensated and corrected, and can realize its high precision control.
When adopting nerve network controller, its training network model is also proved the emulation mode that mostly is of its algorithm validity, and mainly be by obtaining based on " empty machine " or " void " system, as Ideas, MATLAB/SimuLink, ADAMS, tool software such as SolidWorks and EASY-ROB, and limitation " empty machine " or " void " system.This neural network model that obtains based on emulation and the gap of actual machine human model make its controller be difficult to well be used in engineering practice.On the other hand, if adopt the words of actual model machine neural network training, in view of its initial model and robot model's difference and the complicated uncertainty of training process, because the mistake equipment that makes of controller model meets accident in training process, this cost will be difficult to burden probably.
Summary of the invention:
The present invention is intended to improve the Trajectory Tracking Control performance of high speed parallel robot, at the deficiencies in the prior art, proposition is based on high speed parallel robot neural network formula computed moment control device training platform and the training method of virtual machine real electric, by building the virtual machine real electric training platform quiet with the actual machine robot mechanism, that dynamic property is consistent, and by this platform training neural network, obtain the neural network formula computed moment control model of the model of realistic robot control largely, for the design of controller is offered help.
For solving the problems of the technologies described above, basic design of the present invention is: build at utmost consistent with mock-up high speed parallel robot virtual machine real electric system, drive empty machine by real electricity, and from real electricity and empty machine, gather prearranged signals as training sample, neural network training formula computed moment control device, obtain the computed moment control model that the neural network of empty machine is represented, respectively connect weights in the neural network after namely training is finished and threshold value has characterized the computed moment control model.
A kind of robot provided by the present invention neural network formula computed moment control device training platform is reached by upper control module, open motion control driver element, high speed parallel robot virtual reality unit, data acquisition and communication unit to be formed based on the adaptive model based control of artificial neural network; Described upper control module is finished motion planning and process monitoring; Described open motion control driver element comprises motion controller, driver, servomotor, is to finish the electric part of reality that upper control module steering order transforms to automatically controlled driving; Described high speed parallel robot virtual reality unit is used for realizing the structure of the empty machine of high speed parallel robot and relevant three-dimensional scenic; Described data acquisition and communication unit are realized mutual data communication between described upper control module, open motion control driver element and high speed parallel robot virtual reality unit three, and can gather the required training sample of artificial neural network; Described adaptive model based control based on artificial neural network approaches device by parallel neural network formula robot inverse dynamics model and the linear scaling differential feedback controller is formed.
Described adaptive model based control based on artificial neural network is the operand of this training platform, in training period, by the empty machine actual path of study high speed parallel robot and corresponding control moment sample, and with desired trajectory relatively, press weights adjustment rule and adjust its artificial neural network connection weights, progressively approach the empty machine inverse dynamics model of high speed parallel robot, in the runtime, it is embedded into upper control module as the core of controller, its kernel is equal to the inverse dynamics model that the high speed parallel robot has compensated the nonlinear uncertain item, realizes by the mapping of high speed parallel robot desired trajectory to control moment.
The training method concrete steps of a kind of robot provided by the present invention neural network formula computed moment control device are as follows:
(1) is ingredient with described upper control module, open motion control driver element, high speed parallel robot virtual reality unit and data acquisition and communication unit, builds high speed parallel robot virtual machine real electric control system;
(2) foundation is based on the adaptive model based control of artificial neural network, this model approaches device by parallel neural network formula robot inverse dynamics model and the linear scaling differential feedback controller is formed, neural network formula robot inverse dynamics model approaches the output torque of device and the output torque sum of linear scaling differential feedback controller is imported as the signal of controller, drives empty each joint motions of machine of high speed parallel robot;
(3) the described adaptive model based control based on artificial neural network of step (2) is embedded the described high speed parallel robot of step (1) virtual machine real electric control system, obtain high speed parallel robot virtual machine real electric training platform;
(4) from the described high speed parallel robot of step (3) virtual machine real electric training platform, obtain some groups of data of the actual motion track (position actual value, speed actual value) of the movement locus (position expectation value, speed expectation value, acceleration expectation value) of expectation and the empty machine of corresponding high speed parallel robot, as training sample;
(5) the described adaptive model based control based on artificial neural network of the described training sample input step of step (4) (2) is trained it, and adjust with reference to index as it with the output torque of the described linear scaling differential feedback controller of step (2), constantly adjust connection weights and the threshold value of artificial neural network, the control action for the treatment of described linear scaling differential feedback controller is during to insignificant degree, and training process finishes;
(6) cast out step (5) training based on the linear scaling differential feedback controller in the adaptive model based control of artificial neural network, keep described neural network formula robot inverse dynamics model and approach device, and solidify connection weights and the threshold value that described neural network formula robot inverse dynamics model approaches artificial neural network in the device, namely get neural network formula computed moment control device;
(7) with the adaptive model based control based on artificial neural network described in step (6) the described neural network formula computed moment control device replacement step (3), obtain the high speed parallel robot control system of Torque Control.
For described training method, it should be noted that the purpose of using described linear scaling differential feedback controller is to have track following performance preferably in order to guarantee in the initial stage robot of artificial neural network training; About the training to described adaptive model based control based on artificial neural network, basic process is, the training initial stage, a little less than the effect of neural network, joint of robot outgoing position and velocity amplitude are failed the expectation value of pursuit movement planner planning, big tracing deviation occurs, by the regulating action of described linear scaling differential feedback controller, and along with going deep into gradually of training, this deviate reduces gradually; Meanwhile, getting the output of described linear scaling differential feedback controller measures as the reference that neural network formula robot inverse dynamics model approaches device, set up the error function of this neural network, and revise network along the direction that error function reduces network connection weights gradient by the learning rate of appointment and respectively connect weights.
Why make the choice of step (6), be because, training latter stage, robot trajectory's tracking performance significantly improves, and it is very faint that the effect of described linear scaling differential feedback controller becomes, in the extent of the error that allows, can ignore its effect, learn accordingly, the described neural network formula of step (6) robot inverse dynamics model approaches the inverse dynamics characteristic that device has reacted the empty machine of high speed parallel robot, like this, obtained the model of neural network formula computed moment control device.
The present invention has following characteristics:
1, the experiment method of virtual machine real electric, the high speed parallel robot adopts virtual prototype, can be at dissimilar different freedom robot mechanisms, only need rebulid dummy model, economy and the time avoiding rebuliding solid model drop into, and fitness is wide, and system's design cost is low.
2, this training method is compared with traditional training method in kind, owing to directly do not drive material object, the mistake that can avoid model to set up causes the accident of physical mechanism in training process, has guaranteed the security of training.
3, this design system formation is simple relatively, drives empty machine by real electricity, as driving true mechanism, can guarantee the validity of training, but and the fast construction training platform, reasonably neural network model makes up and can shorten the training time, from controller exploitation angle, guaranteed high efficiency.
4, through neural metwork training, the computed moment control device has compensated uncertain of model error, external interference etc., improved robustness and the overall progressive stability of controller, can realize the Trajectory Tracking Control of high speed parallel robot preferably, accelerated its engineering and use.
Description of drawings:
Fig. 1 is training platform structural representation of the present invention;
Fig. 2 is inverse dynamics model neural metwork training scheme synoptic diagram among the present invention;
Fig. 3 is to be the neural network training structural representation of inverse dynamics model among the present invention;
Fig. 4 is that the empty machine of high speed parallel robot of the present invention drives the interface synoptic diagram;
Fig. 5 is the high speed parallel robot virtual machine real electric control system structural representation that carries neural network formula computed moment control device.
Embodiment:
Below in conjunction with accompanying drawing the present invention is further described:
As shown in Figure 1, a kind of robot neural network formula computed moment control device training platform is by upper control module, open motion control driver element, high speed parallel robot virtual reality unit, data acquisition and communication unit and form based on the adaptive model based control of artificial neural network; Described upper control module is finished motion planning and process monitoring; Described open motion control driver element comprises motion controller, driver, servomotor, is to finish the electric part of reality that upper control module steering order transforms to automatically controlled driving; The structure of the empty machine of high speed parallel robot and relevant three-dimensional scenic can be realized in described high speed parallel robot virtual reality unit; Described data acquisition and communication unit are realized mutual data communication between described upper control module, open motion control driver element and high speed parallel robot virtual reality unit three, and can gather the required training sample of artificial neural network.As shown in Figure 2, described adaptive model based control based on artificial neural network approaches device by parallel neural network formula robot inverse dynamics model and the linear scaling differential feedback controller is formed.The purpose of usage ratio derivative controller is to have track following performance preferably in order to guarantee in the initial stage robot of artificial neural network training.
Described adaptive model based control based on artificial neural network is the operand of this training platform, the present invention obtains high speed parallel robot inverse dynamics model by neural network training, for controller design lays the foundation, claim that the controller that obtains thus is neural network formula computed moment control device.
The training method concrete steps of a kind of robot provided by the present invention neural network formula computed moment control device are as follows:
(1) is ingredient with described upper control module, open motion control driver element, high speed parallel robot virtual reality unit and data acquisition and communication unit, builds high speed parallel robot virtual machine real electric control system;
(2) foundation is based on the adaptive model based control of artificial neural network, this model approaches device by parallel neural network formula robot inverse dynamics model and the linear scaling differential feedback controller is formed, and neural network formula robot inverse dynamics model approaches the output torque τ of device
NOutput torque τ with the linear scaling differential feedback controller
fSum τ drives empty each joint motions of machine of high speed parallel robot as the signal input (with reference to Fig. 2) of controller;
(3) the described adaptive model based control based on artificial neural network of step (2) is embedded the described high speed parallel robot of step (1) virtual machine real electric control system, obtain high speed parallel robot virtual machine real electric training platform;
(4) from the described high speed parallel robot of step (3) virtual machine real electric training platform, obtain movement locus (the position expectation value q of expectation
d, the speed expectation value
The acceleration expectation value
) and actual motion track (position actual value q, the speed actual value of the empty machine of corresponding high speed parallel robot
) some groups of data (with reference to Fig. 2), as training sample;
(5) the described adaptive model based control based on artificial neural network of the described training sample input step of step (4) (2) is trained it, and adjust with reference to index as it with the output torque of the linear scaling differential feedback controller described in the step (2), constantly adjust connection weights and the threshold value of artificial neural network, the control action for the treatment of described linear scaling differential feedback controller is during to insignificant degree, and training process finishes;
(6) cast out step (5) training based on the linear scaling differential feedback controller in the adaptive model based control of artificial neural network, keep described neural network formula robot inverse dynamics model and approach device, and solidify connection weights and the threshold value that described neural network formula robot inverse dynamics model approaches artificial neural network in the device, namely get neural network formula computed moment control device.
For described training method, it should be noted that the purpose of using described linear scaling differential feedback controller is to have track following performance preferably in order to guarantee in the initial stage robot of artificial neural network training; About the training to described adaptive model based control based on artificial neural network, basic process is, the training initial stage, and a little less than the effect of neural network, joint of robot outgoing position q and speed
Fail the expectation value q of pursuit movement planner planning
dAnd
, big tracing deviation appears, the regulating action of passing ratio differentiation element, and along with going deep into gradually of training, this deviate reduces gradually; Meanwhile, get τ
fAs neural network formula robot inverse dynamics model approach device with reference to amount, set up the error function of this neural network, and revise network along the direction that error function reduces network connection weights gradient by the learning rate of appointment and respectively connect weights.
Why make the choice of (6), be because, training latter stage, robot trajectory's tracking performance significantly improves, and it is very faint that the effect of described linear scaling differential feedback controller becomes, in the extent of the error that allows, can ignore its effect, learn accordingly, (6) described neural network formula robot inverse dynamics model approaches the inverse dynamics characteristic that device has reacted the empty machine of high speed parallel robot, like this, obtained the model of neural network formula computed moment control device.Treat next step, can be used as controller and use.
Particularly, for training platform proposed by the invention and the realization of training method, can adopt as shown in Figure 3 BP neural network as the kernel of described adaptive model based control based on artificial neural network, obtain the neural network training structure of inverse dynamics model, this network contains two hidden layers, the number of input can be determined according to the number in high speed parallel robot driving joint and the training sample type of choosing, be output as the single output of moment values, the number of hidden layer neural unit determines according to actual conditions, further adjusts definite under also can be after the training undesirable situation.
Between the empty machine of high speed parallel robot and neural network, be the performance of bringing into play real electric control system and the transmission that realizes signal, exist and drive interface, specifically as shown in Figure 4, distribution changes into steering order to control moment τ through motion controller, send to driver, signal (analog quantity to driver output, Torque Control employing analog quantity) carries out the A/D conversion, be convenient to Computer Processing, with the input of the signal after conversion computing machine, drive each the kinematic axis motion of the empty machine of high speed parallel robot by high speed parallel robot virtual reality module soft interface, realize that real electricity is to the butt joint of empty machine.
Through above training process, neural network formula robot inverse dynamics model after the training end is approached the neural network structure of device and respectively connects weights curing, and in the embedding torque controller, carrying as shown in Figure 5 on the high speed parallel robot virtual machine real electric control system experiment porch of neural network formula computed moment control device, implement the experiment of Torque Control high speed parallel robot, also can make entity high speed parallel robot model machine correspondingly further, adopt the neural network formula computed moment control device and the open motion control drive system that obtain, finish the exploitation of its control system.
For satisfying the high-quality Torque Control demand of high speed parallel robot, set up the controller training experiment platform of " robot is virtual; automatically controlled actualization ", proposition is based on the computed moment control device training method of neural network, have a strong impact on when having compensated parallel robot high-speed motion in the computed moment control control accuracy the time become non-linear negative factor.This design is further shortened its controller construction cycle under the condition that satisfies high speed parallel robot high precision control requirement, reduce design cost.Building and the training method of controller of the virtual machine real electric experiment porch that the design proposes, can extend to entire machine people controller design field, comprise serial machine people, parallel robot and the design of series-parallel robot controller, also comprise different control models, as the design of the controller under the control modes such as position control, speed control, Torque Control and the control of full cut-off ring, or even other causes the plant equipment controller design field of control accuracy difference because of control model out of true.