CN104516268A - Robot calibrate error compensation method based on fuzzy nerve network - Google Patents

Abstract

The invention discloses a robot calibrate error compensation method based on a fuzzy nerve network; the method comprises the following steps: compiling a homogeneous transformation matrix between adjacent link rods; calculating a kinetics equation and an error equation general formula of a robot end performer according to the transformation matrix; generating a kinetics equation and an error equation of an angle [theta i] according to a geometry parameter nominal value, the kinetics equation and the error equation general formula; compensating a first error compensation for the robot according to the kinetics equation and error equation; applying a fuzzy nerve network model to carry out a second error compensation for the robot; the method can enable the robot error compensation model to be faster, and more accurate with strong robustness.

Description

A kind of Robot calibration error compensating method based on fuzzy neural network

Technical field

The invention belongs to field of machining and machinebuilding technology, relate generally to a kind of Robot calibration error compensating method based on fuzzy neural network.

Background technology

Its repetitive positioning accuracy of robot of robot manufacturer production is higher now, and absolute fix precision is lower.Along with robot off-line programming technology is more and more extensive, improve robot absolute fix precision and become a wherein key technical problem.Industrial robot is demarcated and is generally divided into modeling, measurement, parameter identification and compensation four step, and wherein compensating is the final step of demarcating.Because robot also exists positioning error, alignment error and operating error (transmission, distortion equal error) in manufacture and installation process, make active machine people absolute fix precision low, need to compensate error, usually take to separate Inverse Kinematics Problem to compensate joint angle, general employing Newton-Raphson method solves inverse compensation problem, but the method has following 2 deficiencies: geometric error must be little, otherwise can not ensure algorithm convergence; Calculated amount is large, not easily online real-Time Compensation.

For the deficiency of this method, researchist proposes some new methods and realizes robot localization compensation of error.Application number is 201110113246.6, disclose a kind of three-dimensional grid precision compensation method for industrial robot, the method utilizes industrial robot to have the characteristic of higher repetitive positioning accuracy, adopt the relation that laser tracker theorizes between coordinate and actual location coordinate, for any point in certain the cube grid divided in the enveloping space, , the theoretical coordinate of Spatial Interpolation Method to robot is adopted to revise, complete the absolute fix accuracy compensation of robot at this point, the shortcoming of this kind of method is the coordinate of major part point in robot kinematics by minority by measuring accurate point coordinate obtains through interpolation, because robot inaccuracy has nonlinear feature, so interpolation point cannot meet the precision of robot off-line programming.Document is called " BP neural networks compensate parallel robot positioning error " (" optical precision engineering ", 2008,16th volume the 5th phase, 878-883 page), the error of nerual network technique to 6DOF high-precision parallel robot end pose is adopted to compensate, but neural network does not have classification capacity, the sample newly added will affect the network of learning success, so require that the number portraying the feature of each input amendment also must be identical.

Summary of the invention

Fundamental purpose of the present invention is to provide a kind of Robot calibration error compensating method based on fuzzy neural network, and it can overcome existing defect, realizes the Robot calibration error compensation of high precision and robustness.

For achieving the above object, the present invention adopts following technical scheme:

Write the homogeneous transform matrix between adjacent links;

According to kinematical equation and the error equation general formula of described transformation matrix calculating robot end effector;

Generate about angle θ according to geometric parameter nominal value, described kinematical equation and described error equation general formula _ikinematical equation and error equation;

According to described kinematical equation and described error equation, first time error compensation is carried out to robot;

Application fuzzy neural network model carries out second time error compensation to robot.

Preferably, application fuzzy neural network model carries out second time error compensation to robot, comprising:

The linguistic variable obtained from kinematics parameters is input to respectively in corresponding node;

Described node is represented with a subordinate function respectively;

The radial basis unit of each node is exported according to the number of fuzzy rules of system;

The number of fuzzy rules calculating each node accounts for the number percent of the total number of fuzzy rules of system;

According to the error after described percentage calculation second compensation;

If the error amount after described second compensation meets parameter predetermined in advance, then retain this model; If do not met, then model is optimized further.

Preferably, the homogeneous transform matrix between adjacent links writes according to robot motion's theory.

Preferably, according to kinematical equation and error equation, error compensation is for the first time carried out to robot and be specially: the geometric parameter error being solved robot by the method for least square method solve linear equations, parameter identification is carried out to correlation parameter.

Beneficial effect of the present invention:

Fuzzy reasoning does not need to set up target compensation mathematical models, is using the experience of people as knowledge model, using fuzzy logic inference as the intelligent algorithm of compensation model; Neural network has stronger self-learning function, can Approximation of Arbitrary Nonlinear Function, has adaptive ability and do not rely on the characteristic of model as error compensation.Second combination is used for the error compensation in Robot calibration process, the advantage of two kinds of modeling strategy can be played, make robot inaccuracy compensation model more fast, accurately, strong robustness.

Accompanying drawing explanation

Fig. 1 is a kind of Robot calibration error compensating method process flow diagram based on fuzzy neural network of the present invention.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Shown in figure 1, the invention discloses a kind of Robot calibration error compensating method based on fuzzy neural network, the method comprises the steps:

S10: write the homogeneous transform matrix between adjacent links;

The homogeneous transform matrix between adjacent links is write according to robot motion's theory.

S20: according to kinematical equation and the error equation general formula of described transformation matrix calculating robot end effector;

S30: generate about angle θ according to geometric parameter nominal value, described kinematical equation and described error equation general formula _ikinematical equation and error equation;

Due to the length of connecting rod nominal value of robot , connecting rod offset nominal value , torsional angle nominal value be constant with end effector length l, therefore, first can input above-mentioned geometric parameter nominal value in error equation general formula, one can be generated about joint angle θ _ikinematical equation and error equation:

T＝f ₁(θ _i) （1）

Δp＝f ₂(θ _i) （2）

S40: first time error compensation is carried out to robot according to described kinematical equation and described error equation;

By the group end actual position coordinate p of laser tracker robot measurement in work space ^c, and the theoretical joint angle value corresponding to each measuring point of record is input in described kinematical equation and error equation and calculates end nominal position p ⁿ, thus obtain deviations of actual position amount Δ p; Solved the geometric parameter error of robot by the method for least square method solve linear equations, parameter identification is carried out to correlation parameter.

S50: application fuzzy neural network model carries out second time error compensation to robot;

By the geometric parameter error compensation that obtains in the kinematics parameters in robot controller, application fuzzy neural network model carries out second time error compensation to robot, described fuzzy neural network model adopts 5 Rotating fields, is respectively input layer, subordinate function layer, T-norm layer, normalization layer and output layer.Concrete compensation process is:

S501: the linguistic variable obtained from kinematics parameters is input to respectively in corresponding node;

At input layer, each linguistic variable x of input r ₁, x ₂..., x _rbe input to respectively in corresponding node and go, in this embodiment, the linguistic variable of input is robot end pose parameter x, y, z, θ _x, θ _y, θ _z, wherein r is 6.

S502: described node is represented with a subordinate function respectively;

Subordinate function layer, this layer of each node represents a subordinate function respectively, and the following Gaussian function of this subordinate function represents:

${\mathrm{\μ}}_{\mathrm{ij}}\left(x_j\right)=\exp [-\frac{{(x_i-c_{\mathrm{ij}})}^2}{{\mathrm{\σ}}_j^2}],i=\mathrm{1,2},...,r,j=\mathrm{1,2},...,u---\left(3\right)$

Wherein: μ _ijx _ia jth subordinate function, c _ijx _ithe center of a jth subordinate function, σ _jx _ithe width of a jth Gauss member function, r is input variable number, and u is the quantity (number of fuzzy rules that also representative system is total) of subordinate function;

S503: the radial basis unit exporting each node according to the number of fuzzy rules of system;

T-norm layer, each node represents the IF-part in a possible fuzzy rule respectively, therefore this node layer reflects the number of fuzzy rules of system, and the output of jth rule can be expressed as:

${\mathrm{\φ}}_j=\exp [-\frac{\underset{i=1}{\overset{r}{\mathrm{\Σ}}}{(x_i-c_{\mathrm{ij}})}^2}{{\mathrm{\σ}}_j^2}]=\exp [-\frac{\left|\right|X-C_j\left|\right|}{{\mathrm{\σ}}_j^2}],j=\mathrm{1,2},...,u---\left(4\right)$

In formula, X=(x ₁, x ₂..., x _r), C _j=(x _1j, x _2j..., x _rj) be the center of a jth radial basis unit, as can be seen from formula, each node on behalf of this layer a radial basis unit, and namely radial basis unit number is identical with number of fuzzy rules.

S504: the number of fuzzy rules calculating each node accounts for the number percent of the total number of fuzzy rules of system;

${\mathrm{\Ψ}}_j=\frac{{\mathrm{\φ}}_j}{\underset{k=1}{\overset{u}{\mathrm{\Σ}}}{\mathrm{\φ}}_k},j=\mathrm{1,2},...,u---\left(5\right)$

S505: according to the error after described percentage calculation second compensation;

Output layer, each node represents an output variable, and output variable is error compensation parameter (Δ x, Δ y, Δ z, the Δ θ of end pose _x, Δ θ _y, Δ θ _z), this output variable is the superposition of all input signals

$y_p\left(X\right)=\underset{k=1}{\overset{u}{\mathrm{\Σ}}}{w}_k^p{\mathrm{\Ψ}}_k,p=\mathrm{1,2},...,z---\left(6\right)$

In formula, y is the output of variable, and the error amount namely after second compensation is the connection weight of THEN-part (result parameter) or a kth rule.And

$w_k^p={\mathrm{\α}}_{k0}^p+{\mathrm{\α}}_{k1}^p{x}_1+...+{\mathrm{\α}}_{\mathrm{kr}}^p{x}_r,k=\mathrm{1,2},...,u---\left(7\right)$

In formula, ${\mathrm{\α}}_{\mathrm{kr}}^p(k=\mathrm{1,2},...,u;i=\mathrm{0,1},...,r;p=\mathrm{1,2},...,z)$ It is real-valued parameter.By formula (4), (5) and formula (7) substitute into formula (6):

$y_p\left(X\right)=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}[({\mathrm{\α}}_{i0}^p+{\mathrm{\α}}_{i1}^p{x}_1+...+{\mathrm{\α}}_{\mathrm{ir}}^p{x}_r)\exp (-\frac{{\left|\right|X-C_i\left|\right|}^2}{{\mathrm{\σ}}_i^2})}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\exp (-\frac{{\left|\right|X-C_i\left|\right|}^2}{{\mathrm{\σ}}_i^2})}---\left(8\right)$

S506: if the error amount after described second compensation meets parameter predetermined in advance, then retain this model; If do not met, then model is optimized further.

Observe the error after compensating and whether meet parameter request predetermined in advance, if do not met, revise model parameter, model is optimized further, if met, retain this model; By large sample measurement data to this model training, until be less than in certain limit for the error of these samples, then demarcate end.

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (4)

1., based on a Robot calibration error compensating method for fuzzy neural network, it is characterized in that, comprise the steps:

Write the homogeneous transform matrix between adjacent links;

According to kinematical equation and the error equation general formula of described transformation matrix calculating robot end effector;

Generate about angle θ according to geometric parameter nominal value, described kinematical equation and described error equation general formula _ikinematical equation and error equation;

According to described kinematical equation and described error equation, first time error compensation is carried out to robot;

Application fuzzy neural network model carries out second time error compensation to robot.

2. a kind of Robot calibration error compensating method based on fuzzy neural network as claimed in claim 1, is characterized in that: described application fuzzy neural network model carries out second time error compensation to robot, comprising:

The linguistic variable obtained from kinematics parameters is input to respectively in corresponding node;

Described node is represented with a subordinate function respectively;

The radial basis unit of each node is exported according to the number of fuzzy rules of system;

The number of fuzzy rules calculating each node accounts for the number percent of the total number of fuzzy rules of system;

According to the error after described percentage calculation second compensation;

If the error amount after described second compensation meets parameter predetermined in advance, then retain this model; If do not met, then model is optimized further.

3. a kind of Robot calibration error compensating method based on fuzzy neural network as claimed in claim 1, is characterized in that: the homogeneous transform matrix between described adjacent links writes according to robot motion's theory.

4. a kind of Robot calibration error compensating method based on fuzzy neural network as claimed in claim 1, it is characterized in that: describedly according to kinematical equation and error equation, first time error compensation is carried out to robot and be specially: the geometric parameter error being solved robot by the method for least square method solve linear equations, parameter identification is carried out to correlation parameter.