CN103381603B - The active obstacle planing method of deicing high-voltage power transmission line inspection robot - Google Patents

Abstract

The invention discloses the active obstacle planing method of deicing high-voltage power transmission line inspection robot, comprise the steps: step 1: utilize the laser radar detection environmental information being arranged on mechanical arm tail end, obtain robot and run preceding object signal; Step 2: according to the obstacle signal under the difference of mechanical arm tail end current location and desired locations and present case, utilize fuzzy programming device to cook up the fuzzy programming angle running front mechanical arm; Step 3: utilize particle cluster algorithm to fuzzy programming angle on-line optimization, obtains the population fuzzy programming angle running front mechanical arm; Step 4: utilize neural network adaptive controller to obtain each joint control moment, instruct mechanical arm action.The Fuzzy Programming adopted according to deicing inspection robot the present situation, can carry out obstacle detouring programmed decision-making in real time, overcomes inaccuracy and the hysteresis quality of perception information.Meanwhile, particle cluster algorithm can on-line optimization fuzzy programming angle, makes track more smooth and redundancy is less.

Description

The active obstacle planing method of deicing high-voltage power transmission line inspection robot

Technical field

The present invention is mainly concerned with the independent navigation field of industrial special robot, particularly the active obstacle planing method of deicing high-voltage power transmission line inspection robot.

Background technology

Ultra-high-tension power transmission line is responsible for the important task of electric power transfer, and its safe and reliable operation is directly connected to the stable sound development of a national economy.For ensureing its safe and reliable operation, need to carry out periodic detection and maintenance.Monthly maked an inspection tour by tower with ascender line by two people by regulatory requirements ultra-high-tension power transmission line, annual human and material resources costly.And in the winter time under severe hazardous environment, the ultra-high-tension power transmission line accumulated snow that freezing sleet can cause and icing, usually cause line tripping, insulator arc-over, broken string, fall the accident such as bar and communication disruption.China is one of higher country of ultra-high-tension power transmission line Harm Accident probability of happening, for ensureing that transmission of electricity safety alleviates ultra-high-tension power transmission line accumulated snow icing problem in winter, usually two kinds of measures are taked: great current deicing technology and deicing of manually reaching the standard grade, the former causes serious energy loss, and not only labour intensity is large for the latter, operating efficiency low but also have larger danger.Therefore, a kind of safety of exploitation, economic, the deicer deicing that replaces manually reaching the standard grade efficiently extremely have realistic meaning.

Deicing inspection robot has line walking and deicing two kinds of functions concurrently, reaches the standard grade compared with deicing with Traditional Man, the work of deicing inspection robot without the need to stop supply of electric power, security performance high, can work continuously, do not affect operation of power networks, there is boundless application prospect.When deicing inspection robot carries out harmless deicing operation in adverse circumstances, there is special requirement to aspects such as its structure, function, power, communications.In addition, realize that deicing inspection robot carries out on a large scale in bad weather environment, for a long time online deicing operation, it just must be made to possess autonomous walking online and the function of obstacle detouring.Autonomy-oriented and the intelligent concentrated reflection of deicing inspection robot are accurate cognitive disorders thing also steady obstacle detouring in particular surroundings.

As plan objects, deicing inspection robot mechanical arm be one become when having, the nonlinear system of the Multiinputoutput of close coupling, its Consideration is more, and programmed decision-making is very complicated.In prior art, deicing inspection robot is usually operated on the severe ultra-high-tension power transmission line of weather conditions, outdoor unstructured moving grids has diversity, complexity, randomness and uncertainty, scenery has different scenes in different environments, as staggered chaotic at shaft tower place in ultra-high-tension power transmission line, rise and fall indefinite, electrical equipment random distribution, electrical equipment and line ice coating exist each other and block.In addition, the dynamic change of illumination, background and weather all has a huge impact the normal work of deicing inspection robot.Traditional Visual servoing control planning strategy is difficult to deal with so complicated working condition, and sensor-based method for planning track adopts repeatedly fitting of a polynomial usually, although the method can generate level and smooth planned trajectory, but barrier on path is not taken into account, do not have and keep away barrier obstacle crossing function.Therefore, obtain ambient condition information by Laser Radar Scanning, design a kind of simple and reliable, real-time good, the deicing high-voltage power transmission line inspection robot obstacle detouring planing method of being convenient to realize, can deal with multiclass complex environment is the key technology difficult problem ensureing that deicing inspection robot normally effectively works.

Summary of the invention

The invention provides the active obstacle planing method of deicing high-voltage power transmission line inspection robot, its object is to overcome the deficiencies in the prior art, adopt the online obstacle detouring planing method of particle cluster algorithm and fuzzy logic, under realizing accessible situation, desired trajectory is effectively followed the tracks of, ensure the accessibility in each orientation of deicing inspection robot mechanical arm, when deicing inspection robot has the danger of chance barrier, can implement effectively to keep away barrier according to human expert's experience, obstacle detouring, thus reach the active obstacle planing method of the deicing high-voltage power transmission line inspection robot had compared with high-intelligentization.

The active obstacle planing method of deicing high-voltage power transmission line inspection robot, comprises the steps:

Step 1: utilize the laser radar detection ambient condition information be arranged on mechanical arm tail end, obtains robot and runs preceding object signal B;

Step 2: deicing robot observes by the camera carried the position running front cable, 30cm place, robot forearm front cable is set to the desired locations (x of mechanical arm tail end current time _g, y _g); According to the difference (e of mechanical arm tail end current location and desired locations _x, e _y) and current obstacle signal B, utilize fuzzy programming device to cook up the fuzzy programming angle in each joint of mechanical arm before traffic direction;

Step 3: can principle be reached according to shortest path and target, utilize particle cluster algorithm to carry out on-line optimization to fuzzy programming angle, obtain the population fuzzy programming expected angle in each joint of mechanical arm before traffic direction;

Step 4: using the input of each joint population fuzzy programming angle as neural network adaptive controller, neural network adaptive controller exports each joint control moment τ, instructs mechanical arm action.

Obtain robot in described step 1 to run preceding object signal B and refer to that the obstacle detouring according to setting judges critical distance A, judge that the obstacle detouring whether distance of the Obstacle Position that mechanical arm tail end current location and Airborne Lidar measure exceedes setting judges critical distance A:

e _xb＝|x-x _b| (1)

e _yb＝|y-y _b| (2)

Wherein, x _b, y _bfor the two-dimensional coordinate position of barrier, x, y are mechanical arm tail end current location, work as e _xb≤ A and e _ybduring≤A, judge that deicing inspection robot runs front and has chance barrier dangerous, export obstacle signal B=1, otherwise B=0.

According to the difference (e of mechanical arm tail end current location and desired locations in described step 2 _x, e _y) and present case under obstacle signal B, utilize fuzzy programming device to cook up mechanical arm each joint fuzzy programming angle when deicing robot travels forward, detailed process is as follows:

1) determine fuzzy programming device input information, and Fuzzy processing carried out to input data:

By mechanical arm tail end control information (e _x, e _y) and obstacle signal B input as fuzzy programming device:

$\left\{\begin{array}{c}{e}_x=x-x_g,e_x\∈\[-0.4,0.36\]\\ e_y=y-y_g,e_y\∈\[-0.4,0.56\]\end{array}\right.,B\∈\[0,1\]$

(x _g, y _g) be set to the desired locations of mechanical arm tail end current time for 30cm place, robot forearm front cable;

Fuzzy processing is carried out, mechanical arm tail end error of coordinate (e to fuzzy programming device input information _x, e _y) corresponding 3 Fuzzy Linguistic Variable { N, Z, P}={ " bear ", " zero ", " just " }, Fuzzy Linguistic Variable N, Z and P respectively with membership function μ _n, μ _zand μ _pone_to_one corresponding, the codomain scope of membership function is [0,1]; The input variable of membership function is mechanical arm tail end error of coordinate (e _x, e _y):

$\begin{array}{cc}{\mathrm{\&mu;}}_N\left(e_x\right)=\{\begin{array}{cc}1& e_x\&le;-0.03\\ \frac{e_x}{-0.03}& -0.03<e_x\&le;0\\ 0& e_x>0\end{array}---\left(3\right)& {\mathrm{\&mu;}}_N\left(e_y\right)=\left\{\begin{array}{cc}1& e_y\&le;-0.03\\ \frac{e_y}{-0.03}& -0.03<e_y\&le;0\\ 0& e_y>0\end{array}\right.\end{array}---\left(6\right)$

$\begin{array}{cc}{\mathrm{\&mu;}}_P\left(e_x\right)=\{\begin{array}{cc}0& e_x\&le;0\\ \frac{e_x}{0.03}& 0<e_x\&le;0.03\\ 1& e_x>0.03\end{array}---\left(5\right)& {\mathrm{\&mu;}}_P\left(e_y\right)=\left\{\begin{array}{cc}0& e_y\&le;0\\ \frac{e_y}{0.03}& 0<e_y\&le;0.03\\ 1& e_y>0.03\end{array}\right.\end{array}---\left(8\right)$

Choosing membership function value is the non-zero Fuzzy Linguistic Variable Fuzzy Linguistic Variable as input information;

2) output Fuzzy Linguistic Variable is obtained according to fuzzy rule;

According to Fuzzy Linguistic Variable and the obstacle signal B of input information, inquiry fuzzy rule design table, as shown in table 1, exported Fuzzy Linguistic Variable { None accordingly, N, P}, wherein, None represents that mechanical arm is failure to actuate, and N represents that mechanical arm operating angle is negative, and P represents that mechanical arm operating angle is just;

Table 1 fuzzy rule designs

3) fuzzy reasoning is carried out to output Fuzzy Linguistic Variable, obtain and export membership function value corresponding to Fuzzy Linguistic Variable;

Adopt Mamdani reasoning, the membership function exporting Fuzzy Linguistic Variable corresponding is as follows:

$\begin{array}{cc}{\mathrm{\&mu;}}_N\left({\mathrm{\&theta;}}_{f1}\right)=\{\begin{array}{cc}\frac{{\mathrm{\&theta;}}_{f1}}{-10}& \left(-10\&le;{\mathrm{\&theta;}}_{f1}<0\right)\\ 0& \left(0\&le;{\mathrm{\&theta;}}_{f1}\&le;10\right)\end{array}---\left(9\right)& {\mathrm{\&mu;}}_N\left({\mathrm{\&theta;}}_{f2}\right)=\{\begin{array}{cc}\frac{{\mathrm{\&theta;}}_{f2}}{-10}& \left(-10\&le;{\mathrm{\&theta;}}_{f2}<0\right)\\ 0& \left(0\&le;{\mathrm{\&theta;}}_{f2}\&le;10\right)\end{array}---\left(11\right)\end{array}$

$\begin{array}{cc}{\mathrm{\&mu;}}_P\left({\mathrm{\&theta;}}_{f1}\right)=\{\begin{array}{cc}0& \left(-10\&le;{\mathrm{\&theta;}}_{f1}<0\right)\\ \frac{{\mathrm{\&theta;}}_{f1}}{10}& \left(0\&le;{\mathrm{\&theta;}}_{f1}\&le;10\right)\end{array}---\left(10\right)& {\mathrm{\&mu;}}_P\left({\mathrm{\&theta;}}_{f2}\right)=\{\begin{array}{cc}0& \left(-10\&le;{\mathrm{\&theta;}}_{f2}<0\right)\\ \frac{{\mathrm{\&theta;}}_{f1}}{10}& \left(0\&le;{\mathrm{\&theta;}}_{f2}\&le;10\right)\end{array}---\left(12\right)\end{array}$

With e _xand e _ycorresponding membership function value under each self-corresponding Fuzzy Linguistic Variable, respectively to e _xand e _ythe union of corresponding all membership function values is tried to achieve the first common factor and second and is occured simultaneously; First common factor, the second common factor and B are asked the 3rd common factor, obtains an exact value; Occur simultaneously respectively to the μ of two outputs with the 3rd _n(θ _f1) and μ _p(θ _f1) cut, obtain the membership function value exporting Fuzzy Linguistic Variable;

4) adopt sharpening interface to obtain and export angle;

Because initial conditions may take multiple input Fuzzy Linguistic Variable simultaneously, according to the derivation of fuzzy rule, multiple Output rusults will be formed, therefore need to carry out precision to the result of fuzzy reasoning;

First the maximum membership degree functional value got in each Output rusults in two fuzzy programming angles compares, the Output rusults at the maximum place of membership function value is selected to export as final result, if in more each Output rusults process, occur that the maximum membership degree function of multiple result is maximum side by side and equal, then select to get less the exporting as final result of fuzzy language amount in Output rusults; Then the membership function value in this output fuzzy language result is substituted into formula 9-12 to invert function, obtain two joint actual output angle; Fuzzy language amount presses N, None, P ascending order.

It is as follows that particle cluster algorithm in described step 3 carries out on-line optimization process to fuzzy programming angle:

1): optimization object is selected: arrange self-correcting parameter k=(k ₁, k ₂) k ₁, k ₂∈ [0,2], the fuzzy programming after optimization is expected, and angle is for (k ₁θ _f1, k ₂θ _f2);

2) update mode is determined: adopt elementary particle group to optimize update mode, be expressed as:

v _ij(n+1)＝wv _ij(n)+r ₁(n)c ₁[pbest _ij(n)-x _ij(n)]+r ₂(n)c ₂[gbest _ij(n)-x _ij(n)] (13)

x _ij(n+1)＝x _ij(n)+v _ij(n+1) (14)

Wherein: i is for initializing correction parameter group number, and j is the dimension often organizing correction parameter, v _ijn () is the diverse vector of correction parameter group i jth dimension in n-th iteration, x _ijn () is the numerical value of correction parameter group i jth dimension in n-th time, r ₁, r ₂for independent random function, c ₁, c ₂for acceleration weight, c ₁, c ₂equal value is 1.2, w is inertia weight;

The mode that inertia weight w linearly successively decreases upgrades, and is shown below:

$w=w_{max}-n\frac{w_{\max }-w_{\min }}{n_{max}}---\left(15\right)$

Wherein: n is current iteration number of times, n _maxfor total iterations, w _maxbe 0.9, w _minbe 0.4;

3) fitness function design:

Min:f(t)＝s ₁f ₁(t)+s ₂f ₂(t) (16)

Wherein, s ₁be 0.4, s ₂be 0.6;

In any t, integration is minimum in time for the distance of mechanical arm tail end and initial point position, fitness function f ₁(t) be:

$Min:f_1\left(t\right)=\underset{0}{\overset{t}{\&Integral;}}\sqrt{{\&lsqb;x\left(t\right)-x_o\&rsqb;}^2+{\&lsqb;y\left(t\right)-y_o\&rsqb;}^2}dt---\left(17\right)$

Impact point attracts then to be presented as that between mechanical arm tail end and impact point, air line distance is minimum, fitness function f ₂(t) be;

$Min:f_2\left(t\right)=\sqrt{{\&lsqb;x\left(t\right)-x_g\&rsqb;}^2+{\&lsqb;y\left(t\right)-y_g\&rsqb;}^2}---\left(18\right)$

Wherein: x _o, y _ofor mechanical arm tail end initial point position, x (t), y (t) are t mechanical arm tail end position, x _gand y _grepresent aiming spot;

4) population on-line optimization:

Step1: population initializes, produces initial value and the diverse vector of correction parameter group, by (the k of current time t ₁θ _f1, k ₂θ _f2) substitute into fitness function formula 16 and calculate;

Step2: at iterations n _maxin, according to formula 13, formula 14, formula 15 upgrades correction parameter group numerical value.Each iteration all selects local optimum parameter group pbest and global optimum parameter group gbest;

Step3: after iteration completes, exports the correction parameter (k as moment t using global optimum parameter group gbest ₁, k ₂).Excitation set mechanical arm is by (k ₁θ _f1(t), k ₂θ _f2(t)) action;

Step4: mechanical arm tail end sensor obtains subsequent time t+1 terminal position, turns Step1.

Neural network adaptive controller in described step 4 comprises ratio control-K _ps and neural network control

Described neural network adaptive controller is shown below:

$\mathrm{\&tau;}=-K_{p}S+{\hat{W}}^{T}g\left(\mathrm{\&chi;}\right)---\left(19\right)$

$\stackrel{\&CenterDot;}{\hat{W}}=-\mathrm{\&Gamma;}g\left(\mathrm{\&chi;}\right)S^T---\left(20\right)$

Wherein: τ is control moment, K _pbe symmetric positive definite matrix with Γ, K _pfor ratio controling parameters is set to $\left[\begin{array}{cc}5& 0\\ 0& 5\end{array}\right],$ Γ is set to the unit matrix of 8 × 8, g (χ)=[g ₁(χ), g ₂(χ) ..., g _n(χ)] ^t, g _j(χ) (1≤j≤r) is RBF vector, for hidden layer is to the connection weight value matrix of output layer, initial value is set to the 8 dimension matrixes in [-1,1] interior random distribution;

S is synovial membrane face e=k θ _f-θ, wherein k θ _ffor population fuzzy programming angle, k θ _f=[k ₁θ _f1, k ₂θ _f2], θ is mechanical arm current angular, θ=[θ ₁, θ ₂], Λ is symmetric positive definite matrix $\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right];$

Adaptive neural network is shown below:

$\hat{h}={\hat{W}}^{T}g\left(\mathrm{\&chi;}\right)---\left(21\right)$

Wherein: $\mathrm{\&chi;}={\&lsqb;k_1{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_{f1},k_1{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_{f1},k_2{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_{f2},k_2{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_{f2},{\mathrm{\&theta;}}_1,{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_1,{\mathrm{\&theta;}}_2,{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_2\&rsqb;}^T$ For neutral net input, be the first joint planning angular speed, be the first joint planning angular acceleration, for second joint planning angular speed, for second joint planning angular acceleration, θ ₁be the first joint action angle, be the first joint action angular speed, θ ₂for second joint operating angle, for second joint operating angle speed, for hidden layer is to the connection weight value matrix of output layer, initial value is set to the unit matrix of 8 × 8, and the attitude different according to mechanical arm and each joint angles desired value real-time update, to adapt to different demands for control; G (χ)=[g ₁(χ), g ₂(χ) ..., g _n(χ)] ^tfor RBF vector, g _j(χ) (1≤j≤r) has following form:

$g_j\left(\mathrm{\&chi;}\right)=\exp (-\frac{{||\mathrm{\&chi;}-{\mathrm{\&mu;}}_j||}^2}{{\mathrm{\&sigma;}}_j^2})---\left(22\right)$

Wherein: r is the number of hidden layer neuron; μ _jcentered by vector, σ _jfor width vector;

The output of adaptive neural network target of approaching be compensation term, compensation term is shown below:

$M\left(\mathrm{\&theta;}\right)\&lsqb;{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f)\&rsqb;+C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})\&lsqb;{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_f)\&rsqb;+G\left(\mathrm{\&theta;}\right);---\left(23\right)$

In adaptive neural network, hidden layer number r is 20, center vector μ _jbe set to the random function being uniformly distributed [-10,10], width vector σ _jall be set to 2.

The output of neutral net is the compensation term matrix in each joint, and neutral net is input as namely be go real-time to approach compensation term according to the current angular information (angle and angular speed) in population fuzzy programming angle information (angular speed and angular acceleration) and each joint $M\left(\mathrm{\&theta;}\right)\&lsqb;{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f)\&rsqb;+C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})\&lsqb;{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_f)\&rsqb;+G\left(\mathrm{\&theta;}\right).$

In control law synovial membrane face S is made up of e, and S is tending towards 0 and e just can be caused to be tending towards 0, and just can ensure tracking performance in theory, this provides in stability proof part.In this control law, first half is that ratio controls, and latter half is the compensation term that neutral net is approached, if the perfection of neutral net convergence approaches, this complication system just becomes simple ratio and controls.In formula being weight matrix, is real-time update, provides the formula of renewal s is synovial membrane face, is an approximate item of error, can according to current synovial membrane face S and input and output real-time update.Whole process is a dynamic neural network control process.

Beneficial effect

The active obstacle planing method of deicing high-voltage power transmission line inspection robot provided by the invention, adopt laser radar detection deicing inspection robot ambient condition information, overcome Changes in weather when deicing inspection robot works, illumination condition is changeable, the deficiencies such as cabinet shake, effectively can detect the barrier on ultra-high-tension power transmission line; Compared with in the past traditional visual servo obstacle detouring strategy, the Fuzzy Programming that the present invention adopts can according to deicing inspection robot the present situation, effectively carry out in real time following the tracks of, obstacle detouring programmed decision-making, overcome inaccuracy and the hysteresis quality of perception end obtaining information, make deicing inspection robot have better independence and intelligent; Particle swarm optimization algorithm is adopted to carry out real-time optimization to fuzzy programming device planning angle, when ensureing that impact point can reach, achieve mechanical arm tail end travels along path the shortest, effectively reduce the operating power consumption of deicing inspection robot, improve the flying power of deicing inspection robot; Neural network adaptive controller is adopted to implement to control to deicing inspection robot mechanical arm, this controller can real-time online self study, there is good adaptivity, particularly good control performance is revealed to this kind of nonlinear Control of multi-joint mechanical arm subject, the effective tracking to various types of signal can be realized, meet deicing inspection robot active obstacle demand.

Accompanying drawing explanation

Fig. 1 is the population fuzzy obstacle detouring Structure Planning schematic diagram that the present invention adopts;

Fig. 2 is the deicing inspection robot frame for movement schematic diagram that the present invention adopts;

Fig. 3 is deicing inspection robot obstacle detouring schematic diagram;

Fig. 4 is fuzzy logic planner internal structure schematic diagram of the present invention;

Fig. 5 is deicing inspection robot mechanical arm tail end confusion region schematic diagram;

Fig. 6 is particle cluster algorithm on-line optimization schematic flow sheet of the present invention;

Fig. 7 is neural network adaptive controller cut-away view of the present invention;

Fig. 8 is deicing inspection robot mechanical arm eight directional reachability test figure, wherein, figure (a) is top impact point, figure (b) is front upper place impact point, and figure (c) is objects ahead point, and figure (d) is front lower place impact point, figure (e) is below impact point, figure (f) is back lower place impact point, and figure (g) is rear area target point, figure (h is) back upper place impact point;

Fig. 9 is deicing inspection robot mechanical arm continuous signal tracing figure;

Figure 10 is deicing inspection robot mechanical arm obstacle performance resolution chart.

Detailed description of the invention

Below with reference to accompanying drawing and concrete case study on implementation, the present invention will be further described.

As shown in Figure 1, the present invention is the active obstacle planing method of deicing high-voltage power transmission line inspection robot, comprises the steps:

Step 1: utilize the laser radar detection ambient condition information be arranged on mechanical arm tail end, obtains robot and runs preceding object signal B;

Step 2: deicing robot, by being arranged on the camera on cabinet, being observed the position running front cable, 30cm place, robot forearm front cable is set to the desired locations (x of mechanical arm tail end current time _g, y _g); According to the difference (e of mechanical arm tail end current location and desired locations _x, e _y) and present case under obstacle signal B, utilize fuzzy programming device to cook up the fuzzy programming angle in each joint of mechanical arm before traffic direction;

Step 3: can principle be reached according to shortest path and target, utilize particle cluster algorithm to carry out on-line optimization to fuzzy programming angle, obtain the population fuzzy programming expected angle in each joint of mechanical arm before traffic direction;

Step 4: using the input of each joint population fuzzy programming angle as neural network adaptive controller, neural network adaptive controller exports and obtains each joint control moment τ, thus instruct each joint action of mechanical arm, realize the normal walking on deicing inspection robot line and obstacle detouring action.

In this example, each arm has two joints, and neural network adaptive controller exports the control moment (τ in two joints obtained on each arm ₁, τ ₂).

As shown in Figures 2 and 3, deicing inspection robot reach the standard grade work time can run into the barriers such as such as stockbridge damper, strain clamp, insulator.Mechanical mechanism according to deicing inspection robot designs, and it is as follows that the obstacle in above-mentioned steps (2) judges that link is arranged:

Deicing inspection robot is without the environmental knowledge of priori, laser radar is adopted to detect mechanical arm tail end surrounding environment, to the memoryless function of the environmental information detected, any instant, mechanical arm can only detect centered by its terminal position, radius is the information in the region of d, if find obstacle, returns corresponding complaint message.The obstacle detouring demand concrete according to deicing inspection robot arranges as follows:

The scope arranging the detection of mechanical arm is 0.03 meter, when the coordinate of mechanical arm tail end and the distance of barrier are less than 0.06 meter, judge that deicing inspection robot has and meets barrier danger.All the other situations are all safety.

e _xb＝|x-x _b| (1)

e _yb＝|y-y _b| (2)

Wherein, x _b, y _bfor the two-dimensional coordinate position of barrier, x, y are mechanical arm tail end current location, work as e _xb≤ 0.06 and e _ybwhen≤0.06, judge that deicing inspection robot runs front and has chance barrier dangerous, export obstacle signal B=1, otherwise B=0.

As shown in Figure 3 and Figure 4, in above-mentioned steps (2), fuzzy programming device partial design in Fuzzy particle swarm artificial obstacle detouring planning strategy:

1. determine fuzzy programming device input information, and Fuzzy processing carried out to input data:

Select mechanical arm tail end control information (e _x, e _y) and obstacle judge information B as fuzzy programming device input, mechanical arm each joint subsequent time operating angle (θ _f1, θ _f2) export as fuzzy programming device.Because deicing inspection robot mechanical arm is subject to environmental constraints, its end coverage has a definite limitation:

$\left\{\begin{array}{c}x\&Element;\&lsqb;-0.2,0.56\&rsqb;\\ y\&Element;\&lsqb;0,0.96\&rsqb;\end{array}\right.---\left(3\right)$

Consider that the initial point of deicing inspection robot mechanical arm tail end is (0.2,0.4) in the position of two-dimensional coordinate system, therefore, the margin of error domain of mechanical arm tail end is:

$\left\{\begin{array}{c}{e}_x=x-x_g,e_x\&Element;\&lsqb;-0.4,0.36\&rsqb;\\ e_y=y-y_g,e_y\&Element;\&lsqb;-0.4,0.56\&rsqb;\end{array}\right.---\left(4\right)$

Design obstacle judges that information B is switching variable, and domain is B ∈ [0,1].

For the consideration of security, the output angle of fuzzy programming device is unsuitable excessive.Therefore, designing fuzzy programming device output language variable field is:

u ₁,u ₂∈[-10,10] (5)

Fuzzy processing is carried out, mechanical arm tail end error of coordinate (e to fuzzy programming device input information _x, e _y) corresponding 3 Fuzzy Linguistic Variable { N, Z, P}={ " bear ", " zero ", " just " }, Fuzzy Linguistic Variable N, Z and P respectively with membership function μ _n, μ _zand μ _pone_to_one corresponding, the codomain scope of membership function is [0,1]; The input variable of membership function is mechanical arm tail end error of coordinate (e _x, e _y):

$\begin{array}{cc}{\mathrm{\&mu;}}_N\left(e_x\right)=\{\begin{array}{cc}1& e_x\&le;-0.03\\ \frac{e_x}{-0.03}& -0.03<e_x\&le;0\\ 0& e_x>0\end{array}---\left(6\right)& {\mathrm{\&mu;}}_N\left(e_y\right)=\left\{\begin{array}{cc}1& e_y\&le;-0.03\\ \frac{e_y}{-0.03}& -0.03<e_y\&le;0\\ 0& e_y>0\end{array}\right.\end{array}---\left(9\right)$

$\begin{array}{cc}{\mathrm{\&mu;}}_P\left(e_x\right)=\{\begin{array}{cc}0& e_x\&le;0\\ \frac{e_x}{0.03}& 0<e_x\&le;0.03\\ 1& e_x>0.03\end{array}---\left(8\right)& {\mathrm{\&mu;}}_P\left(e_y\right)=\left\{\begin{array}{cc}0& e_y\&le;0\\ \frac{e_y}{0.03}& 0<e_y\&le;0.03\\ 1& e_y>0.03\end{array}\right.\end{array}---\left(11\right)$

Choosing membership function value is the non-zero Fuzzy Linguistic Variable Fuzzy Linguistic Variable as input information;

Mechanical arm each joint subsequent time operating angle (θ _f1, θ _f2) corresponding two fuzzy language collection { N, P}={ " bear ", " just " }, membership function adopts ramp function, is expressed as:

${\mathrm{\&mu;}}_N\left({\mathrm{\&theta;}}_{f1}\right)=\left\{\begin{array}{cc}\frac{{\mathrm{\&theta;}}_{f1}}{-10}& \left(-10\&le;{\mathrm{\&theta;}}_{f1}<0\right)\\ 0& \left(0\&le;{\mathrm{\&theta;}}_{f1}\&le;10\right)\end{array}\right.---\left(12\right)$

${\mathrm{\&mu;}}_P\left({\mathrm{\&theta;}}_{f1}\right)=\left\{\begin{array}{cc}0& \left(-10\&le;{\mathrm{\&theta;}}_{f1}<0\right)\\ \frac{{\mathrm{\&theta;}}_{f1}}{10}& \left(0\&le;{\mathrm{\&theta;}}_{f1}\&le;10\right)\end{array}\right.---\left(13\right)$

Wherein, θ _f2membership function design same θ _f1.

2) output Fuzzy Linguistic Variable is obtained according to fuzzy rule;

According to Fuzzy Linguistic Variable and the obstacle signal B of input information, inquiry fuzzy rule design table, as shown in table 1, exported Fuzzy Linguistic Variable { None accordingly, N, P}, wherein, None represents that mechanical arm is failure to actuate, and N represents that mechanical arm operating angle is negative, and P represents that mechanical arm operating angle is just;

Table 1 fuzzy rule designs

3) fuzzy reasoning is carried out to output Fuzzy Linguistic Variable, obtain and export membership function value corresponding to Fuzzy Linguistic Variable;

Adopt Mamdani reasoning, the membership function exporting Fuzzy Linguistic Variable corresponding is as follows:

$\begin{array}{cc}{\mathrm{\&mu;}}_N\left({\mathrm{\&theta;}}_{f1}\right)=\{\begin{array}{cc}\frac{{\mathrm{\&theta;}}_{f1}}{-10}& \left(-10\&le;{\mathrm{\&theta;}}_{f1}<0\right)\\ 0& \left(0\&le;{\mathrm{\&theta;}}_{f1}\&le;10\right)\end{array}---\left(14\right)& {\mathrm{\&mu;}}_N\left({\mathrm{\&theta;}}_{f2}\right)=\{\begin{array}{cc}\frac{{\mathrm{\&theta;}}_{f2}}{-10}& \left(-10\&le;{\mathrm{\&theta;}}_{f2}<0\right)\\ 0& \left(0\&le;{\mathrm{\&theta;}}_{f2}\&le;10\right)\end{array}---\left(16\right)\end{array}$

$\begin{array}{cc}{\mathrm{\&mu;}}_P\left({\mathrm{\&theta;}}_{f1}\right)=\{\begin{array}{cc}0& \left(-10\&le;{\mathrm{\&theta;}}_{f1}<0\right)\\ \frac{{\mathrm{\&theta;}}_{f1}}{10}& \left(0\&le;{\mathrm{\&theta;}}_{f1}\&le;10\right)\end{array}---\left(15\right)& {\mathrm{\&mu;}}_P\left({\mathrm{\&theta;}}_{f2}\right)=\{\begin{array}{cc}0& \left(-10\&le;{\mathrm{\&theta;}}_{f2}<0\right)\\ \frac{{\mathrm{\&theta;}}_{f2}}{10}& \left(0\&le;{\mathrm{\&theta;}}_{f2}\&le;10\right)\end{array}---\left(17\right)\end{array}$

With e _xand e _ycorresponding membership function value under each self-corresponding Fuzzy Linguistic Variable, respectively to e _xand e _ythe union of corresponding all membership function values is tried to achieve the first common factor and second and is occured simultaneously; First common factor, the second common factor and B are asked the 3rd common factor, obtains an exact value; Occur simultaneously respectively to the μ of two outputs with the 3rd _n(θ _f1) and μ _p(θ _f1) cut, obtain the fuzzy membership function value exporting Fuzzy Linguistic Variable;

4) adopt sharpening interface to obtain and export angle;

Because initial conditions may take multiple input Fuzzy Linguistic Variable simultaneously, according to the derivation of fuzzy rule, multiple Output rusults will be formed, therefore need to carry out precision to the result of fuzzy reasoning.

First the maximum membership degree functional value got in each Output rusults in two fuzzy programming angles compares, the Output rusults at the maximum place of membership function value is selected to export as final result, if in more each Output rusults process, occur that the maximum membership degree function of multiple result is maximum side by side and equal, then select to get less the exporting as final result of fuzzy language amount in Output rusults; Then the fuzzy membership function value in this output fuzzy language result is substituted into formula 14-17 to invert function, obtain two joint actual output angle; Fuzzy language amount presses N, None, P ascending order.

Analytic explanation is carried out below with one of situation that may run in the real work of deicing inspection robot:

If mechanical arm tail end error (e _x, e _y) be (-0.02 ,-0.02), obstacle judges that B=1 is as dangerous.

Fuzzy inputing method: by (e _x, e _y) substitute into respective membership function formula 3-8, obtain e _xmembership function value in fuzzy language N situation is 0.667, and the membership function value in fuzzy language Z situation is 0.333, and the membership function value in fuzzy language P situation is 0.Due to e _ywith e _xerror is identical, and membership function is identical, therefore e _ymembership function value in fuzzy language N situation is 0.667, and the membership function value in fuzzy language Z situation is 0.333, and the membership function value in fuzzy language P situation is 0.

Fuzzy rule instructs: because the membership function value in fuzzy language P situation is 0, mamdani the reasoning results is afterwards 0, so do not do to consider.Therefore, the different fuzzy language values that output variable takies, (e are considered _x, e _y) there are four kinds of combinations, be respectively (N, N), (N, Z), (Z, N), (Z, Z).According to the fuzzy rule in table 1, obtain output variable (θ _f1, θ _f2) there are four kinds of fuzzy language results, be respectively (N, None), (None, P), (None, P), (None, None).

Fuzzy reasoning: for the first situation, according to Mamdani reasoning, (e _x, e _y) when being in (N, N) situation, the membership function value of input is (0.667,0.667), and B=1, during reasoning, first will input e _xget membership function value μ during fuzzy language amount N _n(e _x)=0.667 and e _xall membership function μ _n(e _x) ∪ μ _z(e _x) ∪ μ _p(e _x)=1 is done first and is occured simultaneously, and obtains α _ex=0.667, in like manner, e _ywith e _yall membership functions are done second and are occured simultaneously, and can obtain α _ey=0.667, in addition, the height α of B _bbe 1, by α _ex, α _eyand α _bmake the 3rd common factor, obtain α _ex∩ α _ey∩ α _b=0.667, then the 3rd exact value occured simultaneously is cut output membership function formula 9-12, obtain the exact value exporting each fuzzy programming angle.This process is known μ _n(θ _f1), μ _p(θ _f1), μ _n(θ _f2) and μ _p(θ _f2) be 0.667, formula 9-12 is inverted function, obtains the first joint θ _f1to export fuzzy language be angle value under N be angle value under-6.67, P is 6.67, second joint θ _f2to export fuzzy language be angle value under N be angle value under-6.67, P is 6.67.Guide according to fuzzy rule, the first situation exports as output time (N, N) is (N, None), therefore, and output variable (θ _f1, θ _f2) membership function value be (0.667,0), export angle correspond to (-6.67,0).

In like manner can be input as (N, Z), (Z, N), output language variable during (Z, Z) situation is respectively (None, P), (None, P) (None, None), its output variable (θ _f1, θ _f2) membership function value be respectively (0,0.333), (0,0.333), (0,0).

Sharpening interface: adopt SOM clarification method, in these Output rusults, get maximum the comparing of membership function in two fuzzy programming angles, the result at the place selecting membership function maximum is as final output.In this example, be input as (N, the membership function value of output (N, None) corresponding time N) is maximum (0.667>0.333=0.333>0), does not occur the situation that output function is maximum and equal.Therefore get conduct under being input as (N, N) situation to export, therefore the output fuzzy language of mechanical arm is (N, None), corresponding output angle determined value is (-6.67,0).

As shown in Figure 6, in above-mentioned steps (3), in Fuzzy particle swarm artificial obstacle detouring planning strategy, population on-line optimization partial design is as follows:

1. optimization object is selected: by arranging self-correcting parameter k=(k after fuzzy programming device ₁, k ₂) k ₁, k ₂∈ [0,2], utilizes particle cluster algorithm to carry out on-line optimization to it, and the fuzzy programming after optimization is expected, and angle is for (k ₁θ _f1, k ₂θ _f2).

2. update mode is determined: adopt elementary particle group to optimize update mode, be expressed as:

v _ij(n+1)＝wv _ij(n)+r ₁(n)c ₁[pbest _ij(n)-x _ij(n)]+r ₂(n)c ₂[gbest _ij(n)-x _ij(n)] (18)

x _ij(n+1)＝x _ij(n)+v _ij(n+1) (19)

Wherein: i is for initializing correction parameter group number, and j is the dimension often organizing correction parameter, v _ijn () is the diverse vector of correction parameter group i jth dimension in n-th iteration, x _ijn () is the numerical value of correction parameter group i jth dimension in n-th time, r ₁, r ₂for independent random function, c ₁, c ₂for acceleration weight, c ₁, c ₂equal value is 1.2, w is inertia weight;

The mode that inertia weight w linearly successively decreases upgrades, and is shown below:

$w=w_{max}-n\frac{w_{max}-w_{min}}{n_{max}}---\left(20\right)$

Wherein: n is current iteration number of times, n _maxfor total iterations.W _maxbe 0.9, w _minbe 0.4.

3. fitness function design:

In order to reduce the redundancy of mechanical arm tail end track, devising the fitness function of integration schedules length and impact point attraction, realizing path planning at guarantee mechanical arm when impact point can be arrived the shortest.

Mechanical arm can be expressed as through shortest path: in any t, and integration is minimum in time for the distance of mechanical arm tail end and initial point position, has fitness function:

$Min:f_1\left(t\right)=\underset{0}{\overset{t}{\&Integral;}}\sqrt{{\&lsqb;x\left(t\right)-x_o\&rsqb;}^2+{\&lsqb;y\left(t\right)-y_o\&rsqb;}^2}dt---\left(21\right)$

Wherein: x _o, y _ofor mechanical arm tail end initial point position, x (t), y (t) are t mechanical arm tail end position.

Impact point attracts then to be presented as that between mechanical arm tail end and impact point, air line distance is minimum.

$Min:f_2\left(t\right)=\sqrt{{\&lsqb;x\left(t\right)-x_g\&rsqb;}^2+{\&lsqb;y\left(t\right)-y_g\&rsqb;}^2}---\left(22\right)$

Therefore, comprehensive fitness degree function is the weighting that shortest path and impact point attract, and is expressed as:

Min:f(t)＝s ₁f ₁(t)+s ₂f ₂(t) (23)

When for ensureing that target can reach, realize path planning the shortest.Therefore the weight factor of Offered target point attraction is larger.Make s ₁be 0.4, s ₂be 0.6.

4. population on-line optimization step design:

According to fuzzy programming output and particle group optimizing mode, online design Optimization Steps is as follows:

Step1: at a time t, fuzzy programming device, according to the working environment of mechanical arm, cooks up fuzzy operating angle (θ _f1(t), θ _f2(t)), meanwhile, population initializes, and produces initial value and the diverse vector of correction parameter group.

Step2: each correction parameter group forms (k by two correction parameters ₁(t), k ₂(t)), in conjunction with present Fuzzy operating angle (θ _f1(t), θ _f2(t)), substitute into fitness function formula 21 and calculate.

Step3: at iterations n _maxin, according to formula 18, formula 19, formula 20 upgrades correction parameter group numerical value.Each iteration all selects local optimum parameter group pbest and global optimum parameter group gbest.

Step4: after iteration completes, exports the correction parameter (k as moment t using global optimum parameter group gbest ₁, k ₂).Excitation set mechanical arm is by (k ₁θ _f1(t), k ₂θ _f2(t)) action.

Step5: mechanical arm tail end sensor obtains subsequent time t+1 terminal position, turns Step1.

As shown in Figure 7, in above-mentioned steps (4), neural network adaptive controller partial design is as follows:

The output of adaptive neural network can be expressed as:

$\hat{h}={\hat{W}}^{T}g\left(\mathrm{\&chi;}\right)---\left(24\right)$

Wherein: $\mathrm{\&chi;}={\&lsqb;k_1{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_{f1},k_1{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_{f1},k_2{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_{f2},k_2{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_{f2},{\mathrm{\&theta;}}_1,{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_1,{\mathrm{\&theta;}}_2,{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_2\&rsqb;}^T$ For neutral net input, for hidden layer is to the connection weight value matrix of output layer, it can according to the different attitude of mechanical arm and each joint angles desired value real-time update, to adapt to different demands for control; G (χ)=[g ₁(χ), g ₂(χ) ..., g _n(χ)] ^tfor RBF vector.G _j(χ) (1≤j≤N) has following form

$g_j\left(\mathrm{\&chi;}\right)=\exp (-\frac{{||\mathrm{\&chi;}-{\mathrm{\&mu;}}_j||}^2}{{\mathrm{\&sigma;}}_j^2})---\left(25\right)$

Wherein: N is the number of hidden layer neuron; μ _jcentered by vector, σ _jfor width vector.

Definition: e=k θ _f-θ and sliding-mode surface Λ is symmetric positive definite matrix.Then the control inputs of mechanical arm can be expressed as:

$\mathrm{\&tau;}=-K_{p}S+{\hat{W}}^{T}g\left(\mathrm{\&chi;}\right)---\left(26\right)$

$\stackrel{\&CenterDot;}{\hat{W}}=-\mathrm{\&Gamma;}g\left(\mathrm{\&chi;}\right)s^T---\left(27\right)$

Wherein: K _pbe symmetric positive definite matrix with Γ.

2. the stability of a system proves:

The kinetics equation of mechanical arm can be expressed as:

$\mathrm{\&tau;}=M\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}+C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})\stackrel{\&CenterDot;}{\mathrm{\&theta;}}+G\left(\mathrm{\&theta;}\right)---\left(28\right)$

Wherein, M (θ) is inertia matrix, for centripetal acceleration coefficient and Corioli's acceleration coefficient, G (θ) is gravity item.

Consider $S=\stackrel{\&CenterDot;}{e}+\mathrm{\&Lambda;}e,$ Therefore,

$S=\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f+\mathrm{\&Lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_f)---\left(29\right)$

$\stackrel{\&CenterDot;}{S}=\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}-\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}+\mathrm{\&Lambda;}(\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f)---\left(30\right)$

In conjunction with the Neural Network Adaptive Control input of design, formula (28) can be expressed as:

$\begin{array}{c}M\left(\mathrm{\&theta;}\right)\&lsqb;\stackrel{\&CenterDot;}{S}-{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}\left(\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f\right)\&rsqb;+C\left(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}}\right)\&lsqb;S-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}\left(\mathrm{\&theta;}-{\mathrm{\&theta;}}_f\right)\&rsqb;+G\left(\mathrm{\&theta;}\right)=\mathrm{\&tau;}\\ M\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}+C\left(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}}\right)S=\mathrm{\&tau;}-h\end{array}---\left(31\right)$

Wherein, h is $h=M\left(\mathrm{\&theta;}\right)\&lsqb;{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f)\&rsqb;+C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})\&lsqb;{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_f)\&rsqb;+G\left(\mathrm{\&theta;}\right),$ Under normal circumstances, h neural network adaptive controller exports and evaluated error sum represents.

$h=\hat{h}+\tilde{h}={\hat{W}}^{T}g\left(\mathrm{\&chi;}\right)+{\tilde{W}}^{T}g\left(\mathrm{\&chi;}\right)---\left(32\right)$

Convolution (24) and formula (26), the output of Neural Network Adaptive Control can be expressed as:

$\mathrm{\&tau;}=-k_{p}S+\hat{h}---\left(33\right)$

Formula (31) and formula (32) are substituted into formula (33),

$M\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}+C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})S=-k_{p}S+{\tilde{W}}^{T}g\left(\mathrm{\&chi;}\right)---\left(34\right)$

Get Liapunov functions:

$V=\frac{1}{2}{S}^{T}M\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}+\frac{1}{2}tr\left(\tilde{W}{\mathrm{\&Gamma;}}^{-1}\tilde{W}\right)---\left(35\right)$

Differentiate obtains:

$\begin{array}{c}\stackrel{\&CenterDot;}{V}=S^{T}M\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}+\frac{1}{2}{S}^T\stackrel{\&CenterDot;}{M}\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}+tr\left(\tilde{W}{\mathrm{\&Gamma;}}^{-1}\stackrel{\&CenterDot;}{\tilde{W}}\right)\\ =S^T\left(\mathrm{\&tau;}-f-C\left(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}}\right)S\right)+\frac{1}{2}{S}^T\stackrel{\&CenterDot;}{M}\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}+tr\left(\tilde{W}{\mathrm{\&Gamma;}}^{-1}\stackrel{\&CenterDot;}{\tilde{W}}\right)\end{array}---\left(36\right)$

Utilize skew symmetry know $S^T\stackrel{\&CenterDot;}{M}\left(\mathrm{\&theta;}\right)\stackrel{\&CenterDot;}{S}=2C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})S,$ Formula (32) and formula (33) are substituted into formula (36) can obtain:

$\begin{array}{c}\stackrel{\&CenterDot;}{V}=S^T\left(\mathrm{\&tau;}-h\right)+tr\left(\tilde{W}{\mathrm{\&Gamma;}}^{-1}\stackrel{\&CenterDot;}{\tilde{W}}\right)\\ =S^T\&lsqb;\mathrm{\&tau;}-\hat{h}-{\tilde{W}}^{T}g\left(\mathrm{\&chi;}\right)\&rsqb;+tr\left(\tilde{W}{\mathrm{\&Gamma;}}^{-1}\stackrel{\&CenterDot;}{\tilde{W}}\right)\\ =-S^T{k}_{p}S-S^T{\tilde{W}}^{T}g\left(\mathrm{\&chi;}\right)+tr\left(\tilde{W}{\mathrm{\&Gamma;}}^{-1}\stackrel{\&CenterDot;}{\tilde{W}}\right)\\ =-S^T{k}_{p}S\end{array}---\left(37\right)$

Due to negative semidefinite, and K _ppositive definite, then when time, S ≡ 0 substitutes into formula (29), known from LaSalle theorem, deicing inspection robot forearm Globally asymptotic.Namely from arbitrary initial conditions set out, all have θ → θ _d,

The output of neutral net is the compensation term matrix in each joint, and neutral net is input as namely be go real-time to approach compensation term according to the current angular information (angle and angular speed) in population fuzzy programming angle information (angular speed and angular acceleration) and each joint $M\left(\mathrm{\&theta;}\right)\&lsqb;{\stackrel{\&CenterDot;\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\stackrel{\&CenterDot;}{\mathrm{\&theta;}}-{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f)\&rsqb;+C(\mathrm{\&theta;},\stackrel{\&CenterDot;}{\mathrm{\&theta;}})\&lsqb;{\stackrel{\&CenterDot;}{\mathrm{\&theta;}}}_f-\mathrm{\&Lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_f)\&rsqb;+G\left(\mathrm{\&theta;}\right).$

In control law s is made up of e, and S is tending towards 0 and e just can be caused to be tending towards 0, and just can ensure tracking performance in theory, this provides in stability proof part before.In this control law, first half is that simple ratio controls, and latter half is the compensation term that neutral net is approached, if neutral net perfection approaches, the system of this complexity just can become simple ratio and control. being weight matrix, is real-time update, provides the formula of renewal s can find out it is an approximate product of error, can according to error current S and input and output real-time update.A dynamic neural network control process that Here it is.

Below with a concrete application example, operation of the present invention is described in detail, the mechanical arm active obstacle planing method of this patent is applied to deicing high-voltage power transmission line inspection robot and reaches the standard grade in work, mainly plans that two aspects are to embody its validity from accessible track following and chance barrier obstacle detouring.Specifically arrange as follows:

1. arrange neural network adaptive controller parameter, arranging hidden layer number N is 20, width vector μ _jbe set to the random function being uniformly distributed [-10,10], center vector σ _junification is set to 2, connects weights initial value and is arranged on [-1,1] interior random distribution.

2. arrange deicing inspection robot mechanical arm tail end initial point coordinate for (0.2,0.4), particle group optimizing part arranges correction parameter (k ₁, k ₂) scope is [± 0.5, ± 1.5], particle dimension is 2, and particle position initialisation range is [-2,2].In order to meet deicing inspection robot trajectory planning real-time demand, arranging population iterations is 5, and population is 30, acceleration weight c ₁=c ₂=1.5.Design mechanical arm is to eight direction actions, each orientation assigned address is respectively: top---and (0.2,0.5), front upper place---(0.3,0.5)---(0.3,0.4), front lower place---(0.3,0.3), below---(0.2, front, 0.3), the back lower place---(0.18,0.38), rear---(0.18,0.4) and back upper place---(0.18,0.5).Arrange not containing barrier in mechanical arm motion space, under namely deicing inspection robot works in safe mode.

As shown in Figure 8, give deicing inspection robot mechanical arm in accessible situation to around the result of eight orientation reachability tests, marked the starting point in each task and impact point.Can just find out from figure, the arrival impact point that mechanical arm tail end can be level and smooth in each task, meets the reachability requirements of the accessible track following of deicing inspection robot mechanical arm.

3. arrange deicing inspection robot mechanical arm to follow the tracks of continuous signal, particle group optimizing part arranges the same, and design tracking signal is with (0.31,0.4) be the center of circle, initial point is (0.21,0.4), radius is the circumference of 0.1, is expressed as:

$\left\{\begin{array}{cc}{x}_{signal}\left(t\right)=-0.1\cos t+0.31& t\&GreaterEqual;0\\ y_{signal}\left(t\right)=0.1\sin t+0.4& t\&GreaterEqual;0\end{array}\right.---\left(38\right)$

As shown in Figure 9, give deicing inspection robot mechanical arm tracking situation to continuous signal in accessible situation, in two curves in Fig. 9,---represent continuous signal variation track,------represents the pursuit path of deicing inspection robot mechanical arm tail end, as can be seen from the figure, continuous signal is without mechanical arm tail end initial point, mechanical arm can follow the tracks of continuous signal under the effect of population fuzzy programming device, and error gradually zero, keep stable.Therefore illustrate that this population fuzzy programming strategy can ensure the good tracking performance of mechanical arm.

4. the obstacle performance of this planning strategy is tested, deicing inspection robot mechanical arm tail end starting point is set for (0.2,0.4), be arranged on (0.35,0.4) there is obstacle at place, and the expanded radius of circle of this obstacle is 0.3, the detection range d=0.3 of mechanical arm tail end.Front arm tip terminal is (0.5,0.4).Particle group optimizing part arranges the same

As shown in Figure 10, give two-dimensional space information and the deicing inspection robot mechanical arm obstacle detouring geometric locus of barrier on the original state of deicing inspection robot mechanical arm and final state, ultra-high-tension power transmission line and circuit,------represents mechanical arm tail end variation track.As can be seen from the figure, when providing target location and obstacle location information, deicing inspection robot mechanical arm can make decisions on one's own, plan obstacle detouring action, has higher independence and intelligent.

Claims (5)

1. the active obstacle planing method of deicing high-voltage power transmission line inspection robot, is characterized in that, comprise the steps:

Step 1: utilize the laser radar detection ambient condition information be arranged on mechanical arm tail end, obtains robot and runs preceding object signal B;

Step 2: deicing robot observes by the camera carried the position running front cable, 30cm place, robot forearm front cable is set to the desired locations (x of mechanical arm tail end current time _g, y _g); According to the difference (e of mechanical arm tail end current location and desired locations _x, e _y) and current obstacle signal B, utilize fuzzy programming device to cook up the fuzzy programming angle in each joint of mechanical arm before traffic direction;

Step 3: can principle be reached according to shortest path and target, utilize particle cluster algorithm to carry out on-line optimization to fuzzy programming angle, obtain the population fuzzy programming expected angle in each joint of mechanical arm before traffic direction;

Step 4: using the input of each joint population fuzzy programming angle as neural network adaptive controller, neural network adaptive controller exports each joint control moment τ, instructs mechanical arm action.

2. the active obstacle planing method of deicing high-voltage power transmission line inspection robot according to claim 1, it is characterized in that, obtain robot in described step 1 to run preceding object signal B and refer to that the obstacle detouring according to setting judges critical distance A, judge that the obstacle detouring whether distance of the Obstacle Position that mechanical arm tail end current location and Airborne Lidar measure exceedes setting judges critical distance A:

e _xb=|x-x _b| (1)

e _yb=|y-y _b|(2)

Wherein, x _b, y _bfor the two-dimensional coordinate position of barrier, x, y are mechanical arm tail end current location, work as e _xb≤ A and e _ybduring≤A, judge that deicing inspection robot runs front and has chance barrier dangerous, export obstacle signal B=1, otherwise B=0.

3. the active obstacle planing method of deicing high-voltage power transmission line inspection robot according to claim 2, is characterized in that, according to the difference (e of mechanical arm tail end current location and desired locations in described step 2 _x, e _y) and present case under obstacle signal B, utilize fuzzy programming device to cook up mechanical arm each joint fuzzy programming angle when deicing robot travels forward, detailed process is as follows:

1) determine fuzzy programming device input information, and Fuzzy processing carried out to input data:

By mechanical arm tail end control information (e _x, e _y) and obstacle signal B input as fuzzy programming device:

(x _g, y _g) be set to the desired locations of mechanical arm tail end current time for 30cm place, robot forearm front cable;

Fuzzy processing is carried out, mechanical arm tail end error of coordinate (e to fuzzy programming device input information _x, e _y) corresponding 3 Fuzzy Linguistic Variable { N, Z, P}={ " bear ", " zero ", " just " }, Fuzzy Linguistic Variable N, Z and P respectively with membership function μ _n, μ _zand μ _pone_to_one corresponding, the codomain scope of membership function is [0,1]; The input variable of membership function is mechanical arm tail end error of coordinate (e _x, e _y):

Choosing membership function value is the non-zero Fuzzy Linguistic Variable Fuzzy Linguistic Variable as input information;

2) output Fuzzy Linguistic Variable is obtained according to fuzzy rule;

According to Fuzzy Linguistic Variable and the obstacle signal B of input information, inquiry fuzzy rule design table, as shown in table 1, exported Fuzzy Linguistic Variable { None accordingly, N, P}, wherein, None represents that mechanical arm is failure to actuate, and N represents that mechanical arm operating angle is negative, and P represents that mechanical arm operating angle is just;

Table 1 fuzzy rule designs

3) fuzzy reasoning is carried out to output Fuzzy Linguistic Variable, obtain and export membership function value corresponding to Fuzzy Linguistic Variable; Adopt Mamdani reasoning, the membership function exporting Fuzzy Linguistic Variable corresponding is as follows:

With e _xand e _ycorresponding membership function value under each self-corresponding Fuzzy Linguistic Variable, respectively to e _xand e _ythe union of corresponding all membership function values is tried to achieve the first common factor and second and is occured simultaneously; First common factor, the second common factor and B are asked the 3rd common factor, obtains an exact value; Occur simultaneously respectively to the μ of two outputs with the 3rd _n(θ _f1) and μ _p(θ _f1) cut, obtain the membership function value exporting Fuzzy Linguistic Variable;

4) adopt sharpening interface to obtain and export angle;

First the maximum membership degree functional value got in each Output rusults in two fuzzy programming angles compares, the Output rusults at the maximum place of membership function value is selected to export as final result, if in more each Output rusults process, occur that the maximum membership degree function of multiple result is maximum side by side and equal, then select to get less the exporting as final result of fuzzy language amount in Output rusults; Then the membership function value in this output fuzzy language result is substituted into formula 9-12 to invert function, obtain two joint actual output angle; Fuzzy language amount presses N, None, P ascending order.

4. the active obstacle planing method of deicing high-voltage power transmission line inspection robot according to claim 3, is characterized in that, it is as follows that the particle cluster algorithm in described step 3 carries out on-line optimization process to fuzzy programming angle:

1): optimization object is selected: arrange self-correcting parameter k=(k ₁, k ₂) k ₁, k ₂∈ [0,2], the fuzzy programming after optimization is expected, and angle is for (k ₁θ _f1, k ₂θ _f2);

2) update mode is determined: adopt elementary particle group to optimize update mode, be expressed as:

vi _j(n+1)=wvi _j(n)+r ₁(n)c ₁[pbesti _j(n)-xi _j(n)]+r ₂(n)c ₂[g ^besti _j(n)-xi _j(n)] (13)

xi _j(n+1)=xi _j(n)+vi _j(n+1) (14)

Wherein: i is for initializing correction parameter group number, and j is the dimension often organizing correction parameter, v _ijn () is the diverse vector of correction parameter group i jth dimension in n-th iteration, x _ijn () is the numerical value of correction parameter group i jth dimension in n-th time, r ₁, r ₂for independent random function, c ₁, c ₂for acceleration weight, c ₁, c ₂equal value is 1.2, w is inertia weight;

The mode that inertia weight w linearly successively decreases upgrades, and is shown below:

Wherein: n is current iteration number of times, n _maxfor total iterations, w _maxbe 0.9, w _minbe 0.4;

3) fitness function design:

Min:f(t)=s ₁f ₁(t)+s ₂f ₂(t) （16）

Wherein, s ₁be 0.4, s ₂be 0.6;

In any t, integration is minimum in time for the distance of mechanical arm tail end and initial point position, fitness function f ₁(t) be:

Impact point attracts then to be presented as that between mechanical arm tail end and impact point, air line distance is minimum, fitness function f ₂(t) be;

Wherein: x _o, y _ofor mechanical arm tail end initial point position, x (t), _yt () is t mechanical arm tail end position, x _gand y _grepresent aiming spot;

4) population on-line optimization:

Step1: population initializes, produces initial value and the diverse vector of correction parameter group, by (the k of current time t ₁θ _f1, k ₂θ _f2) substitute into fitness function formula 16 and calculate;

Step2: at iterations n _maxin, according to formula 13, formula 14, formula 15 upgrades correction parameter group numerical value, and each iteration all selects local optimum parameter group pbest and global optimum parameter group gbest;

Step3: after iteration completes, exports the correction parameter (k as moment t using global optimum parameter group gbest ₁, k ₂), excitation set mechanical arm is by (k ₁θ _f1(t), k ₂θ _f2(t)) action;

Step4: mechanical arm tail end sensor obtains subsequent time t+1 terminal position, turns Step1.

5. the active obstacle planing method of deicing high-voltage power transmission line inspection robot according to claim 4, is characterized in that, the neural network adaptive controller in described step 4 comprises ratio control-K _ps and neural network control

Described neural network adaptive controller is shown below

Wherein: τ is control moment, K _pbe symmetric positive definite matrix with Γ, K _pfor ratio controling parameters is set to be set to the unit matrix of 8 × 8, g (χ)=[g ₁(χ), g ₂(χ) ..., g _n(χ)] ^t, g _j(χ) (1≤j≤r) is RBF vector, for hidden layer is to the connection weight value matrix of output layer, initial value is set to 8 × 8 matrixes in [-1,1] interior random distribution;

S is synovial membrane face e=k θ _f-θ, wherein k θ _ffor population fuzzy programming angle, k θ _f=[k ₁θ _f1, k ₂θ _f2], θ is mechanical arm current angular, θ=[θ ₁, θ ₂], Λ is symmetric positive definite matrix

Adaptive neural network is shown below:

Wherein: for neutral net input, be the first joint planning angular speed, be the first joint planning angular acceleration, for second joint planning angular speed, for second joint planning angular acceleration, θ ₁be the first joint action angle, be the first joint action angular speed, θ ₂for second joint operating angle, for second joint operating angle speed, for hidden layer is to the connection weight value matrix of output layer, initial value is set to 8 dimension unit matrixs, and the attitude different according to mechanical arm and each joint angles desired value real-time update, to adapt to different demands for control; G (χ)=[g ₁(χ), g ₂(χ) ..., g _n(χ)] ^tfor RBF vector, g _j(χ) (1≤j≤r) has following form:

Wherein: r is the number of hidden layer neuron; μ _jcentered by vector, σ _jfor width vector;

The output of adaptive neural network target of approaching be compensation term, compensation term is shown below:

in adaptive neural network, hidden layer number r is 20, center vector μ _jbe set to the random function being uniformly distributed [-10,10], width vector σ _jall be set to 2.