CN105643157B - Automatic girder welding obstacle predicting method for optimizing GRNN based on correction type fruit fly algorithm - Google Patents

Abstract

The invention discloses an automatic girder welding obstacle predicting method for optimizing a GRNN based on a correction type fruit fly algorithm. According to the method, an ultrasonic sensor is adopted to collect obstacle information, two factors of pheromones and sensitivity are introduced into a traditional fruit fly algorithm, an optimizing strategy and the substitute mode of the fruit fly position are improved, the global optimization characteristics of fruit fly optimization are corrected, and the GRNN is subject to parameter optimization; the optimized GRNN is trained by the collected obstacle information, and an optimal automatic girder welding obstacle prediction model is built; and then according to a prediction result given by the prediction model, automatic gun retreating and returning of a girder production line welding gun is achieved. The correction type fruit fly algorithm is adopted to optimize the GRNN so as to perform obstacle prediction on automatic girder welding, the prediction speed and precision of obstacles are improved, meanwhile, the production efficiency of automatic girder welding is greatly improved, the production cost is saved for an enterprise, and giant economic benefits are brought.

Description

Crossbeam automatic welding obstacle Forecasting Methodology based on amendment type fruit bat algorithm optimization GRNN

Technical field

The invention belongs to field of automatic welding, and in particular to a kind of crossbeam based on amendment type fruit bat algorithm optimization GRNN Automatic welding obstacle Forecasting Methodology.

Background technology

Crossbeam is widely used in the industries such as bridge, container, hoisting machinery as a kind of welding structural element. and it is common for I-shaped The girder constructions such as beam, H type beams, box, generally there are the barriers such as gusset, dividing plate and cavity between crossbeam and web. due to big The species of beam is various, and the clamping precision of workpiece is not high, and the position of the barrier of workpiece has larger randomness, it is difficult to by list One method carries out forecasting-obstacle.At present, the welding of crossbeam still adopts semi-manual automanual welding method, works as welding process In when running into barrier, can only the switching signal of manually button exit welding gun, manually opened after clearing the jumps again OFF signal makes welding gun return to welding position, restarts welding, this severely limits the production efficiency of crossbeam welding, constrains enterprise The development of industry, therefore the predictive study of barrier is carried out, that realizes welding gun moves back rifle and time rifle automatically, to realizing that crossbeam automatization welds Connect significant.

The content of the invention

It is contemplated that the speed and degree of accuracy in improving crossbeam automatic welding to forecasting-obstacle, so as to improve crossbeam certainly The production efficiency of dynamic weldering, is that enterprise saves production cost, increases economic efficiency.

The present invention is employed the following technical solutions：A kind of crossbeam automatic welding obstacle based on amendment type fruit bat algorithm optimization GRNN Forecasting Methodology, comprises the following steps that：

Step 1：The information of collection container crossbeam barrier (gusset, gutter channel and auxiliary upper flange), including obstacle The size of thing, three indexs of supersonic sensing measured value and workpiece height：Wherein barrier size is different because of product.

Step 2：It is that unified Analysis are carried out to different types of sample, obtains preferable prediction effect, must be to sample data Make following normalized：

Wherein, X^kRepresent initial data X_kInput sample Jing after normalization, X_minAnd X_maxIn representing initial data respectively Minima and maximum.

Step 3：Smoothing factor σ in optimal GRNN is searched using amendment type fruit bat algorithm, so as to obtain optimum GRNN.

Step 4：It is using the data after normalization as training sample, for training optimum GRNN, optimal big so as to obtain The forecast model of beam automatic welding barrier.

In step 3, described amendment type fruit bat algorithm is that pheromone and sensitivity two are introduced in traditional fruit bat algorithm The individual factor.

Sensitivity and pheromone in described amendment type fruit bat algorithm is defined as follows：

The optimal fruit bat bestSmell of flavor concentration is found out first, calculates i-th individual food information element P of fruit bat (i)；

Retain the position (Xo (i), Yo (i)) of fruit bat body position (X (i), Y (i)) and previous generation, spirit is calculated according to following formula Sensitivity judges the value (the same to Rx (i) of computing formula of Ry (i)) of factor R x (i)；

The factor is judged according to the sensitivity that pheromone is individual with the conformity relation of sensitivity and fruit bat, each fruit bat is calculated The corresponding sensitivity S (i) of body；

Wherein：S_min=P_min,S_max=P_max

The fruit bat that pheromone is matched with sensitivity is found out, that is, is met P (i)≤S (i), is determined the search starting point of next round：

Wherein：(X_best,Y_best) and (X_worst,Y_worst) it is respectively in the good fruit bat of olfactory function flavor concentration most preferably and most Poor coordinate.

The specific implementation step of described amendment type fruit bat algorithm is as follows：

Step one：Initialization

1) according to object function, initial value, population scale Sizepop, maximum iteration time Maxgen are searched in setting；

2) random initializtion fruit bat position (X_axis, Y_axis)；

3) fruit bat individuality random direction and distance (X, Y) are given；

4) random coordinates according to fruit bat position, calculate the distance (Dist) and flavor concentration decision content (SM) with origin；

5) flavor concentration (Smell) of fruit bat individuality position is calculated according to object function (Function)；

Smell (i)=Function (SM (i)) (9)

6) fruit bat initial information element and sensitivity are produced according to formula (1)-(3), and finds out pheromone and matched with sensitivity Fruit bat, retain the coordinate of wherein concentration highest and lowest individual；

Step 2：Iteration optimizing

7) starting point that fruit bat individuality next round is searched is determined according to formula (4) and formula (5)；

8) repeat 4) and 5) to calculate the individual flavor concentration value of fruit bat；Judge current best flavors concentration whether better than front One iteration best flavors concentration：If so, next step, otherwise execution step are entered 2) then；

9) fruit bat initial information element and sensitivity are produced according to formula (1)-(3), and finds out pheromone and matched with sensitivity Fruit bat, retain the coordinate of wherein concentration highest and lowest individual；

10) into iteration optimizing, repeat step 7) -9)；

Step 3：Terminate judging, until current iteration number of times is equal to maximum iteration time Maxgen, or reached target essence Degree requires or theoretially optimum value that optimizing terminates, and exports optimizing result.

The beneficial effect that the present invention reaches：1. amendment type fruit bat algorithm only needs to setting initially compared with traditional fruit bat algorithm Change scope, it is not necessary to which direction and the distance of the search of every generation fruit bat are set, parameter setting is simpler；2. amendment type fruit bat Optimized algorithm is easier to jump out local extremum and find globally optimal solution；3. the crossbeam based on amendment type fruit bat algorithm optimization GRNN The prediction effect of automatic welding obstacle forecast model is stable, and precision of prediction is high.

Description of the drawings

Fig. 1 is the crossbeam automatic welding obstacle Forecasting Methodology flow chart based on amendment type fruit bat algorithm optimization GRNN；

Flight paths of the Fig. 2 for FOA-GRNN model fruit bats

Flight paths of the Fig. 3 for AFOA-GRNN model fruit bats

Fig. 4 is searching process of the AFOA algorithms to SPREAD values

Fig. 5 is the root-mean-square error of two kinds of model predictions

Fig. 6 predicts the outcome for AOFA-GRNN models

Specific embodiment

The present invention is described in further detail with reference to the accompanying drawings and examples.

Embodiment 1, referring to Fig. 1, the present invention is employed the following technical solutions：It is a kind of to be based on amendment type fruit bat algorithm optimization GRNN Crossbeam automatic welding obstacle Forecasting Methodology, comprise the following steps that：

Step 1：The information of collection container crossbeam barrier (gusset, gutter channel and auxiliary upper flange), including obstacle The size of thing, three indexs of supersonic sensing measured value and workpiece height：Wherein barrier size is different because of product.

Step 2：It is that unified Analysis are carried out to different types of sample, obtains preferable prediction effect, must be to sample data Make following normalized：

Wherein, X^kRepresent initial data X_kInput sample Jing after normalization, X_minAnd X_maxIn representing initial data respectively Minima and maximum.

Step 3：Smoothing factor σ in optimal GRNN is searched using amendment type fruit bat algorithm, so as to obtain optimum GRNN.

Step 4：It is using the data after normalization as training sample, for training optimum GRNN, optimal big so as to obtain The forecast model of beam automatic welding barrier.

In step 3, described amendment type fruit bat algorithm is that pheromone and sensitivity two are introduced in traditional fruit bat algorithm The individual factor.

Sensitivity and pheromone in amendment type fruit bat algorithm is defined as follows：

The optimal fruit bat bestSmell of flavor concentration is found out first, calculates i-th individual food information element P of fruit bat (i)；

Retain the position (Xo (i), Yo (i)) of fruit bat body position (X (i), Y (i)) and previous generation, spirit is calculated according to following formula Sensitivity judges the value (the same to Rx (i) of computing formula of Ry (i)) of factor R x (i)；

The factor is judged according to the sensitivity that pheromone is individual with the conformity relation of sensitivity and fruit bat, each fruit bat is calculated The corresponding sensitivity S (i) of body；

Wherein：S_min=P_min,S_max=P_max

The fruit bat that pheromone is matched with sensitivity is found out, that is, is met P (i)≤S (i), is determined the search starting point of next round：

Wherein：(X_best,Y_best) and (X_worst,Y_worst) it is respectively in the good fruit bat of olfactory function flavor concentration most preferably and most Poor coordinate.

The specific implementation step of described amendment type fruit bat algorithm is as follows：

Step one：Initialization

11) according to object function, initial value, population scale Sizepop, maximum iteration time Maxgen are searched in setting；

12) random initializtion fruit bat position (X_axis, Y_axis)；

13) fruit bat individuality random direction and distance (X, Y) are given；

14) random coordinates according to fruit bat position, calculate the distance (Dist) and flavor concentration decision content (SM) with origin；

15) flavor concentration (Smell) of fruit bat individuality position is calculated according to object function (Function)；

Smell (i)=Function (SM (i)) (9)

16) fruit bat initial information element and sensitivity are produced according to formula (1)-(3), and finds out pheromone and matched with sensitivity Fruit bat, retain the coordinate of wherein concentration highest and lowest individual；

Step 2：Iteration optimizing

17) starting point that fruit bat individuality next round is searched is determined according to formula (4) and formula (5)；

18) repeat 4) and 5) to calculate the individual flavor concentration value of fruit bat；Judge whether current best flavors concentration is better than Preceding iteration best flavors concentration：If so, next step, otherwise execution step are entered 2) then；

19) fruit bat initial information element and sensitivity are produced according to formula (1)-(3), and finds out pheromone and matched with sensitivity Fruit bat, retain the coordinate of wherein concentration highest and lowest individual；

20) into iteration optimizing, repeat step 7) -9)；

Step 3：Terminate judging, until current iteration number of times is equal to maximum iteration time Maxgen, or reached target essence Degree requires or theoretially optimum value that optimizing terminates, and exports optimizing result.

Embodiment 2, by taking certain container crossbeam forecasting-obstacle as an example, the barrier of prediction has a gusset, gutter channel and auxiliary Helping upper flange three types. prediction index includes the size of barrier, three fingers of supersonic sensing measured value and workpiece height Mark：Wherein barrier size is different because of product, and data herein include the barrier size of 6 kinds of products；Obstacle information leads to Ultrasonic sensor collection is crossed, ultrasound wave is with position changing machine clamp centerline same level and parallel with workpiece web；Workpiece is high The difference in height of the centrage for workpiece and position changing machine clamp centrage is spent, the data that ultrasonic sensor is gathered first are can be considered. Experiment, part sample data such as table 1 are predicted to 200 groups of truthful datas of certain container company container crossbeam product collection It is shown, wherein specifying obstacle identity:1 is gusset, and 2 is gutter channel, and 3 are auxiliary upper flange.

The initial data of 1 part sample of table

In order to verify the performance of proposed amendment type fruit bat algorithm optimization GRNN forecasting-obstacle models, using MATLAB Newgrnn functions in neutral net GRNN workbox carry out forecasting-obstacle, and pass through amendment type fruit bat optimized algorithm AFOA Optimal spreading parameter SPREAD is searched, while contrast experiment is carried out with FOA-GRNN models. two kinds of models are joined using identical Number is arranged：If fruit bat population be 10, maximum iteration time is 200, fruit bat initial position be [0,1], fruit bat random flight in FOA Range direction is [- 1,1] with distance range, and optimizing end condition is GRNN neural network forecast root-mean-square errors RMSE<1e-12. Jing Used as training sample, 20 groups used as forecast sample afterwards for front 180 groups of data of normalized. the fruit bat searching process of two kinds of models As shown in Figures 2 and 3.

Contrast Fig. 2 and Fig. 3 can be seen that AFOA-GRNN models compare FOA-GRNN models, and the individual point searched of fruit bat is more Few, search path is more random, and search space is bigger. therefore, improved fruit bat optimized algorithm is easier to jump out local extremum and look for To globally optimal solution.

Can be seen that from Fig. 4 and Fig. 5：Root-mean-square error RMSE=7.4786e-13 of FOA-GRNN model predictions, repeatedly Generation number is 137, takes 265.736789S；And root-mean-square error RMSE=1.7549e-16 of AFOA-GRNN model predictions, Iterationses are 10, take 17.731818S. and repeat to test, as a result stable. therefore, amendment type AFOA-GRNN model relative to Unmodified FOA-GRNN models, predetermined speed is faster and precision of prediction is higher.

The optimal spreading factor SPREAD values that amendment type fruit bat optimized algorithm AFOA finds are updated in GRNN models, are entered Row is left the forecasting-obstacle of 20 groups of samples, as a result as shown in Figure 6.

From in Fig. 6, AFOA-GRNN models are up to many conversion of 100%. Jing to the forecasting accuracy of obstacle identity Experiment, prediction effect is stable, compares that FOA-GRNN model errors rates are low, and precision of prediction is higher.

Claims (3)

1. the crossbeam automatic welding obstacle Forecasting Methodology based on amendment type fruit bat algorithm optimization GRNN, it is characterised in that concrete steps are such as Under：

Step 1：The information of collection container crossbeam barrier, including the size of barrier, supersonic sensing measured value and work Part three indexs of height：Wherein barrier size is different because of product；

Step 2：It is that unified Analysis are carried out to different types of sample, obtains preferable prediction effect, sample data must be made such as Lower normalized：

$X^k=\frac{X_k-X_{\min }}{X_{\max }-X_{\min }},k=1,2,...,n$

Wherein, X^kRepresent initial data X_kInput sample Jing after normalization, X_minAnd x_maxRespectively represent initial data in minimum Value and maximum；

Step 3：Smoothing factor σ in optimal GRNN is searched using amendment type fruit bat algorithm, so as to obtain optimum GRNN；

Step 4：Using the data after normalization as training sample, for training optimum GRNN, so as to obtain optimal crossbeam certainly The forecast model of dynamic weldering barrier.

2. the crossbeam automatic welding obstacle Forecasting Methodology based on amendment type fruit bat algorithm optimization GRNN according to claim 1, It is characterized in that：In step 3, amendment type fruit bat algorithm be introduce in traditional fruit bat algorithm pheromone and sensitivity two because Son.

3. the crossbeam automatic welding obstacle Forecasting Methodology based on amendment type fruit bat algorithm optimization GRNN according to claim 2, It is characterized in that：Sensitivity and pheromone in amendment type fruit bat algorithm is defined as follows：

The optimal fruit bat bestSmell of flavor concentration is found out first, calculates i-th individual food information element P (i) of fruit bat；

$P\left(i\right)=\left\{\begin{array}{ccc}\frac{Smell\left(i\right)}{bestSmell},& Smell\left(i\right)\&GreaterEqual;0,& bestSmell\&NotEqual;0;\\ \frac{bestSmell}{Smell\left(i\right)},& Smell\left(i\right)<0;& \end{array}\right.$

Retain the position (Xo (i), Yo (i)) of fruit bat body position (X (i), Y (i)) and previous generation, according to following formula meter sensitivity Judge the value of factor R x (i), the same to Rx (i) of computing formula of Ry (i)；

$Rx\left(i\right)=\left\{\begin{array}{cc}\frac{X\left(i\right)-Xo\left(i\right)}{X_{best}-X_{worst}},& P\left(i\right)\&le;S\left(i\right);\\ \frac{X\left(i\right)+B}{A},& P\left(i\right)>S\left(i\right);\end{array}\right.$

$Ry\left(i\right)=\left\{\begin{array}{cc}\frac{Y\left(i\right)-Yo\left(i\right)}{Y_{best}-Y_{worst}},& P\left(i\right)\&le;S\left(i\right);\\ \frac{Y\left(i\right)+B}{A},& P\left(i\right)>S\left(i\right);\end{array}\right.$

The factor is judged according to the sensitivity that pheromone is individual with the conformity relation of sensitivity and fruit bat, each fruit bat individuality is calculated right The sensitivity S (i) answered；

$\left\{\begin{array}{c}S\left(i\right)=S_{min}+(S_{max}-S_{\min })*R\left(i\right)\\ R\left(i\right)=\frac{Rx\left(i\right)+Ry\left(i\right)}{2}\end{array}\right.$

Wherein：S_min=P_min, S_max=P_max

The fruit bat that pheromone is matched with sensitivity is found out, that is, is met P (i)≤S (i), is determined the search starting point of next round：

$X\left(i\right)=\left\{\begin{array}{c}{X}_o\left(i\right)+(X_{best}-X_{worst})*unifrnd(0,1,1),\\ P\left(i\right)\&le;S\left(i\right);\\ A*rand\left(\right)-B,P\left(i\right)>S\left(i\right);\end{array}\right.$

$Y\left(i\right)=\left\{\begin{array}{c}{Y}_o\left(i\right)+(Y_{best}-Y_{worst})*undifrnd(0,1,1),\\ P\left(i\right)\&le;S\left(i\right);\\ A*rand\left(\right)-B,P\left(i\right)>S\left(i\right);\end{array}\right.$

Wherein：(X_best,Y_best) and (X_worst,Y_worst) be respectively the good fruit bat of olfactory function in flavor concentration it is optimal and worst Coordinate.