Artificial intelligence application in the bending process.

Groze, Florica Mioara ; Achimas, Gheorghe ; Lazarescu, Lucian 等

Abstract: The accurate prediction of springback is essential for the design of the tools used in automotive sheet forming operations. This paper describes a novel method for predicting the springback in bending operations using artificial neural networks. The accuracy of the model is verified by comparisons with the results of finite element simulations (ABAQUS).

Key words: Springback, Artificial Neural Networks (ANN), Bending Process, Finite Element Simulation

1. INTRODUCTION

The term Artificial Intelligence (AI) was first used by John McCarthy as a denomination of "the science and engineering of making intelligent machines". Research in AI is concerned with producing machines that perform automated tasks requiring intelligent behavior. Examples include control, planning and scheduling, the ability to answer diagnostic and consumer questions, handwriting, natural language, speech, and facial recognition. As such, the study of AI has also become an engineering discipline, focused on providing solutions to real life problems. AI is studied in overlapping fields of computer science, psychology, philosophy, neuroscience, and engineering, dealing with intelligent behavior, learning, and adaptation and usually developed using customized machines or computers. AI is subdivided into 3 main fields: Neural networks, Fuzzy systems and Evolutionary computation. This research overlaps with cognitive science, cybernetics & robotics. Many hybrid intelligent systems have also appeared. In this paper we will use Artificial Neural networks in order to predict the springback angle in the V-bending process.

Hambli, and Guerin (Hambli & Guerin 2003) use artificial neural networks in order to predict the optimum clearance in sheet metal blanking processes. Achimas and Lazarescu (Achimas & Lazarescu 2004) also used neural networks for the springback prediction of the bent parts. Chan et al. (Chan et al, 2003) performed finite element analyses of springback in the V-bending metal forming process. Samek & Sykorova (Samek&Sykorova 2007) used a multilayer Feed-Forward Neural Network for the verification of the predictor of laser micro-machining input parameters. Ruffini & Cao (Ruffini & Cao 1998) used neural networks for springback minimization in a channel forming process.

2. SPRINGBACK PHENOMENON

A large amount of the metal production results from forming operations of sheet metals. In nearly all of the sheet forming operations, the metal is subjected to bending in at least one area. Examples of this include forming of aluminum rain gutters, aircraft skins, auto bodies, appliance shells, soda cans, fan blades, etc. Because of the extremely high volume of sheet metal forming operations, considerable attention has been focused on improving the quality of the sheet metal bending. Specifically, the issue of springback has been thoroughly studied and documented for the case of cold working of sheets. In the specific case of sheet bending, half of the cross section is strained in compression, while the other half is stretched.

Somewhere between the compression and tension zones it is a neutral axis which is completely unstrained. After releasing the bending load, the metal sheet 'springs back', thereby relieving much of the elastic fraction of the strain. The amount of this springback that occurs after bending is dependent upon the material's elastic modulus, yield stress, thickness, cross sectional shape, as well as the bending angle and the radius over which the material was bent.

The existence of springback can add complexity to many industrial bending operations, and it must be corrected. Correction methods include overbending. Overbending a metallic sheet compensates for springback by bending it farther than what is required for the finished product. Overbending requires knowledge of how many degrees one must overbend. The springback ratio is defined by the ratio of the initial bending radius divided by the final radius. This knowledge may be gained through some simple bending tests, through mathematical calculations or through the use of a new method, namely the artificial neural networks.

3. ARTIFICIAL NEURAL NETWORK

A neural network is a parallel processing architecture consisting of a large number of inter-connected processing elements called neurons organized in layers.

In order to solve a problem using neural network, it is necessary to run trough two main steps: a) training and b) generalization.

The system is trained by a large set of training samples based on a supervised learning procedure. They learn by adaptively updating the synaptic weights that characterize the strength of the connections. A network is considered successfully trained if it can closely approximate the taught values for the trained data space and can provide smooth interpolations for the untrained data space.

Generalization represents the answer phase of the neural network, when giving its the new input vectors or vectors from the training data, the output vector will have the expected value, with a very little possible error.

3.1 Predictions of the neural network

The objective of the work was to develop a new method for prediction of springback, using Artificial Neural Networks (ANNs). The springback parameter is predicted as a function of the bending parameters: thickness g, angle as, punch radius [r.sub.p], and die radius [r.sub.pl]. Initially, finite element simulations of V bending processes was conducted. A back propagation training neural network model was taught by using the numerical results. After this stage, the network has been employed for substituting the finite element code needed for the springback prediction. The ANN has an input layer, an output layer, and at least two hidden layers. The neurons of the input layer receive the input data to be processed; the output neurons give the processed data and the neurons of the hidden layer provide intermediate processing support. For the prediction of springback, the authors have developed a feedforward network using EasyNN-Plus Software (fig.1).

[FIGURE 1 OMITTED]

The network has an input layer, an output layer and two hidden layers. The input layer has a number of four neurons, the same as the number of the input parameters. The output layer has one neuron, corresponding to the output parameter (springback angle [beta]).

[FIGURE 2 OMITTED]

The training process of the ANN was stopped when the average training error became less than 0.01. This level was attained in 3069 learning cycles.

After the training, the network was tested in order to find out whether the calculated results are in agreement with the values obtained by finite element simulation. For this purpose, the so-called validation test was performed using input parameters not included in the training procedure of the neural network. The input data is presented in Table 1.

On the basis of the new input data, the neural network calculated the springback angle. The comparison between the predicted values provided by ANN and FEM is presented in figure 3.

[FIGURE 3 OMITTED]

4. CONCLUSIONS

In order to predict the springback angle, the authors used many artificial neural network architectures, but the best results were provided by the artificial neural network with two hidden layers which is presented in figure 1.

In this paper, a neural network with four inputs and one output was used for predicting the springback angle in the V bending process.

The results provided by ANN are in good agreement with the FEM results, which means that the neural network was well trained and it has a good capacity of generalization. Therefore, the network can be used for the springback prediction in practical applications.

The predictive capabilities of the ANN can be further improved. In the present paper, the authors considered that the input parameters has an equal influence on springback. A more refined approach would have to assign an influence weight to each of the input parameters. Another subject of future research will be the solution of the inverse problem, consisting in the use of the springback angle [beta] and thickness g as input data, in order to find out the angle [[alpha].sub.s], punch radius [r.sub.p], die radius [r.sub.pl] (as output parameters).

5. REFERENCES

Achimas, Gh. & Lazarescu, L. (2004) A new method for springback prediction of bent parts, Academic Journal of Manufacturing Engineering, Editura Politehnica, Timisoara, vol. 2 nr. 1/2004, pag. 53-57, ISSN 1583-7904

Chan, W.M.; Chew, H.I.; Lee, H.P. & Cheok, B.T. (2004). Finite element analysis of spring-back of V-bending sheet metal forming processes, Available from: http://www.sciencedirect.com/ Accesed:2006-06-14

Hambli, R. & Guerin, F. (2003): Application of a neural network for optimum clearance prediction in sheet metal blanking processes, Available from: http://www.sciencedirect.com/ Accesed:2005-03-08

Ruffini, R. & Cao, J. (1998) Using neural network for springback minimization in a channel forming process,, Available from: http://www.sciencedirect.com/ Accesed:2005-01-20

Samek, D. & Sykorova, L. (2007). Verification of the predictor of laser micro-machining input parameters, Microcad 2007 International Scientific Conference, Miskolc pp. 149-154, ISBN 978-963-661-753-0

*** ABAQUS User's Guide (release 6.4). Electronic documentation http://en.wikipedia.org/wiki/Portal:Artificial_intelligence

Table 1. Bending parameters as input data for the tests

thickness g [[alpha].sub.s] [r.sub.p] [r.sub.pl] [beta]
[mm] [[degrees]] [mm] [mm] [[degrees]]

0.6 90 6 8 0.216
1 120 4 8 0.173
1.6 60 6 4 1.132
0.6 60 5 6 0.243