The Failures Forecast Of Muffler's Pre-mixed Cavity Based On Artificial Neural Networks

Stamping is an important method of sheet metal forming, especially in the automobile manufacturing. It's the technology of simulation and analysis of sheet metal that holds the balance not only to a higher vehicle design level, manufacturing quality and efficiency, but also to the improvement of market competitiveness of automobiles in our country. If the traditional methods were simply used to analyze the formability, the simulation results of the software are very precise. However, in order to avoid the various failures during the forming process, the workers have to modify some forming parameters or die shape, which makes more cost for the process, longer period for product development. Consequently, it's very practical to research how to apply the artificial intelligence technology to the formality analysis of the stamping part and the parameter optimization.In this paper, the new type of the pre-mixed cavity of vehicle muffler has been researched, including the characteristics, the stamping process analysis, the material properties analysis, etc. The product is a kind of three way conversion device. The shape looks like a muffler and it has the ability of noise reduction as the muffler. The shell is made of stainless steel which is resistant to high temperature, and the catalyst is coated on the honeycomb channels. When the automobile exhaust gas passed the channels, the carbon monoxide and hydrocarbon will be converted into harmless water and carbon dioxide, and the Oxides of Nitrogen will converted into oxygen and nitrogen. For the material, although the austenitic stainless steel SUS304 have the properties of high corrosion resistant ability and fine weld ability, it is hard to draw because fracturing and wrinkling may appear during the drawing process. In order to forecast the failures under different process parameters'impact before the real stamp, the method of combining artificial neural networks (ANNs) with numerical simulation has been used in this paper.The software of DYNAFORM has been used to simulate and analyze the part's drawing process which is demonstrated in this paper. Moreover, the influence main factors affecting the formality are summed up, and the impacts of the process parameters are researched emphatically. How to decide the numeric values of the parameters and the failures'type andcriteria are demonstrated, too. In the meantime, the samples'values for net training and testing are collected. By means of designing the table of orthogonal experiment, the test frequency can be reduced, and it makes the number and distribution of the samples more reasonable.On the base of researching the basic theory of ANN, the model of forming failures forecast for muffler's pre-mixed cavity is established by using the ANN toolbox of MATLAB. The steps are given as followed: selecting the network model (BP network and RBF network); deciding the input and output; designing the latent (not for the RBF network); training and testing the network by the samples collected during the numerical simulation period. There are 9 inputs for the network, which are blank holding force, friction coefficient, relative radius of the die and six draw bead forces. There are 2 outputs, which are wrinkling factor P and fracturing factor Q. According to the forming failures judgment criteria previously mentioned, when P is greater than 0.05, wrinkling occurs; whereas when Q is less than 0, fracturing occurs. Use the 37 vectors of samples to train the network. After that, by observing the output, you can get the estimation results. Use the other five vectors that didn't train the network to test the trained network's ability and show the reliability of it.In this paper, the algorithm of BP network has been researched. As far as BP neural network (Back-Propagation Network) is concerned, the gradient descent algorithm (traingd) is traditional as well as the gradient descent with momentum algorithm is improved. Research shows that the gradient descent algorithm has the weaknesses to encounter local minimum, slow convergence rate and convergence instability, etc, so it's necessary to use other optimized algorithms to train the network. In this paper, six different algorithms has been introduced and applied. It turns out that not all the algorithms are doable for this subject. And on the other hand, some algorithms, such as L-M algorithm, are very satisfying because of the advantages including fast convergence rate and high forecast precision.As far as RBF neural network (Radial Basis Function Network) is concerned, there are many advantages such as simple structure, easy training process, fast convergence rate and the ability to approach any unlinear functions, etc. The model of forming failures forecast of pre-mixed cavity based on RBF network is established in this paper. By using the same samples as the BP network, the training and testing results are satisfying. The results show us that any vectors of process parameters in the range ofthe samples, ANN models could replace the numerical software to forecast the failures because of the characteristics of interpolation, but not the characteristics of extrapolation.The main functions of the blank holder force (BHF) include material flow control, wrinkling suppression, especially for the flange region. So, it's one of the most important process parameters which need to be decided during the stamping process. This paper combines finite element numerical simulation with the RBF neural network to establish the BHF optimizing model. According to the analysis, the input and output numbers are reduced, and therefore the net structure is simplified. The inputs include wrinkling factor P, fracturing factor Q and any group of the initial BHF curve F0. The outputs are optimized BHF values. The samples for net training and testing are collected by using the DYNAFORM software. As for the input values, they are not obtained by using the different process parameter sets to simulate for many times, but because during the whole forming period, the strain calculation of every phases are independent from each other, therefore we can take the values of every phases as the sample vectors. As for the output values, simulate by the relatively reasonable set of parameters based on the former works, which are the optimized BHF values.There are 100 phases in the simulation period, and choose the 18 vectors after the very time as the training samples. Afterward, use three vectors to test the ability of the network, the results are also satisfying, which show us that the feasibility of the BHF optimization by combining ANN technology with numerical simulation technology, and it can be used for the practical production.