| NEURAL NETWORKS IN ADVANCED MANUFACTURING TECHNOLOGY ABSTRACT In this paper we are proposing a method of intelligent controlling of injection molding using neural network based control unit. Neural networks have been recently expanding into many areas of engineering science and commerce is an appropriate tool for this task. The characteristics of neural networks that make them promising for manufacturing applications are – After being trained by observed data, neural networks offer pattern recognition based intelligent control. Neural network process data in a parallel manner and thus they are very fast. Since neural networks learn solely based on training data it is possible to obtain inverse models of manufacturing process by simply obtaining the training data from a process model and presenting it to the neural networks in an inverse manner. This simulation of neural network control scheme may be done by using soft wares like MATLAB, TG MOLD etc, We are suggesting that it is better to use “FORWARD MODELLING”of neural network rather than the “INVERSE MODELLING” |
DIAGRAMATIC REPRESENTATION OF NEURAL NETWORK
INTELLIGENT MANUFACTURING CONCEPT APPLIED TO INJECTION MOLDING PROCESS
Injection molding makes a wide variety of plastic parts. When used with thermoplastics the process is basically consists of plastic compound, which isin powdered or granular form and then injecting the resultant melt into a mold either by a hydraulic plunger or by rotating screw system of an extruder. Injection molding of thermosets differs in the way, the melt is hardened in the mould.
APPLICATION OF NEURAL NETWORK BASED CONTROL TO INJECTION MOLDING PROCESS
The aim is to identify the feasibility of obtaining an intelligent control scheme that can learn the inverse model of a manufacturing process and react in a real-time to take corrective actions during fabrication.
The primary reasons for this are-
Control of flow progression are, during the molding process required using the process model information in an inverse manner since there is no way of relating a desired flow progressionscheme to an inlet flow rate profile that will result in it and Injection molding process with two inlet gates are proposed during the present thesis.The steps that should be followed in obtaining the training and test data for the neural network proposed are:
Control of flow progression during the molding process requires fast acting control unit since mold filled times are very short.
Start with an empty mold and run the simulation code until the mold was filled, randomly changing the inlet flow rates every ten seconds.
From the simulation, every two consecutive flow fronts plus the inlet flowrates that combine them constituted atraining (a test) pattern for the neural network.
FORWARD MOLDING WITH A BACK PROPAGATING NEURAL NETWORK
In the forward modeling case the neural network is given apresent flow front and an inlet flow rate combination and is expected to predict the shape and location of the next flow front
The control method can be summerised as follows:
At any given time, sense the present flow front.
Using a ‘C’ programme compare this present flow front with flow fronts along the desired flow progression part identified earlier and find the desired flow front that comes closest to and after the present flow front. This can be called the desired flow front of interest for the given time step.
Run the neural network many times, each time with the input set to the neural network being the present flow front plus, a different flow rate combination.
Compare the resulting predicted flow fronts obtained as neural networkoutputs with the desired flow fronts of interest and find the predicted flow rate combination that gave this predicted flow front is the best choice for this time step.
The four-step procedure is repeated until the mold cavity is filled.
CONTROL SYSTEM BY USING NEURAL NETWORKS
CONTROL OF FLOW PROGRESSION DURING INJECTION MOLDING PROCESS
The resulting flow front can then be fed into c programme, which can do the following operations:
It should find the desired flow front of interest.
- It should call the neural network 256 times (as specified earlier) compare each time with a different inlet flow rate combination plus the flow front obtained from the simulation software as the present flow front.
- It should compare the each of the resulting 256 possible next flow front with the desired flow front of interest.
- It should identify the possible next flow front that is closest to the desired flow front of interest.
CONCLUSIONS AND FUTURE DIRECTION
On going efforts are directed towards demonstrating the capabilities of this process model/neural Network based control unit during the injection molding and resin transfer molding experiments that integrate the control method with a reliable flow progression sensing sub system . Once this is accomplished we would like to apply this novel process devolopment method to other manufacturing processes for which process models and in-situ schemes exists.
