On-line Quality Control of Injection Molding Using Neural Networks

Plastics is one of the world’s largest and fastest growing industries, ranked as one of the few multi-billion dollar industries. Due to the vast improvements in polymers and production processes, plastics are now amongst the world’s most widely used materials, with the world consumption of polymers surpassing steel in volume and weight. Injection molding represents approximately 32% (by weight) of this industry. Despite major advancements in injection molding manufacturing technology, there still exists significant obstacles in the manufacturing process, most notably the achievement and control of molded part quality. This leads to the production of poor quality or defective parts resulting in material wastage, increased production costs, and reduced customer satisfaction. Controlling part quality directly is the key to further advances in injection molding. This research resulted in the development of a pattern-based process control system for injection molding based on a three-layer, feedforward neural network model of the molding process. This system is based on the premise that measured process patterns act as ‘quality signatures’ and uses process patterns, corresponding to measured machine response, as inputs to determine the necessary control actions. The developed system was integrated with an injection molding process control system currently under development at Moldflow Pty. Ltd. and compared with a Statistical Process Control (SPC) based system. It was expected that the developed pattern-based controller would better model the complex interactions and nonlinearity of the molding process and result in improved quality control. The performance of the neural and SPC system was undertaken by analysing the part weight for the production of a small molded box during a number of tests to simulate typical production scenarios. These tests included set-point tracking and process disturbance tests. The resulting part weight for each test was recorded and analysed, with the best control system achieving the minimum part weight variation. In addition, to provide a measure of the control achieved by each system, each test was also performed using no control or fixed machine set-points. The test results demonstrated the benefits of the neural controller by achieving approximately 43% and 27% better part weight control than the reference system in two of the three tests performed. The neural controller also performed significantly better than SPC in all tests, achieving, on average, approximately 50% better part weight control. In addition, the neural controller was able to quickly restore production as a result of deliberate set-point adjustments. The benefits of the pattern-based approach was also demonstrated by the neural controller showing significant capability at modelling the complex process interactions without requiring assumptions or a priori knowledge about the process. 1 Moldflow  is a registered trademark of Moldflow Pty. Ltd.