ACCELERATED LEARNING NEURAL NETWORK FOR MATERIAL PROCESSING SENSOR DATA ANALYSIS

Abstract

In order to enhance the production of manufactured parts from raw materials, the recent trend has been to improve productivity and quality by increased automation which will require better process control. Improved process control requires that the process be monitored by sensors; however, there are several problems associated with the use of sensors in real-time manufacturing situations. The artificial neural system (ANS) offers a solution to these problems. ANS's have the ability to tolerate significant amounts of noise in the sensor data; in addition, ANS's are inherently a massively parallel processing architecture which allows very high processing speeds. Finally, ANS have the capability to be implemented as integrated circuit chips for miniaturized and low cost sensors. To be applied in true flexible manufacturing environments, ANS's must be able to learn examples of sensor data in very short periods of time. The authors have developed and tested an accelerated learning system for modified back-propagation multi-layer neural networks which has been applied to a variety of sensor data both two and three dimensional.