A neural network to evaluate dynamic systems

This paper concentrates on the structure of a system to be used in the study of products that change some of their properties during the use they are submitted to. The system has been applied to inspect products that have changed some of their main characteristics when used in normal conditions. The color of textiles (clothes, for instance) in an example of this kind of situation. The approach is based on pattern recognition techniques and learning methods are used in this application. The system first evaluates the original characteristics of the properties and then studies a set of images of the material during its use, after some days or months. A classification system is developed to define if the changes during this period are acceptable or not. The evaluation is made according to the classification that the system has learnt. A neural network is used in this system. Starting with the analysis made of the material in the moment it leaves the factory (patterns), we follow the pieces during the use, comparing them with the patterns. The neural network decides if the changes are acceptable or not. The system has an important use to define what actions have to be taken in the production process in terms of quality assurance of the products. The hardware and the software were developed for application in a textile industry in Brazil. Five thousand pieces were tested. Only 1.15% was incorrectly classified (which means that we have acceptable changes that the system has considered as unacceptable), indicating the success of the approach developed