Neural Network-based Approximate Quality Prediction for Parameter Exploration in Industrial Manufacturing

Various control parameters are required for industrial plant operation and product manufacturing. However, the trial-and-error scheme of testing physical equipment for optimal parameter exploration is very costly and time consuming. Therefore, interest in an environment, that can predict quality output parameters in advance by modeling the target plant, has increased. Mathematical modeling is difficult due to the lack of interrelationships between the various parameters applied to the facility and the complex internal behavior. This paper proposes a technique to predict the quality output factor before manufacturing a product using a multilayer perceptron (MLP) neural network. Moreover, we handle the critical parameters of the produced product in duplicate at the input layer and enable the generation of product-dependent inference results. The experiments used input and output data sets from various products extracted from the wire manufacturing process. The proposed scheme generated a quality output of the product not included in the neural network training as average and distribution similar with label. The experimental results showed that the difference from the label average was reduced by up to 50.26%, similar to the label distribution in the various product cases used for training,