Backpropagation neural network model with statistical inference in manufacturing processes

Abstract

Nowadays, there is a growing need for tools to model the complex characteristics of manufacturing processes to support decision-making and optimize production and quality. This study proposes using a Backpropagation neural network (BPNN) to model manufacturing processes, leveraging its ability to capture complex and nonlinear relationships. Additionally, integrating statistical inference techniques from Generalized Linear Models (GLM) with the neural network is suggested. This integration combines the predictive capabilities of the BPNN with the statistical tools of GLMs, enhancing result interpretability and analysis accuracy. The proposed approach was applied to two manufacturing processes. In the die-casting process, the BPNN with a logit function showed a lower deviance (0.0399) compared to the probit model (0.0875) and a greater deviance difference (6.1280) with a p-value of 0.0201. Confidence intervals confirmed the significance of these results. Metal temperature and solidification time were significant predictors, with weights of -1.0375 and -0.9880, respectively. In the machining process, the BPNN model with the probit function had a lower deviance (0.0140) compared to the logit model (0.0175) and a slight precision advantage with a deviance difference of 3.9325 and a p-value of 0.0473. Parameters S1 and S2 had significant effects with weights of 72.671 and -54.397, respectively. This approach allows for selecting optimal activation functions for each process, improving efficiency and quality control in manufacturing.