Convolutional Neural Networks for Structured Industrial Data

Abstract

Regression methods aim to predict a numerical value of a target variable given some input variables by building a function $f:\mathbb {R}^n \rightarrow \mathbb {R}$. In Industry 4.0 regression tasks, tabular data-sets are especially frequent. Decision Trees, ensemble methods such as Gradient Boosting and Random Forest, or Support Vector Machines are widely used for regression tasks with tabular data. However, Deep Learning approaches are rarely used with this type of data, due to, among others, the lack of spatial correlation between features. Therefore, in this research, we propose two Deep Learning approaches for working with tabular data. Specifically, two Convolutional Neural Networks architectures are tested against different state of the art regression methods. We perform an hyper-parameter tuning of all the techniques and compare the model performance in different industrial tabular data-sets. Experimental results show that both Convolutional Neural Network approaches can outperform the commonly used methods for regression tasks.