Image as a Way of Processing Multidimensional Production Data for Product Quality Prediction Using Deep Learning

The article concerns the processing of multidimensional data sets that are obtained from production processes and are used as training data for predictive models. Preparing such training data often requires performing complex feature engineering operations to remove highly correlated features or select only the most important features. The article proposes an approach that minimizes the number of feature engineering operations by using encoding feature values in the form of images. This approach was tested on data about the parameters of the manufacturing process in a glassworks. This complex data set, which contains many highly correlated features, was transformed into images. Two image-coding methods were tested. The sets of images were used to train convolutional neural networks (CNNs), which perform the task of predicting the level of defective products. The research conclusions show that the method of encoding features influences the selection of the optimal CNNs architecture. The CNN classifiers were compared with benchmark models (BMs) trained on original data, not transformed into images. The results obtained showed greater effectiveness of CNNs compared to BMs. The article may be an inspiration for further research on simplifying and automating the feature engineering process, especially for multidimensional data sets from manufacturing processes.