Machine learning-based prediction of processing time in furniture manufacturing to estimate lead time and pricing

Abstract

Furniture manufacturing plants are mainly small to medium enterprises (SMEs) and must merge customized mass production into their schedule to meet the market demand. Furniture plants produce a diverse array of models, with each process uniquely adding to the costs. In this multiproduct, multipart and multi-process manufacturing, it is difficult to accurately predict the processing time of new models and the lead time for highly customized orders. The processing time of parts is critical for optimizing, estimating the lead time and pricing the products, particularly for new models. Machine Learning (ML) is a useful tool to analyze and control manufacturing parameters and could be applied to furniture factories too. In this study the authors demonstrated the use of a ML-based framework to predict the processing time of wooden furniture based on the design of parts and actual manufacturing data. Specifically, the objectives are to define the accuracy of Convolutional Neural Networks (CNN) in classifying furniture parts according to their design characteristics into categories such as Plain, 2D, and 3D curved, and define the accuracy of Artificial Neural Networks (ANNs) in taking CNN data along with real manufacturing processing time data for identifying and analyzing intricate correlations between parts and manufacturing processes, thereby facilitating precise prediction of processing time. Images of the furniture’s parts design and data from a time and motion study in mass production in a plant were used to develop the models. The models' R², Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) were calculated as a criterion for defining accuracy. Random Forest and Gradient Boosting regression models were developed to compare and validate against ANN for predicting processing time, ensuring the robustness and reliability of the ML-based framework. All four models showed successful performance with R² scores above 0.90, MSE below 1, and MAPE below 10, except 10.26 in Random Forest and 11.15 in Gradient Boosting. However, ANN showed significantly higher accuracy than other traditional regression models comparing MAPE of 1.63 to 10.26 in ANN and Random Forest respectively demonstrating its better performance in analyzing intricate relationships of input features and outputs.