Prediction of air handling unit module assembly times with machine learning

Abstract

Finding the optimal scheduling for a production of a manufacturing system is crucial if the system's profitability is to be maximised. Also, scheduling production and adhering to promised delivery times is an important part of customer satisfaction. But, optimisation of the production schedule is in many cases extremely difficult as manufacturing systems have become more intricate, dynamic and connected than ever before. In addition, data necessary for accurately estimating the processing times of individual processes in a manufacturing system may be missing or incomplete, in which case the estimations must be based on a process study. Traditionally, the processing times are estimated by applying analytical modelling methods or simulation. However, as the amount of collected data from manufacturing systems has increased, applying data-driven approaches for estimating the processing times has become more common. Therefore, applying the data-driven approaches might be a better option than the traditional method, if the data being collected from the manufacturing system is consistent and includes more relevant information than noise.

The goal of this study is to explore whether supervised machine learning methods can predict the processing times of an air handling unit module assembly process more accurately than two process study-based analytical modelling methods. Another goal of this study is to develop a method, which utilizes the collected data to predict the times required to complete various working stages of the assembly process. The machine learning methods that are applied in the problem of this study are Multiple linear regression, K-Nearest Neighbor, Random forest and Multi-layer neural network. In this study, data analysis is used to examine the quality of the collected data from the assembly process and the suitability of the data to apply the machine learning methods for the problem of this study.

Results of the study show that among the aforementioned machine learning methods, the Random forest model is the most suitable for predicting the processing times of the assembly process. The Random forest model predicts the processing times of air handling unit modules in the test data with a mean absolute error of 1 hour and 6 minutes, whereas a coarse analytical modelling method model achieves a mean absolute error of 2 hours and 49 minutes. Thus, it can be concluded that the Random forest model can outperform the coarse analytical modelling method model. The times required to complete working stages of the assembly process cannot be extracted from the available data using only supervised machine learning methods. Hence, a hybrid model was developed, combining a fine analytical modelling method with the Random forest method, to estimate the durations of the various working stages.