Machine Learning Model for a Sustainable Drilling Process

Abstract

Drilling process is one of the most performed machining processes. Across many industries, drilling process directly influences the product quality as it is usually used in the final production steps before assembly. Drilling quality depends on the process parameters such as the spindle speed and feed. Improper selection of these parameters can lead to several defects like high surface roughness and burr formation. Consequently, the final product will not function properly, resulting in higher wastage of material, cost, and time. Additionally, reworking results in the loss of many resources. From a sustainability point of view, rework has a negative environmental impact as it increases electricity consumption and carbon emissions. Thus, the optimization of drilling process parameters is essential to produce high-quality products and make the process more cost-effective, efficient, and sustainable. Many experiments have been done to model the drilling process responses in terms of the input parameters. As a result, there is large amount of data available in the literature for drilling input parameters and their responses. This project aims to use big data analytics to make use of the data gathered from previous studies to model and optimize the responses. The collected data have some missing values because of the different input parameters and responses chosen for each experiment. To handle these missing values, deletion and imputation methods are used. For data analysis, various machine learning algorithms are used to model the process responses. The analysed responses are the surface roughness, thrust force, and drilling time. Further analyses are done including features selection and partial dependence. Moreover, several optimization runs are performed to assess different drilling cases. According to the results, the best algorithms for predicting the surface roughness, thrust force, and drilling time are bagged trees with 6 parameters, exponential GPR with 6 parameters, and fine trees with 5 parameters, respectively. The obtained accuracies are 92.91% for the surface roughness model, 81.33% for the thrust force model, and 90.61% for the time model. The obtained models can be used to find the optimal values of the input parameters that will give the minimum surface roughness, thrust force, or time without the need to conduct any experiments which leads to time, money, and resources savings.