Predicting cutting tool wear with machine learning techniques in automotive industry

In companies engaged in machining operations, developing an effective method for predicting and preventing tool wear is critically important for production efficiency. This study aims to predict tool wear using data obtained from CNC machines in a company operating in the automotive sector, utilizing machine learning. Python programming language was used in the implementation. In the process of organizing the datasets, data were cleaned, outliers were checked, features were determined based on correlation, data scaling was performed, and data were divided into windows. Machine learning models such as Random Forest Regressor (RFR), Gradient Boosting Regressor (GBR), Extreme Gradient Boosting Regressor (XGB), and Adaptive Boosting Regressor (ABR) were used. These models were evaluated on three different scenarios created from different combinations of datasets. Model performances were improved by hyperparameter tuning. Subsequently, the datasets were divided into stages of tool life, and experiments were conducted under the assumption that different machine learning models could better predict different stages of tool wear. The models were evaluated and compared across all scenarios for different window sizes and phases using the Mean Squared Error (MSE) performance metric. In the first phase of Scenario 1, the GBR model was successful with an MSE value of 0.1096, in the second phase the XGB model with an MSE of 0.0242, and in the final phase the GBR model with an MSE of 0.0313. In the first phase of Scenario 2, the GBR model was successful with an MSE value of 0.1615, in the second phase again the GBR with an MSE of 0.0434, and in the final phase the RFR model with an MSE of 0.0404. In Scenario 3, which did not include phase analysis, the best model was the RFR with an MSE value of 0.0041 for a window size of 50. When the models were run without dividing into phases, the best MSE value for Scenario 1 was 0.0614 with GBR, and for Scenario 2 it was 0.0908 with GBR. Dividing tool wear into stages allows for more precise predictions, providing businesses with the opportunity for more accurate and timely interventions.