Dense-Block Structured Convolutional Neural Network-Based Analytical Prediction System of Cutting Tool Wear

Cutting tool wear strongly affects machining processes. Conventionally, multiple types of sensors are installed on the main rotating axis and platform of a machine, and signal analysis is used to determine the condition of the tool in real time through machine learning. With the traditional approach, to measure tool wear, it is required to remove the tool and place it underneath the microscope for measurement to obtain relatively precise results. The time cost, however, is relatively high; it is not cost-effective. With the charge-coupled device (CCD) camera installed in the machine tool, tool wear may be estimated through programed processing. The CCD camera, however, will affect spatial configuration in the machine tool and additional time is needed for tool wear measurement upon completion of processing each time. With the accelerometer sensor combining an offline or online tool wear measurement system and using the training method in this research model upon completion of the dataset, forecasting can be done with the precompleted model, which saves the time needed for offline or online wear measurement. However, the deployment of multiple sensors is difficult and expensive in practice. Used in this study were milling data, namely acoustic emission (AE), vibration, and current data from sensors installed on the rotating axis and platform of a machine. The data were from a National Aeronautics and Space Administration (NASA) dataset. The accelerometer data were used to develop a novel machine-learning algorithm. The vibration signals were integrated with other machining parameters. Signal preprocessing was used to reduce interference from environmental noise, and parameter records and feature signals were analyzed. A 1-D convolutional neural network (CNN) model similar to the DenseNet framework was developed. The model was optimized with framework and parameter adjustment and verified using ${k}$ -fold cross-validation. The model had a mean absolute error (MAE) of 0.06 and a root mean square error (RMSE) of 0.09. Compared with other machine-learning models, the developed model has higher accuracy.