Cost-effective classification of tool wear with transfer learning based on tool vibration for hard turning processes

• A novel application of transfer learning for tool wear detection in turning processes using one-dimensional (1D) convolutional neural network (CNN). • Transfer learning of a source model for tool wear detection to a new CNC turning machine using new parameters. • Applications of low-cost MEMS and IEPE accelerometers in tool wear detection. • Evaluation of the performance of low-cost MEMS accelerometers in transfer learning. • Transfer learning using significantly smaller amount of vibration data for the target model in tool wear detection. This paper presents a novel application of transfer learning for tool wear detection in turning processes using one-dimensional (1D) convolutional neural network (CNN). The work also investigates the applications of low-cost MEMS as well as IEPE accelerometers in tool wear detection. Tool wear detection is performed using a classification method in which two classes of tool wear sizes are defined: class one and class two which correspond to tool wear sizes smaller or equal to 0.1 mm and bigger than 0.1 mm respectively. The advantage of transfer learning is that it utilizes knowledge from previously learned tasks and applies them to the related ones. In the case of CNC machines, tool wear mechanism in the turning process and the working principles are similar thus, transfer learning can be used to generalize the specific cutting condition to a broader use case. The transfer learning model in this paper uses a pretrained tool wear prediction model which was composed of 1D CNN layers with full connection layers and sufficient amount of data on a source CNC machine. The CNN layers of the pretrained model are frozen at first and then the full connection layers are trained with the new data from the target CNC machine with different cutting insert types. In this study, the application of transfer learning in tool wear size classification showed that the pretrained tool wear detection models can be transferred to other similar processes while maintaining a high accuracy. The accuracy of the transfer learning model was evaluated by comparing it with the results of a model developed from scratch using only the target machine data. It is demonstrated that the transfer learning maintained an accuracy of higher than 80%. In addition to this, the transfer learning model significantly increased the tool wear classification accuracy using the single-axis low-cost MEMS accelerometer from 58% to 85%. Moreover, the tool wear classification model using transfer learning significantly reduced the amount of data required for model development. To evaluate this, the accuracy of the transfer learning model was tested by further reducing the amount of the training data by maximum 80% and the model still showed an accuracy higher or equal to 80%.