Data acquisition and CNN-LSTM architecture
Abstract
Effective and precise prediction of tool wear plays a key role in improving machining efficiency, and product quality and reducing production cost. The majority of earlier studies have depended on limited experimental data, which may not be sufficient to estimate tool wear and surface quality. Aiming at these issues, the present study proposed a convolutional neural network (CNN)-long short-term memory (LSTM) hybrid deep neural network model that directly utilizes heterogeneous data including timely captured tool images, working conditions, vibration data, surface roughness, flank wear length, and wear depth. First, experiments were conducted on AISI D2 steel at three levels of spindle speed and feed/tooth, and experimental results for wear length, wear depth, surface roughness, and vibration signals were collected. The time domain vibration signals were processed with a fast Fourier transformer and converted to the frequency domain, and 13 and 5 features were extracted from the time and frequency domain, respectively, and integrated with the heterogeneous data. Second, tool images were annotated using Roboflow software, and wear region information was collected using YOLOv8 and added to heterogeneous data. Third, the CNN-LSTM network was trained with heterogeneous data containing spatial and time-dependent features. The performance and accuracy of the proposed methodology were validated using experimental data collected at different working conditions. The results show that the CNN-LSTM model effectively predicted the tool wear length on the flank, with the root mean square error (RMSE) value of 0.219 mm, and the determination coefficient R2 value of 0.974; wear depth with the RMSE value of 0.018 mm and R2 value of 0.943; surface roughness with the RMSE value of 0.216 μm and R2 value of 0.956. The proposed methodology has significance in metal-cutting applications and provides a solution to predict tool conditions and surface quality accurately.
