Construction of an Online Machine Tool Wear Prediction System by Using a Time-Delay Phase Space Reconstruction-Based Dilation Convolutional Neural Network

Tool status testing is an important technology for enhancing processing efficiency and quality of a work piece. In terms of wear monitoring, the main approaches include offline and online tests. As the sensor technology grows over the past few years, by installing a sensor onto the spindle of a machine tool or one of the common approaches of the working platform, sensor signals go through preprocessing and characteristic/approach processing in this study, with the machine learning method entered in order to create the tool wear monitoring model. For the sake of cost and actual application, single-axis acceleration is adopted in this study for vibration signal extraction. Sensor signals are inputs for the model and the online tool image test system developed in this study is applied, with the tool wear value obtained as the model output, to complete the dataset of the tool wear forecast model. In terms of signal pretreatment, time delay is adopted for phase plane reconstruction. 1-D signals are expanded into 2-D pictures. Then, phase plane characteristics are reinforced by processing images in the picture. Following that, such image is combined with the time process parameters and the on-line image test wearing quantity. Learning takes place through the 2-D dilation convolutional neural network. Meanwhile, model precision is compared with the signal processing method featuring traditional time domain and frequency domain. It has been proven in studies that the time delay approach plus the neural network model framework not only have higher wear forecast precision but also help save lots of training time cost due to the fact that the model is not complex and the model can be trained very fast.

Automated summary

Tool status testing is crucial for improving processing efficiency and quality. This study develops a tool wear monitoring model using sensor signals and machine learning. By preprocessing the signals and applying a neural network model, higher wear forecast accuracy and reduced training time cost can be achieved.