Tool wear monitoring of TC4 titanium alloy milling process based on multi-channel signal and time-dependent properties by using deep learning. (October 2021)

• Record Type: Journal Article
• Authors: Yan, Boling; Zhu, Lida; Dun, Yichao
• Is Part Of: Journal of manufacturing systems. Volume 61(2021)
• Page Start: 495
• Page End: 508
• Publication Date: 2021-10
• Subjects: Tool wear -- Deep learning -- Milling process monitoring -- Convolution neural network -- Signal processing; Manufacturing processes -- Periodicals; Production engineering -- Data processing -- Periodicals; Robots, Industrial -- Periodicals; Production, Technique de la -- Informatique -- Périodiques; Robots industriels -- Périodiques; Electronic journals; 670.42
• DOI: 10.1016/j.jmsy.2021.09.017
• Languages: English
• ISSNs: 0278-6125

Abstract

Highlights: The prediction of tool wear is based on multi-channels signals as well as number of cuts (cutting times actually), and the accuracy is increased remarkable compared with the results only based on signals. The created dataset containing relatively large amounts of signals and tool wear values. According to the measuring experience, the maximum flank wear band width is selected as the labels to save measuring time and increase accuracy of prediction. The preprocessing of signals is to conduct STFT to transform one-dimensional signals into two-dimensional signals. Afterwards, the transformed signals are concatenate with the number of cuts after dimension increment. Abstract: Tool wear in machining of TC4 can largely affect processing efficiency and quality, and unaware of tool condition may cause huge economic losses. This paper presents a monitoring method of tool wear, which can predict the tool wear in real time from the force and acceleration signals that acquired during the milling process. Firstly, the time domain signal is transformed into two-dimensional time-frequency domain signal. Afterwards, the two-dimensional signal is concatenate with the number of cuts after dimension-increment, and then the experimental data and labels are put into the residual network for training. The mean square error (MSE) of real wear and predicted wear is taken as loss function. Comparing with the process of feature extraction, it is found that the prediction of deep learning is time-saving, and it can be used for tool wear prediction with maximum error around 8 μm . Finally, various length and combination signals are input into the trained residual network to test the generalization and transferability of network.