Acoustic emission based grinding wheel wear monitoring: Signal processing and feature extraction. (July 2022)

• Record Type: Journal Article
• Authors: Shen, Chia-Hsuan
• Is Part Of: Applied acoustics. Volume 196(2022)
• Publication Date: 2022-07
• Subjects: Machine learning -- Condition monitoring -- Acoustic emission -- Surface grinding; Acoustical engineering -- Periodicals; Periodicals; 620.2
• DOI: 10.1016/j.apacoust.2022.108863
• Languages: English
• ISSNs: 0003-682X

Abstract

Graphical abstract: Highlights: Frequency spectrum alone produces quality feature sets. Signal segment as short as 1 ms is ample for wear detection. High frequency AE events provides consistent classification performances. Entry or exit AE transient spikes does not affect overall classifications. Abstract: Surface grinding, being one of the later stages of a manufacturing process, is crucial to achieve the desired workpiece quality by maintaining proper wheel condition. Machine learning and acoustic emission (AE) detection have demonstrated potentials for developing an effective on-line wheel wear monitoring system. However there are inconsistencies among existing studies regarding to signal sampling and feature extraction methods. In the present paper, surface grinding experiments using medium carbon steel workpieces were conducted to determine the effects of using different signal analysis window lengths, feature types, and AE sensor bandwidths on the wheel wear detection performances. In addition, the use of AE transient spikes occurred during wheel entry and exit were analyzed. Feature were extracted from time domain (TD), frequency domain (FD), and time–frequency domain (TFD) representations of the measured AE signals, which were then refined using sequential floating forward selection (SFFS) algorithm to reduce feature redundancy. Support vector machine (SVM) algorithm with linear kernel was used for pattern classification. A number of insights for future monitoring system development were obtained. Firstly, while the lower frequency AE sensor (below 100 kHz) provided adequate classification performances, the higher frequency AE sensor (above 100 kHz) produced more robust results. Secondly, 100% classification rate was achievable using features derived from Fast Fourier Transform (FFT) in combination with descriptive statistics. Thirdly, it was determined that a window length as short as 1 ms (1000 data points) could capture enough signal information using FFT with the higher frequency AE sensor. Lastly, the choice sampling instance did not significantly impact classification outcomes.