Global competition and the new focus on the carbon and energy footprints of manufacturing processes have led manufacturers to seek architectures, methods, and algorithms that enable advanced control and decision-making support; specifically, automated systems that reduce the dependence on human resources and error. At the core of these automated systems lie robust monitoring algorithms that work with the entirety or a subset of the information available in the system, through its hardware sensors.
In this communication, we report progress on such a monitoring system developed for the precision machining of metals, with support from the Clean Energy Smart Manufacturing Innovation Institute. Specifically, we study face milling with a multi-point cutting tool that commonly operates inefficiently because of tool wear. The manufactured components can be of poor quality if machining is performed with a worn tool, leading to high scrap yields and energy wastage. Tool wear is also detrimental to the machine operation, as it can lead to other system faults. The proposed automated system for detecting the tool condition is learned using run-to-failure machining tests, conducted with varying machine settings. We use microscope flank wear measurements to assess tool conditions and leverage this information to identify indicators in the available machine measurements that can help with the accurate inference of tool wear. Available real-time sensory information, such as force and vibration, current and power, temperature, are all interrogated for their information as it relates to tool wear. The signals are transformed using wavelet transform, and the developed monitoring system exploits the variability of information in different wavelet subspaces. Compared to the Fast Fourier transform, wavelet analysis provides information about the frequencies existing in the signal and their time of occurrence as well. Localization is an important feature of the wavelet, which makes it suitable for analyzing time-varying signals. The coefficients of up to six levels in the discrete wavelet transform of the collected signals are analyzed for their correlation to the measured tool wear. We find that certain statistical features of the wavelet coefficients are very strong indicators of the tool state. We also observe better information in higher decomposition levels. Feature ranking was conducted, and a machine learning algorithm was developed using subsets of the most highly ranked features, for inferring tool wear. Vibration and force signals are found to be the most informative sensors in this application.