A Data Processing Pipeline for Prediction of Milling Machine Tool Condition from Raw Sensor Data

作     者：Ferguson, M. Bhinge, R. Park, J. Lee, Y. T. Law, K. H. 

作者机构：Stanford Univ Civil & Environm Engn Y2E2 Bldg473 Via Ortega Stanford CA 94305 USA Infinite Uptime Inc 2150 Shattuck Ave Berkeley CA 94704 USA Korea Adv Inst Sci & Technol Ind & Syst Engn 291 Daehak Ro Daejeon 305701 South Korea NIST Syst Integrat Div 100 Bur Dr Gaithersburg MD USA Stanford Univ Civil & Environm Engn Y2E2 Bldg Stanford CA 94305 USA 

出 版 物：《SMART AND SUSTAINABLE MANUFACTURING SYSTEMS》 (Smart Sustain. Manufact. Syst.)

年 卷 期：2018年第2卷第1期

页      面：40-63页

核心收录：

基　　金：NIST [70NANB17H031] 

主　　题：tool condition monitoring Gaussian process regression machine learning sustainable manufacturing predictive model markup language 

摘      要：With recent advances in sensor and computing technology, it is now possible to use real-time machine learning techniques to monitor the state of manufacturing machines. However, making accurate predictions from raw sensor data is still a difficult challenge. In this work, a data processing pipeline is developed to predict the condition of a milling machine tool using raw sensor data. Acceleration and audio time series sensor data are aggregated into blocks that correspond to the individual cutting operations of the Computer Numerical Control milling machine. Each block of data is preprocessed using well-known and computationally efficient signal processing techniques. A novel kernel function is proposed to approximate the covariance between preprocessed blocks of time series data. Several Gaussian process regression models are trained to predict tool condition, each with a different covariance kernel function. The model with the novel covariance function outperforms the models that use more common covariance functions. The trained models are expressed using the Predictive Model Markup Language where possible to demonstrate how the predictive model component of the pipeline can be represented in a standardized form. The tool condition model is shown to be accurate, especially when predicting the condition of lightly worn tools.