SURFACE ROUGHNESS AND CUTTING TOOL-WEAR DIAGNOSIS BASED ON BAYESIAN NETWORKS Jos´ e Vcte Abell ´ an Nebot * Rub´ en Morales-Men´ endez ** Antonio J Vallejo Guevara ** Ciro A. Rodr´ ıguez ** * Universitat Jaume I de Castell´ on, Espa˜ na ** ITESM, Monterrey NL, M´ exico abellan@esid.uji.es, {rmm, avallejo, ciro.rodriguez} @itesm.mx Abstract: Next-generation manufacturing systems demand more intelligent features such as diagnosis, adaptation, etc. One of the main goals in machining systems is to find an appropriate trade-off among cutting tool condition, surface roughness and productivity. However, both surface roughness and cutting tool wear are impractical to measure online. This work proposes a multi-sensor system for indirect monitoring based on Bayesian Networks (BN) framework. Unlike classical modelling techniques, a important advantage of BN models are their ability to deal with the stochastic nature of the machining process. We discuss BN models based on their precision and reliability for different discretization sizes of their continuous variables. Early results show that models with high discretization exhibit a reliability of 89.5% for the surface roughness prediction and 97.3% for the cutting tool wear diagnosis. However, lower discretization presents better reliability but worse diagnosis. Copyright c 2006 IFAC Keywords: Probabilistic models, machining, sensor systems, intelligent manufacturing systems, diagnosis, prediction methods. 1. INTRODUCTION Machining process monitoring for surface roughness and cutting tool condition has received great attention because their industrial implications. Surface roughness is a widely used index of product quality and in most cases a technical requirement for mechanical products (Benardos and Vosniakos, 2003). Cutting tool wear impacts negatively affecting dimen- sions, finish and surface integrity (Liang et al., 2004); also, it represents about 20% of machine tool down- time for failures. Unfortunately, both surface rough- ness and cutting tool-wear are impractical to measure online and depend on several machining parameters. Direct methods include optical techniques as well as radioactivity analysis, workpiece size measurement, etc. One important practical drawback to the indus- trial environment for process monitoring sensors is the high cost and sometimes the intrusive nature in production lines. So, model-based techniques (indi- rect methods) have been proposed. Indirect methods include measurement of alternatives variables (such as cutting force, acoustic emission, sound, vibration, spindle power, cutting temperature, etc.) and combi- nation of them into a mathematical model. Classical mathematical models are based on statistical tech- niques such as time series and regression as well as neural network approaches. Due to the stochastic nature of machining systems, we want to evaluate the feasibility of the Bayesian Network (BN) framework for modelling both surface roughness and cutting tool wear through a multi- sensor system. 408