Real Time Identification of Incipient Surface Morphology Variations in Ultraprecision Machining Process

Abstract

Realtime monitoring and control of surface morphology variations in their incipient stages are vital for assuring nanometric range finish in the ultraprecision machining (UPM) process. A realtime monitoring approach, based on predicting and updating the process states from sensor signals (using advanced neural networks (NNs) and Bayesian analysis) is reported for detecting the incipient surface variations in UPM. An ultraprecision diamond turning machine is instrumented with three miniature accelerometers, a threeaxis piezoelectric dynamometer, and an acoustic emission (AE) sensor for process monitoring. The machine tool is used for faceturning aluminum 6061 discs to a surface finish (Ra) in the range of 15–25 nm. While the sensor signals (especially the vibration signal in the feed direction) are sensitive to surface variations, the extraneous noise from the environment, machine elements, and sensing system prevents direct use of raw signal patterns for early detection of surface variations. Also, nonlinear and timevarying nature of the process dynamics does not lend conventional statistical process monitoring techniques suitable for characterizing UPMmachined surfaces. Consequently, instead of just monitoring the raw sensor signal patterns, the nonlinear process dynamics wherefrom the signal evolves are more effectively captured using a recurrent predictor neural network (RPNN). The parameters of the RPNN (weights and biases) serve as the surrogates of the process states, which are updated in realtime, based on measured sensor signals using a Bayesian particle filter (PF) technique. We show that the PFupdated RPNN can effectively capture the complex signal evolution patterns. We use a meanshift statistic, estimated from the PFestimated surrogate states, to detect surface variationinduced changes in the process dynamics. Experimental investigations show that variations in surface characteristics can be detected within 15 ms of their inception using the present approach, as opposed to 30 ms or higher with the conventional statistical change detection methods tested.