Thumbnail Image

Sensor based real-time process monitoring for ultra-precision manufacturing processes with non-linearity and non-stationarity

Beyca, Omer Faruk
This research investigates methodologies for real-time process monitoring in ultra-precision manufacturing processes, specifically, chemical mechanical planarization (CMP) and ultra-precision machining (UPM), are investigated in this dissertation.
The three main components of this research are as follows: (1) developing a predictive modeling approaches for early detection of process anomalies/change points, (2) devising approaches that can capture the non-Gaussian and non-stationary characteristics of CMP and UPM processes, and (3) integrating multiple sensor data to make more reliable process related decisions in real-time.
In the first part, we establish a quantitative relationship between CMP process performance, such as material removal rate (MRR) and data acquired from wireless vibration sensors. Subsequently, a non-linear sequential Bayesian analysis is integrated with decision theoretic concepts for detection of CMP process end-point for blanket copper wafers. Using this approach, CMP polishing end-point was detected within a 5% error rate.
Next, a non-parametric Bayesian analytical approach is utilized to capture the inherently complex, non-Gaussian, and non-stationary sensor signal patterns observed in CMP process. An evolutionary clustering analysis, called Recurrent Nested Dirichlet Process (RNDP) approach is developed for monitoring CMP process changes using MEMS vibration signals. Using this novel signal analysis approach, process drifts are detected within 20 milliseconds and is assessed to be 3-7 times faster than traditional SPC charts. This is very beneficial to the industry from an application standpoint, because, wafer yield losses will be mitigated to a great extent, if the onset of CMP process drifts can be detected timely and accurately.
Lastly, a non-parametric Bayesian modeling approach, termed Dirichlet Process (DP) is combined with a multi-level hierarchical information fusion technique for monitoring of surface finish in UPM process. Using this approach, signal patterns from six different sensors (three axis vibration and force) are integrated based on information fusion theory. It was observed that using experimental UPM sensor data that process decisions based on the multiple sensor information fusion approach were 15%-30% more accurate than the decisions from individual sensors. This will enable more accurate and reliable estimation of process conditions in ultra-precision manufacturing applications.