| AVS 59th Annual International Symposium and Exhibition | |
| Vacuum Technology | Tuesday Sessions |
| Session VT-TuP |
| Session: | Vacuum Technology Poster Session and Student-built Vacuum System Poster Competition |
| Presenter: | S.H. Nam, KRISS, Korea |
| Authors: | S.H. Nam, KRISS, Korea W.J. Kim, KRISS, Korea J.Y. Lim, KRISS, Korea W.S. Cheung, KRISS, Korea |
| Correspondent: | Click to Email |
This paper introduces unique statistical features extracted from the measured state variables of dry vacuum pumps in the semiconductor processes. They were found to have three distinctive means and overlapped distributions, not a single normal distribution. More specifically, two distinctive distributions near the upper and lower asymptotic bounds are obviously observed from the gas-loaded states of the vacuum pump and the third one from the idle states. These observations have provided new motivations of not only separating the pump operation state into the gas-loaded and idle states but also modeling the upper and lower bounds as a separated distribution. A linear adaptive parametric model (APM) is proposed such that their linear trends of each state variable are shown to be mapped onto their model parameters. Those estimated model parameters are used to construct the batch data obtained after each process. The APM-based batches are also exploited to construct the batches under the normal operating conditions (NOC) such that the major eigenvectors of the NOC batches are used to diagnose the current process batch data. It should be noted that the APM-based batch provides a dramatic reduction of memory usage and computation time (for example, 1~2 % memory usage and 10 times faster computation time) in comparison to the conventional dynamic-time wrapping methods. The feasibility of the proposed APM for the predictive maintenance of dry vacuum pumps is demonstrated to be successful by illustrating test results obtained from the six dry vacuum pumps.
This paper proposes the use of two statistics, the Hostelling’s T2 and the sum of squared residuals, in order to improve the reliability of the predictive maintenance and self-diagnostics of vacuum pumps. The first one is exploited to examine what amount of similarity the current process batch has in reference to the normal operation conditions and, furthermore, the second is to examine what contribution the current process batch provides to the noise space. The two proposed statistics are examined to quantitatively analyze the reliability and improvement of the predictive maintenance and self-diagnostics developed in this work.