Analytical approach to similarity-based prediction of manufacturing system performance

• We present a new method for non-stationary and non-Gaussian time-series prediction time-series prediction.
• It uses similarities of the on-going time-series with its previous realizations.
• The method matches accuracy properties of another recent similarity based method.
• It achieves that accuracy with an order of magnitude faster calculation.
• The computations are accelerated because of the analytical updating of predictions.

In this paper, a new method is proposed that is capable of predicting system condition by comparing the similarity of the most recent performance signatures with the known degradation patterns available in the historical records. For predicting the future performance, the similarities of the current performance signatures to each known degradation pattern are utilized in an analytically tractable manner to slant the prediction distributions toward most similar past degradation patterns. The newly proposed method was applied to prediction of sensor signatures coming from an industrial plasma enhanced chemical vapor deposition (PECVD) tool operating in a major semiconductor manufacturing fab. Results showed that the proposed method significantly improves the long-term time series prediction accuracy in terms of mean squared errors over the traditional autoregressive moving average (ARMA) model and additionally showed comparable mean squared prediction errors to another recently introduced similarity-based algorithm for long-term prediction of non-linear and non-stationary time series. However, the analytical structure of the method proposed in this paper enables computation of the prediction distributions an order of magnitude faster.