Self-adaptive evolving forecast models with incremental PLS space updating for on-line prediction of micro-fluidic chip quality

Abstract

An important predictive maintenance task in modern production systems is to predict the quality of products in order to be able to intervene at an early stage to avoid faults and waste. Here, we address the prediction of the most important quality criteria in micro-fluidics chips: the flatness and critical size of the chips (in the form of RMSE values) and several transmission characteristics. Due to semi-manual inspection, these quality criteria are typically measured only intermittently. This leads to a high-dimensional batch process modeling problem with the goal of predicting chip quality based on the trends in these process values (time series). We apply time-series based transformation for dimension reduction to the lagged time-series space using of partial least squares (PLS), and combine this with a generalized form of Takagi–Sugeno(TS) fuzzy systems to obtain a non-linear PLS forecast model (termed as PLS-fuzzy). The rule consequent functions are robustly estimated by a weighted regularization scheme based on the idea of the elastic net approach. To address particular system dynamics over time, we propose dynamic updating of the non-linear PLS-fuzzy models using new on-line time-series data, with the options 1.) adapt and evolve the rule base on the fly, 2.) smoothly down-weight older samples to increase flexibility of the fuzzy models, and 3.) update the PLS space by incrementally adapting the loading vectors, where processing is achieved in a single-pass stream mining manner. We call our method IPLS-GEFS (incremental PLS combined with generalized evolving fuzzy systems). We applied our predictive modeling approach to data from on-line microfluidic chip production over a time period of about 6 months (July to December 2016). The results show that there is significant non-linearity in the predictive modeling problem, as the non-linear PLS-fuzzy modeling approach significantly outperformed classical PLS for most of the targets (quality criteria). Furthermore, it is important to update the models on the fly with incremental updating of the PLS space and/or with down-weighting older samples, as this significantly decreased the accumulated error trends of the prediction models compared to conventional updating. Reliable predictions of flatness quality (with around 10% error) and of RMSE values and transmissions (with around 15% errors) can be achieved with prediction horizons of up to 4 to 5 h into the future.