Edwin Lughofer

Self-Adaptive Evolving Forecast Models in Predictive Maintenance Systems

Predictive maintenance relies on real-time monitoring and diagnosis of system components, and process and production chains. The primary strategy is to take action when items or parts show certain behaviors that usually result in machine failure, reduced performance or a downtrend in product quality.

In the first stage, it is thus of utmost importance to recognize potentially arising problems as early as possible. Therefore, a core component in predictive maintenance systems is the usage of techniques from the fields of forecasting and prognostics, which can either rely on process parameter settings (static case) or process values recorded over time (dynamic case). We will focus on the latter and demonstrate a robust learning procedure of time-series based forecast models, which can deal with very high-dimensional batch process modeling settings. Furthermore, our approach allows the forecast models to be on-line updated over time and on the fly whenever required due to intrinsic system dynamics (such as, e.g. varying product types, charges, settings, environmental influences) => leading to the paradigm of self-adaptive forecast models. This is achieved i) by recursive adaptation of model parameters to permanent changes and to increase model significance and accuracy and ii) by evolution of new model components (rules) on the fly in order to account for variations in the process, which require a change in the model’s ‘non-linearity degree’. We will also present some enhanced methods in model adaptation for increased flexibility to properly compensate system drift and shifts, such as dynamic forgetting, rule merging and splitting as well as an incremental update of the latent variable sub-space as a variant of incremental feature space transformation.

In the second stage, the evolved and incrementally adaptive forecast models can be used as surrogates in a fully automatized optimization procedure in order to prevent operator’s intervention and time-intensive manual reactions to predicted downtrends. Often, there are some “control wheels”, usually machine parameter settings which are able to change the behavior of the production process in order to meet the quality standards. Such settings may have indeed been optimized before (due to expert knowledge or in a static optimization process), but may not take into account dynamically changing factors during production. In other cases, such settings could not be optimized before at all (as requiring time-intensive design of experiments cycles), such that often a default parametrization is used which is suboptimal for the final product quality. We will define the optimization problem, which typically leads to a many-valued problem with very-high dimensional input parameter space; thus, we will demonstrate possibilities how a reduction to smaller problems can be achieved, and how the reduced problems can then be more quickly and robustly solved with multi-objective evolutionary algorithms.

The talk will be concluded with a real-world application scenario from a (micro-fluidic) chip production site, where the self-adaptive time-series based forecast models together with the process optimization component have been successfully applied in order to detect product quality downtrends at an earlier stage and to even suggest modified process values trends (and associated) machine parameter settings to improve and assure high-level product quality anytime.