| Prognostics focus on failure prediction in order to prevent unexpected machine downtime; which can have major impact on costs in industry. Despite progress made in recent decades, many prognostics techniques have had limited success due to the reliance on ad hoc approaches. The novel adaptive prognostics framework presented in this dissertation can provide robust prognostic information and is capable of being reconfigured for diverse applications. The proposed framework uses a self-organizing map (SOM) based method to quantitatively assess degradation statuses, based on which a reinforcement learning agent is trained to provide guidance for dynamically selecting the appropriate prediction models under various degradation statuses. A new density estimation method, which utilizes a boosting Gaussian mixture model (GMM), is proposed to improve the prediction accuracy, after the most appropriate prediction model is determined, by taking into consideration prediction uncertainties. Case studies show the proposed method achieves the highest accuracy compared to traditional F-test, as well as several auto-regressive moving average (ARMA) models with different orders. In order to address the issue of deploying the right prognostics tools for the right applications, a methodology for designing the architecture of a reconfigurable prognostics platform (RPP) is also proposed. This methodology is validated in two industrial case studies, which demonstrate that the RPP is both feasible and effective. With a reconfigurable architecture, the proposed adaptive prognostics framework can automate the prediction model selection procedure, which enables it to be applied to many rotary machine component applications. With an appropriate definition of the "degradation state", the proposed methodology can be used for general time series prediction in many other areas as well. |
