應用整體學習方法於設備元件剩餘壽命之預測

Abstract

用於生產各式產品的機台一直是工廠中很重要的設備，因此如何預測機台設備零件的剩餘壽命(remaining useful life, RUL)以避免機台的耗損是一個重要的議題。然而機台設備零件的資料相當複雜且龐大，使用單一個模型往往很難得到良好的預測，因此在本論文中以具線性混和效應(linear mixed effects)之參數衰變模型(parametric degradation models)為基礎學習演算法(base learner)，使用三種整體學習方法(ensemble learning):靴拔重抽總合法(bagging)、推升法(boosting)與堆疊法(stacking)，以改善基礎學習法在預測與分析上的準確度。本論文將以一組由生產LED晶片的MOCVD機台上所獲得之感測資料(censoring data)為例，來說明與展示所提出的整體學習方法。為了因應例題機台設備資料的重複測量值(repaeted measurements)型態，我們使用移動區塊靴拔法(moving block bootstrap)取代原來整體學習方法的靴拔法(bootstrap)。

The machine equipment that is used to produce several products is important facility in factory. Therefore, how to predict the remaining useful life (RUL) of equipment components to avoid damage to the machine is an important issue. The censoring data from the machine are usually too huge and complicate to get accurate RUL prediction by one single model. Thus, in this thesis, we adopt a parametric degradation model with linear mixed effects as the base learner and combine three popular ensemble learning approaches: bagging, boosting and stacking to improve the prediction accuracy of the base learner. This thesis analyzes a set of censoring data from the MOCVD machine equipment that produces LED chips, and uses them to demonstrate the usefulness of the proposed ensemble learning approaches. Because the analyzed censoring data contain repeated measurements, instead of using the traditional bootstrap sampling method, we use the moving block bootstrap sampling in the ensemble learning procedures.