Probabilistic Based Classification Techniques for Improved Prognostics Using Time Series Data

Recent progress in data collection has enabled the expansion and availability of raw data to take place at an explosive rate. Most of these data, which have been acquired over months and years, are "time series data". From Prognostics and Health Management (PHM) points of view, it is important to convert raw time series data into useful information quickly and accurately. Therefore, it becomes significant to have comprehensive understanding of data format to support PHM. Most prognostics and health management algorithms hypothesize that the data is a random model drained from a stationary distribution; however, one of the difficult situations for training time series data is a non-stationary environment. Learning in non-stationary environment, also known as learning concept drift, is concerned with interpreting data whose statistical characteristics change over time.;Because of the complex intrinsic characteristics of concept drift, learning from such data requires new understandings, techniques, and algorithms to efficiently transform huge amounts of raw data into useful information for PHM usage. This dissertation provides a comprehensive review of three different approaches: physics based model, a data-driven model, and a combination of the two known as a coupled model. These approaches enable the detection of key classification challenges related to the implementation of a prognostics model on time series data. To overcome the classification and concept drift issues with time series data, a novel approach from the coupled model is introduced by applying advanced probabilistic ensemble techniques adapted to the nature of data.;For better understanding of the proposed method, its application has been explained in three different case studies. The first case study shows the probabilistic classification of healthy and faulty lithium-ion batteries in a pack using the coupled model. The second case study introduces a method for improving the accuracy of mobility performance assessments of electric vehicles based on driving behavior classification. The driving behavior can drastically influence the mobility performance of electric vehicles. Applying the probabilistic classification technique helps us to classify the changes of driving behavior of a driver over time more accurately. Therefore, the model can predict the mobility of the EV based on aggressive or moderate driver more precisely. Finally, the last case study has been applied to some GE industrial machines to identify best time for maintenance activity. This proposed approach efficiently classifies different maintenance activities in order to develop the probability risk assessment model and predict the high risk and low risk time intervals in the upcoming year. The method was applied in the 2014 PHM data challenge and won first rank at the competition.