Book Description

Condition monitoring (CM) data, or simply monitoring data, is defined as a dataset that has been collected from individuals along the time, and it implicitly manifests the underlying unobservable system status. Due to the advances in sensory devices and information technology, prognosis and diagnosis in various fields can take enormous advantages from the rich condition monitoring data. However, at the same time, it also creates new challenges for research in data analytics as to how this vast and complex data could be utilized to retrieve accurate diagnosis and meaningful prognosis. Many existing techniques fall short of addressing this issue because most of them are developed when the data were collected in a well-controlled experimental setting. However, the monitoring data often involves many factors that are uncontrollable and, inevitably, has severe heterogeneity. This research simultaneously addresses multiple challenges that arise from the monitoring data. i. An Individualized model is crucial for effective diagnosis and prognosis based on the monitoring data. The primary focus of collecting the monitoring data is to understand the specific in-service unit rather than studying the population behavior. Therefore, the predictions or diagnostic decisions based on the monitoring data needs to be highly individualized. Collecting monitoring data happens in the on-line stage at the real time. Thus, the model should be able to update or adjust itself according to the newly-collected data points from the specific individual. ii. A non-parametric framework that can account for heterogeneity and handle high dimensionality in the data is needed. CM signals may not follow any parametric form, and if the specified form is far from the truth, the modeling and prognosis results will be misleading. For instance, parametric representations are typically based on physical or chemical theories; however, in most cases, such theories are unknown. Therefore, functional forms should be derived through empirical evaluation or visual observation, making them sensitive to model misspecification iii. A flexible modeling strategy that can handle multiple data types simultaneously is needed. Specifically incorporating qualitative data and introducing a distance measure between such data is essential for better prognosis. The monitoring data comes in various data types. In the literature, there are many statistical methods that are developed for a specific type of data which is not suitable for the monitoring data that includes various data types at the same time. Thus, a novel statistical model fusion needs to be investigated. iv. A scalable approach specifically when the number of CM signals/functional outputs is large. Further, the integrative analysis of multiple outputs implicitly assumes that these outputs share some commonalities. However, if this does not hold, negative transfer of knowledge may occur, which leads to decreased performance relative to learning tasks separately. Therefore, the model needs to possess excellent scalability when the number of outputs is large and simultaneously minimizes the negative transfer of knowledge between uncorrelated outputs. To address those issues listed above, four tasks are investigated in this report. (a) To build a mixture mixed effects model which is able to account for imbalance (early vs late failure) in the data. This technique greatly improves prognostics specifically for systems where most units are reliable and only few tend to fail at early stages of their life cycle. (b) To propose an alternative view on modeling CM data using multivariate Gaussian process. This view draws its roots from multitask learning and is based on treating each CM signal as an individual task. This technique is non-parametric, scalable and is able to account for heterogeneity in the data. (c) To incorporate qualitative features in non-parametric prognostics through a reparametrization technique called hypersphere decomposition. This technique allows incorporating external factors into prognostic models through defining a distance measure based on a unit hypercube. (d) To provide scalability for the multivariate Gaussian process when the number of outputs is large and to minimize the negative transfer of knowledge between uncorrelated outputs. This technique utilizes a distributed estimation scheme which allows scaling to arbitrarily large datasets through parallelization.