Predictive analytics for complex engineering systems using high-dimensional signals
MetadataShow full item record
Tremendous maintenance and operational expenditure are incurred due to unexpected failures, and inefficient maintenance and operational practices. Consequently, many capital-intensive assets used in the energy, manufacturing, and service sectors are equipped with numerous sensors that generate large amounts of high-dimensional data related to the physical performance and the operational characteristics of the asset. One of the key Big Data challenges stems from the need to analyze high-dimensional data, in real-time, to detect faults and predict the future state-of-health of critical assets (i.e., predictive analytics). This doctoral dissertation focuses on addressing several key challenges in predictive analytics for asset management and optimization. The first research challenge revolves around the development of prognostic methodologies (for predicting asset health and remaining operational life) that can scale with the size and complexity of high-dimensional data. In contrast, most existing research focuses on single time-series data applications or multivariate applications where only small-sized time-series vectors are considered. Furthermore, the limited research efforts that involve more complex data structures like profile and/or image data are limited to fault detection, and do not extend to prognostics—two fundamentally different problems. The second research component focuses on computational efficiency of analytic models. Specifically, we pursue fundamental research aimed at speeding up matrix computations of conventional statistical methodologies that enable their application in real-time prognostic applications. Many of the existing models have been validated utilized small-sized data sets, and thus computational challenges have often been overlooked. The third challenge, one that has traditionally been neglected due to lack of real-world data, deals with data quality and its impact on the accuracy and fidelity of the resulting analytics. Harsh industrial environments have a significant impact on the quality of sensor data that range from missing and fragmented data observations to corrupt values and outliers that often result in significant false alarms—a problem that plagues many industries. In this thesis, we focus on developing models that are relatively robust to applications that exhibit poor data quality.