Our student Marcella Miller defends her PhD dissertation on enhancing PHM models
We’re thrilled to share that our graduate student Marcella Miller recently defended her PhD dissertation, titled “Modeling of Complex Visual Patterns in Multivariate Time Series Data for Enhanced PHM” . Her dissertation synthesizes years of cutting-edge research on global anomaly detection models, covering critical applications in semiconductor manufacturing, commercial aero-engine performance, and ICU patient monitoring. Congratulations to Dr. Miller on a brilliant career at Center for Intelligent Metrology & Sensing!
Learn more by reading the dissertation abstract below:
Abstract:
In prognostics and health management (PHM), commonly employed methods often incorporate several potential drawbacks that limit the utility of the approach in quickly producing high quality results. There is a general reliance on prior knowledge of the system physics, background expertise on the problem or solution domain, and/or ample data availability that prevents the democratization of PHM for a wide array of users and applications. Additional challenges also arise in the case of exploring PHM for time series data. Samples may show drift and shift over time, which creates siloed data that must be considered under its individual conditions rather than as a cohesive, all-encompassing dataset. Time-based characteristics of individual samples can vary, which further complicates data processing and increases the gap between data and results. Collected data will not always contain a complete instance, so additional considerations are required to determine how to fit the available data into the overall process. These factors also inhibit optimum realization of the potential of existing data to provide beneficial reports of system status. Therefore, the opportunity exists to create more automated, flexible methods of assessing PHM tasks, particularly for time series data. To achieve this, the proposed method leverages visual patterns in the time series data to re-define the standard PHM process and eliminate the most timeconsuming steps as well as the need for advanced expertise and background information. With dynamic time warping principles serving as a foundation, the developed framework incorporates three main stages of pattern discovery, distillation, and dissemination wherein visual patterns in the data are automatically or semi-automatically identified, grouped, summarized, and deployed as concentrated segments of knowledge for enhanced outcomes and improved insights. Strategies for overcoming time series specific challenges are innately incorporated into the framework. The methodology and approach are demonstrated in three case studies: reducing model requirements for fault detection in a semiconductor etch process, automating labeling for key features of intracranial pressure waveforms, and predicting the remaining useful life of aero-engines. The first two case studies operate only on complete samples while the last case study applies the methods to incomplete samples. Results obtained in the case studies indicate that the proposed methodology is a promising tool for bridging the gap between time series data inputs and insightful PHM outputs.
