Richard J. Povinelli
Machine Learning, Signal Processing, Dynamical Systems and Chaos
Marquette Energy Analytics forecasts approximately 20% of the daily natural gas used in the United States. We design models to use weather data, historical trends in the weather, and past load demand for the purpose of generating accurate predictions of natural gas use in a variety of climates and geographies.
Specific projects have included automatic anomaly detection and correction, probabilistic forecasting, forecasting using deep neural networks, and electricity forecasting.
Classifying Vestibular Disorders
Dizziness and imbalance are among the most common conditions for which patients seek medical evaluation, adults are disabled, and the elderly are injured. Unfortunately, misdiagnosis of vestibular disorders is common; some studies showing misdiagnosis rates of 74-81% in emergency departments and 85% in primary care offices. Such misdiagnosis is costly, inefficient and causes patient harm and dissatisfaction. At root of this misdiagnosis is a profound knowledge gap among non-vestibular specialty providers as to what patient information is relevant, critical, and diagnostic. Complete histories from vestibular patients can predict specific clinical conditions, risk of falls and injury, appropriate diagnostic tests, and cost-effective interventions and treatment. The ultimate goal of our work is to improve the accuracy and efficiency of diagnosing vestibular conditions by acquiring complete and accurate histories directly from the patient.
This project focused on the creation of a stochastic representation for the phase space embeddings of dynamical systems with application to the task of speech classification and recognition. We developed a general stochastic model for such signal embeddings, tested the model through classification simulations, then applied the technique to both isolated and continuous speech recognition. The goal of the research was to discover time-domain analysis techniques using dynamical systems methods that lead to improved analysis of speech signals and to improvements in speech recognition accuracy. This approach represents the integration of two traditionally distinct research fields: statistical signal processing and chaotic systems. Since signal processing is fundamentally based on linear systems theory and the study of chaos is inherently non-linear, these fields have little or no overlap outside of the fact that both attempt to model the behavior of physical systems. This research integrated these different fields by applying stochastic analysis and modeling tools from the signal processing field to the problem of analyzing embedded phase spaces obtained from chaotic systems analysis of time-series signals.
Automatic Identification of Heart Arrhythmias
Changes in the normal rhythmicity of a human heart may result in different cardiac arrhythmias, which may be immediately fatal or cause irreparable damage to the heart when sustained over long periods of time. The ability to automatically identify arrhythmias from ECG recordings is important for clinical diagnosis and treatment, as well as for understanding the electrophysiological mechanisms of the arrhythmias. This project focused on novel approaches to efficiently and accurately identify normal sinus rhythm and various ventricular arrhythmias through phase space reconstruction and machine learning techniques.
Induction Motor Asynchronous Drive Fault Prediction
Induction motor-drive systems are used throughout industry. In many applications they are the motor-drives of choice. They are used in a wide variety of motor-drive applications such as industrial plant control, propulsion systems, and medical diagnostic equipment. Failure of such motor-drive systems have serious impact on the equipment they are part of including shutdown of the larger system.
Accurate diagnostics and fault prediction increase the reliability of motor-drive systems and minimizes the problem of their failure in the field. A dual track approach was proposed to develop a comprehensive diagnostic system for motor-drive systems. We addressed two fundamental problems in developing such a diagnostic system, namely the difficulty and cost of obtaining large amounts of motor-drive system failure data and the difficulty in developing accurate, robust, fast, and effective diagnostic techniques. The dual track was a combination of time stepping coupled finite element state space techniques for generating accurate, but inexpensive, fault data with a reconstructed phase space based approach for modeling system trajectories for creating effective fault signatures.
Predicting stock market behavior is a challenging and sometimes impossible task. There are many factors that influence the changes in value of stocks, and no one knows exactly how each of these factors will impact the market. This project approached this problem with a combination of proven time series analysis methods and novel evaluation techniques. In the strategies used here, only the past stock price information is used to predict the future behavior of the stock. A model of the stock price time series is learned and then evaluated using a metric that determines the confidence level of the accuracy of the model. Stocks can be traded at times when the model predicts success, and the model is evaluated at a high confidence level. This way, the number of trades, and the commissions that go along with them, can be limited, while still employing a successful market prediction strategy.
Richard J. Povinelli, Ph.D., P.E.
Office: EN 221 (Haggerty Hall)
Tuesdays — 2:00 - 3:45 pm
Thursdays — 3:00 - 3:45 pm
If these times do not work for you, please make an appointment on Outlook 365 and schedule a time that works for both of us.