Master of Science, University of Toledo, 2023, Engineering (Computer Science)

Anomaly detection in power grids has become a significant challenge in recent years. The heterogeneous nature and integration of different smart grid appliances make it difficult to detect system faults or energy thefts leading to substantial cumulative losses over long periods. However, implementing the anomaly detection mechanism at every node in the grid network can be costly. Therefore, it is crucial to optimize the anomaly detection technique to not only detect local anomalies but also those further down the network. By doing so, we can ensure minimal resource usage and maximum reliability. In this study, we investigate anomaly detection capabilities at various levels in the distribution power system. We utilize Recurrent Neural Networks (RNNs) for anomaly detection and evaluate their performance compared to other machine learning techniques. The power grid analyzed in this research is a Digital Twin - a digital replica of a real-world power grid modeled using Gridlab-D and Helics. To ensure accurate simulation behavior and simulate with real consumption data and voltage properties. This paper presents two aspects of the study: building the Digital Twin and conducting anomaly detection in the Digital Twin Simulation at various grid levels.

Committee: Ahmad Y Javaid (Committee Chair); Weiqing Sun (Committee Co-Chair); Raghav Khanna (Committee Member) Subjects: Computer Engineering; Electrical Engineering

Master of Science, The Ohio State University, 2022, Mechanical Engineering

Truck drivers idle their trucks for their comfort in the Cab. They might need air conditioning to maintain a comfortable temperature and use the onboard appliances like TV, radio, etc. while they rest during their long journeys. On average idling requires 0.8 gallons of diesel per hour for an engine and up to 0.5 gallons per hour for a diesel APU. For a journey greater than 500 miles, a driver rests for 10 hours for every 11 hours of driving. Drivers tend to leave the truck idling throughout the 10 hours. With today's cost of diesel in the US, for one 10-hour period, the average cost incurred by the owner only on idling is $32. About a million truck drivers idle their trucks overnight for more than 300 days a year. Super Truck II is a 48V mild hybrid class 8 truck with all auxiliary loads powered purely by the battery pack. This offers an opportunity to reduce the idling from the whole 10 hours to whatever is necessary to charge the battery enough to power the auxiliaries. To quantify this “necessary idling” during the hoteling period we need to predict what the power load requirement in the future would be. The total power estimation is divided into two portions, (1) Cabin Hotel loads except HVAC and (2) HVAC load. A physics-based grey box models are developed for components in the vapor compression cycle and cabin using system dynamics which is used to estimate the HVAC power consumption. A special kind of Recurrent Neural Network (RNN) called Long, and Short Term Memory (LSTM) is used to predict the cabin hotel loads by user activity tracking. Synthetic load profiles are synthesized to overcome the limitation of lack of availability of data, about the user activity inside the cabin for training the LSTM algorithm, using rules and observations derived from the existing load profile for the hotel period from a survey conducted for SuperTruck project and literature survey on driver sleeping behavior. Dynamic Time Warping along with pointwise Euclidian distance is us (open full item for complete abstract)

Committee: Qadeer Ahmed Dr (Advisor); Marcello Canova Dr (Committee Member); Athar Hanif Dr (Other) Subjects: Artificial Intelligence; Automotive Engineering; Engineering; Mechanical Engineering; Statistics; Sustainability; Systems Design; Transportation

Master of Science, The Ohio State University, 2022, Computer Science and Engineering

Natural Language Processing techniques are being widely used in the industry these days to solve a variety of business problems. In this work, we experiment with the application of NLP techniques for the use case of understanding the call interactions between customers and customer service representatives and to extract interesting insights from these conversations. We focus our methodologies on understanding call transcripts of these interactions which fall under the category of long document understanding. Existing works in text encoding typically address short form text encoding. Deep Learning models like Vanilla Transformer, BERT and DistilBERT have achieved state of the art performance on a variety of tasks involving short form text but perform poorly on long documents. To address this issue, modifications to the Transformer model have been released in the form of Longformer and BigBird. However, all these models require heavy computational resources which are often unavailable in small scale companies that run on budget constraints. To address these concerns, we survey a variety of explainable and light weight text encoders that can be trained easily in a resource constrained environment. We also propose Hierarchical Self Attention based models that outperform DistilBERT, Doc2Vec and single layer self-attention networks for downstream tasks like text classification. The proposed architecture has been put into production at the local industry organization that sponsored the research (SafeAuto Inc.) and helps the company to monitor the performance of its customer service representatives.

Committee: Eric Fosler-Lussier (Committee Chair); Rajiv Ramnath (Advisor) Subjects: Artificial Intelligence; Computer Science

Bachelor of Science (BS), Ohio University, 2021, Computer Science

There is a strong desire to be able to more easily formalize existing mathematical statements and develop machine-checked proofs to verify their validity. Doing this by hand can be a painstaking process with a steep learning curve. In this paper, we propose a model that could automatically parse natural language proofs written in LaTeX into the language of the interactive theorem prover, Coq, using a recurrent neural network. We aim to show the ability for such a model to work well within a very limited domain of proofs about even and odd expressions and exhibit generalization at test time. We demonstrate the model's ability to generalize well given small variations in natural language and even demonstrate early promising results for the model to generalize to expressions of intermediate lengths unseen at training time.

Committee: David Juedes (Advisor) Subjects: Computer Science

MS, University of Cincinnati, 2021, Engineering and Applied Science: Computer Science

Modeling of bacteria growth under different environmental conditions provides a useful tool to predict food and consumer goods safety. This study introduces a flexible, unique, and data-driven model to predict the bacteria growth under different pH conditions, using a one-to-many Long-Short-Term Memory (LSTM) model. When compared with a benchmark model the proposed model showed a good predictive power for different bacteria behaviors. In addition to its predictive ability, the model architecture is flexible and can be adapted for different bacteria behavior patterns without additional prior assumptions.

Committee: Anca Ralescu Ph.D. (Committee Chair); Kenneth Berman Ph.D. (Committee Member); Mark Maupin Ph.D. (Committee Member); Dan Ralescu Ph.D. (Committee Member) Subjects: Computer Science

PhD, University of Cincinnati, 2020, Engineering and Applied Science: Aerospace Engineering

This thesis is a preliminary investigation into using artificial neural networks to generate surgical trajectories towards space-based surgery in non-near earth environments. The first study focused on the communication-delay associated with tele-operation. This study was performed to find the threshold of the communication delay where tele-surgery would not be feasible. This is important because if humans are going to travel to Mars, semi-autonomous surgical methods must be developed as an alternative means of providing surgical intervention. We found a delay of 1.5 seconds was about the threshold where tele-operation becomes excessively difficult. An IRB study was conducted using surgeons and surgical residents who performed virtual surgical procedures in a virtual dissection simulator to acquire training data to use to train LSTM-RNNs. The data obtained from that study was reduced, because the sampling rate of the simulator was too high and the datasets contained too many data points to train the LSTM-RNNs effectively. The procedure was segmented into three subtasks. Several LSTM-RNN were trained using a custom cost function and evaluated using custom metrics. This method was compared with another algorithm, which was used to generate surgical trajectories. We tested the LSTM-RNN several shifted set of fiducial markers to assess its robustness. We found the LSTM-RNNs were robust enough to handle slight changes in anatomy.

Committee: Catharine McGhan Ph.D. (Committee Chair); Kelly Cohen Ph.D. (Committee Member); Ou Ma Ph.D. (Committee Member); Grant Schaffner Ph.D. (Committee Member) Subjects: Aerospace Materials

Master of Science (MS), Ohio University, 2018, Electrical Engineering & Computer Science (Engineering and Technology)

The main objective of this research is to assess the capability, applicability and versatility of Recurrent Neural Networks for the task of blood glucose prediction. We first experimented with Long-Short Term Memory (LSTM) networks on a real-patient dataset that has been used in previous research at the SmartHealth lab, and evaluated their performance compared to Support Vector Regression (SVR) models that use elaborate hand-engineered features extracted from raw data, physiological models and Autoregressive Integrated Moving Average (ARIMA) models. The results confirm that a simple LSTM network with only 5 nodes and one hidden layer is able to achieve similar or better performance versus the SVR, on both the time horizons of 30 and 60 minutes. The obtained results are also better than those of a T0 baseline, in which the blood glucose level (BGL) is assumed to stay unchanged, ARIMA, and human experts. We then evaluated a larger model of 20 nodes on a synthetic dataset called UVA/Padova, which is one of the most realistic simulators for Type I diabetes. Our UVA dataset constitutes 900 days of data from 10 simulated patients. The obtained results confirmed that LSTM networks are much more capable than ARIMA with exogenous variable (ARIMAX) in terms of incorporating sparse signals such as meal and insulin. We also show that only a few architectural changes are needed to adapt LSTM networks between different datasets. Dropout is shown to be one important factor, along with the early-stop threshold. The last set of experiments were performed on our most recent and best real-patient dataset called OUBasis. The results exhibit previous trends, with LSTMs comfortably beating the baseline T0 and ARIMAX models. We also evaluate the versatility of LSTM networks for seamless incorporation of raw signals obtained from physiological sensors, with results showing that such signals do help performance. Almost no modification was required to the network when new features ar (open full item for complete abstract)

Committee: Razvan Bunescu (Advisor) Subjects: Computer Science