Human Activity Recognition and Identification Driven Automated Deep Learning for Time-Series Classification

The growing emphasis on Digital Engineering (DE) within the U.S. Department of Defense (DoD) demands advanced methods for leveraging vast time-series data generated by sensor-rich environments. Deep learning models offer promising solutions for complex timeseries classification tasks, however their design and optimization remain highly resource intensive, requiring specialized expertise. This dissertation addresses this challenge by developing and evaluating an Automated Machine Learning (AutoML) framework specifically tailored for the time-series classification task of Human Activity Recognition and Identification (HARI). A systematic investigation was conducted using the Design Science Research Methodology (DSRM) comparing traditional search strategies of grid search and random search with advanced automated Neural Architecture Search (NAS) strategies employing Optuna and a novel configurable gamma strategy. Over 800,00 one-dimensional Convolutional Neural Network (1D-CNN) architectures were generated, trained, and evaluated based on model performance for participant identification and activity recognition using the MotionSense dataset. The results demonstrate that automated search methods outperform traditional approaches, consistently yielding higher F1-scores and more efficient exploration of the architectural search space. The study’s findings highlight the impact of neural architecture on model performance and demonstrate the feasibility of AutoML frameworks for time-series classification tasks in secure and resource-constrained environments, addressing limitations of current open-source and cloud-based AutoML solutions. The contributions advance the state of the art for AutoML for time-series data and supports the DoD’s DE transformation by enabling broader access to high performing deep learning models without requiring deep domain expertise.