Data-Driven Machine Learning and Deep Learning Approaches for Smart Sensor Systems and Battery State of Charge Estimation

This dissertation investigates scalable machine learning (ML) and deep learning(DL) frameworks for interpreting complex sensor signals across multiple domains. The research bridges the gap between raw signal acquisition and actionable insights by focusing on computational methods—rather than hardware innovations—to improve preprocessing, modeling, and classification of multimodal data. Three real-world sensing applications are explored. First, wearable spectral color sensing is demonstrated using the SENSIPATCH device, which integrates LED-based spectroscopy and photodiodes for robust color classification. Through classical ML models and deep neural networks, the system achieves stable performance under variable lighting and contact conditions, with implications for biomedical, agricultural, and industrial applications. Second, water pollutant detection is addressed using the Smart Cable Water (SCW) platform, which applies electrochemical impedance spectroscopy (EIS) coupled with novel deep learning architectures, including a Residual Channel and Spatial Attention Network (RCSANet). This approach achieves over 94% accuracy across multiple pollutants, offering a scalable and cost-effective solution for real-time water quality monitoring. Third, battery state of charge (SoC) estimation is tackled using laboratory-grade EIS data, transformed into image representations via Gramian Angular Fields (GAF). Convolutional neural networks (CNNs) and AlexNet are employed to classify SoC levels, supported by a newly released public dataset from 11 lithium-iron-phosphate cells. The thesis makes several contributions: (1) development of robust preprocessing and normalization strategies for diverse sensing modalities, (2) creation of a publicly available EIS dataset for SoC estimation, (3) validation of ML/DL models across optical, electrochemical, and impedance domains, and (4) demonstration of scalable frameworks that generalize across wearable, embedded, and laboratory platforms. Collectively, these contributions advance data-driven sensing, enabling more reliable, interpretable, and deployable solutions for environmental monitoring, energy storage, and wearable health technologies.