Signal fusion and dimensionality reduction for classification and anomaly detection tasks

Abstract

This thesis explores sensor fusion for feature reduction applied to classification and anomaly detection tasks. We developed new signal fusion techniques applied to tactile sensor signals texture classification and dimensionality reduction for anomaly detection in time series data. The first part of this thesis introduces a novel approach dimensionality reduction of tactile signals for texture classification using principal component analysis (PCA) and reducing exploration time without compromising classifier accuracy. Various pipeline configurations demonstrated that a 3-second exploration combined with PCA-fused features achieved up to 98% classification accuracy with a significant reduction in feature input, yielding a reduction factor of 6750 times. This part of the work highlights PCA's advantage over alternative fusion techniques, offering both interpretability and dimensionality reduction that enhance classifier performance. The second part of this thesis examines anomaly detection within thread line signals, comparing the efficacy of PCA-based fusion, averaging, and raw signal methods. Experimental results across three thread lines indicate that PCA-based fusion provides a balanced sensitivity to anomalies, offering a streamlined process by removing the need for individual signal threshold adjustments. This work contributes to advancements in classification and automated anomaly detection by showcasing the effectiveness of PCA-based signal fusion for feature reduction and robust anomaly detection.