91Ë¿¹ÏÊÓƵ

Abstract

Anomaly Detection (AD) is a critical yet challenging task due to the scarcity of abnormal samples. Self-Supervised Learning (SSL) offers a promising solution by enabling effective representation learning from abundant normal data. While SSL-based approaches have achieved significant success in image-based AD, their application to other modalities, particularly temporal data, remains relatively underexplored. This thesis investigates the use of SSL for anomaly detection across increasingly complex data settings, progressing from basic temporal data to more complex and multidimensional data types, while addressing the challenges arising from limited anomalous samples.

We begin by exploring SSL in one of its most accessible temporal modalities: acoustic signals. By representing audio as time–frequency images, we apply contrastive learning with audio-specific augmentations to achieve strong performance in anomalous sound detection. This demonstrates that SSL can effectively capture temporal patterns when the signal is mapped to a suitable feature-based representation. Building on this insight, we introduce Deep Autoencoding Support Vector Data Descriptor (DASVDD), a more general, task-agnostic framework that integrates a self-supervised autoencoder with an SVDD constraint. Through evaluations across multiple modalities, we demonstrate the effectiveness of this joint optimization strategy while also revealing the limitations of modality-agnostic models when confronted with complex temporal dependencies.

The need for more specialized modeling of Multivariate Time-Series leads to the introduction of mVSG-VFP, a framework designed for sensor-based vehicle engine monitoring. By leveraging graph-based modeling, mVSG-VFP captures latent dependencies across multiple interdependent sensors. Notably, like its predecessors, this model operates within the representation space to mitigate the impact of sensor noise. This progression culminates in ARTA, an adversarial self-supervised framework designed to operate directly on raw, high-dimensional time-series signals. Unlike previous iterations, ARTA is inherently insensitive to noise, bypassing the need for intermediate feature extraction while ensuring the detector remains robust and interpretable.

Through this trajectory, from simple acoustic models to specialized multivariate systems, this thesis develops a series of increasingly sophisticated SSL frameworks. These contributions demonstrate how addressing the specific structural constraints of temporal data expands the applicability of self-supervised learning, advancing the state of the art in anomaly detection across multiple challenging real-world domains.