Accurate detection and localization of anomalies in multivariate time series is essential for maintaining the safety and stability of modern industrial systems. Despite the critical need for timely anomaly detection, most existing methods rely on fixe...
Accurate detection and localization of anomalies in multivariate time series is essential for maintaining the safety and stability of modern industrial systems. Despite the critical need for timely anomaly detection, most existing methods rely on fixed-size sliding windows, which limit their ability to capture long-range temporal dependencies and cause positional bias—making precise localization challenging. This thesis proposes a novel framework that addresses these limitations through (1) a 2D structural representation that maps the entire multivari- ate time series into an image-like format, enabling convolutional neural net- works (CNNs) to jointly model local and global temporal relationships, and (2) a multi-scale, patch-wise scoring mechanism that evaluates each time point through overlapping receptive fields to produce stable, context-aware anomaly scores. A new padding strategy, Time Series Image Padding (TSIP), is also introduced to preserve temporal continuity and mitigate boundary artifacts during the 2D transformation. The proposed design inherently enhances robustness to sensor noise, as the 2D CNN dilutes transient perturbations while multi-patch aggregation further stabilizes the anomaly scores. Comprehensive experiments on diverse bench- mark datasets show that the proposed approach outperforms recent baselines, achieving state-of-the-art detection accuracy and noise robustness, while also exhibiting promising diagnostic capability. While primarily developed for anomaly detection, the framework also ex- tends naturally to anomaly diagnosis, a critical downstream task. In practical settings, identifying when an abnormal event occurs is only the first step—the ultimate goal is to determine which sensor or channel is responsible for the failure so that operators can take immediate and targeted action. Without any architectural modifications, the model leverages patch-wise evaluations across channels to highlight which sensors contribute most to abnormal behavior, demonstrating strong diagnostic capability and showing that the proposed rep- resentation and scoring mechanisms have substantial potential for channel-level anomaly analysis in real-world applications.