Analysis of High-Dimensional Time-Series Data

Abstract

Modern software systems generate large volumes of high-dimensional multivariate time-series data during testing and operation. In industrial settings such as continuous integration and continuous deployment (CI/CD) pipelines, manually diagnosing recurring failures across thousands of performance metrics is time-consuming and error-prone. This thesis investigates whether representation learning can produce stable, interpretable, and diagnostically useful groupings of anomalous behaviour in such data. This allows for recurring behavioural patterns to be identified and related to meaningful system characteristics, thereby supporting failure investigation at scale. We develop and compare several autoencoder architectures, including standard, variational, Long Short-Term Memory-based (LSTM), and Transformer-based variants, against traditional dimensionality reduction baselines and state-of-the-art anomaly detection models. Each architecture encodes multivariate time windows into a compact latent representation, from which reconstruction error is derived as an anomaly score and clustering is subsequently performed. We evaluate these methods across four dimensions: anomaly detection performance, clustering stability under perturbation, interpretability of the resulting cluster structure, and computational cost. Experiments are conducted on three established benchmark datasets: Server Machine Dataset (SMD), Secure Water Treatment System (SWaT), and Soil Moisture Active Passive (SMAP) and an unlabelled industrial dataset from an Ericsson CI/CD testing pipeline. No single model consistently outperforms across all datasets. LSTM-based and probabilistic architectures achieve the highest anomaly detection scores, yet traditional methods retain approximately 90% of peak performance at a fraction of the computational cost. Clustering stability is high on the benchmark datasets, with simpler models frequently matching more complex architectures under perturbation. This means that the models capture meaningful structure instead of noise which is important for downstream analysis. The LSTM encoder combined with Uniform Manifold Approximation and Projection (UMAP) with K-Means clustering produces the most interpretable groupings, with cluster prototypes showing clearly distinct feature-level patterns. In the industrial case study, the approach groups test runs by performance-metric dynamics rather than textual error similarity, offering a complementary perspective for fault diagnosis. Across all tasks, the relationship between model complexity and performance proves non-linear: moving beyond linear projections yields meaningful gains, but further architectural complexity delivers diminishing returns.