Time series anomaly detection is a challenging task and is applicable in many scenarios. Especially in industrial environments, anomaly detection is widely underexplored due to the combination of individual applications and high-frequency operation. The often multivariate nature and the contextual and temporal dependencies of time series data require modern models that can identify novel patterns from historical data. This work primarily explores whether novel architectures utilizing generative models and modern deep learning techniques can reliably detect anomalies in industrial environments in real time. While current deep learning approaches fall short of investigating the model’s performance under varying input window sizes, we believe that a scalable model benefits from broader context windows, leading to better detection results. We design different novel architectures that leverage generative adversarial networks, the transformer architecture, and diffusion models. We focus on reconstruction-based approaches trained solely on nominal data. Anomalous patterns cannot be accurately reconstructed during inference, leading to high anomaly scores. The scoring mechanism is thereby not only restricted to the difference between the reconstructed and the original signal, but can also incorporate other metrics, such as similarity measurements. Benchmark datasets are essential for new methods, as they make different models and configurations comparable. While multiple datasets for multivariate time-series anomaly detection exist, we find that only a few are representative and can be used to evaluate different models with large input windows. We evaluate our and other often-used models on a subset of those datasets and investigate their performance across different window sizes, thereby showing that current approaches lack scalability and that their performance is misleading. The path from anomaly score to detection result is by no means standardized, which impairs the comparability of algorithms. We discuss different thresholding techniques and result calculations in this thesis and compare them using our and various other state-of-the-art detection algorithms on different datasets. In particular, our novel transformer-based architecture demonstrates scalability, low latency, and robust detection performance in our experiments, enabling its use in production environments. The simulated deployment in two distinct real-world industrial use cases demonstrates the efficiency of our approach. Because long-lived machinery rarely exhibits anomalies within a constrained monitoring period, we generate synthetic anomalies and demonstrate our model’s capabilities.     «

Time series anomaly detection is a challenging task and is applicable in many scenarios. Especially in industrial environments, anomaly detection is widely underexplored due to the combination of individual applications and high-frequency operation. The often multivariate nature and the contextual and temporal dependencies of time series data require modern models that can identify novel patterns from historical data. This work primarily explores whether novel architectures utilizing generative...     »