Real-time anomaly detection in time series using supervised deep learning techniques
This thesis presents a hybrid anomaly detection system for operational time series data from ITnow, designed to enhance the performance of a baseline self-supervised model. The methodology employs a self-supervised Convolutional Neural Network (CNN) to generate pseudo-labels, which, after being enri...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/451054 |
| Acceso en línea: | https://hdl.handle.net/2117/451054 |
| Access Level: | acceso abierto |
| Palabra clave: | Neural networks (Computer science) Time-series analysis Anomaly detection Deep learning Time series analysis Xarxes neuronals (Informàtica) Sèries temporals--Anàlisi Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic |
| Sumario: | This thesis presents a hybrid anomaly detection system for operational time series data from ITnow, designed to enhance the performance of a baseline self-supervised model. The methodology employs a self-supervised Convolutional Neural Network (CNN) to generate pseudo-labels, which, after being enriched through data augmentation, are used to train various supervised architectures, including LSTM, TCN, and Transformer models, among others. The evaluation demonstrates the hybrid method's superiority, consistently outperforming the self-supervised baseline. The LSTM and TCN architectures achieved the most robust performance, with average F1-scores nearing 0.90 across all signals, significantly surpassing the baseline's average of 0.84. Despite varying training complexities, all models proved viable for real-time deployment due to their rapid inference speeds. |
|---|