Multivariate anomaly detection with LSTM layered Variational Autoencoder

Abstract

The aim of this thesis was to develop and evaluate the effectiveness of a recurrent neural network layered autoencoder model for detecting anomalies in multivariate time-series data, with a focus on improving the accuracy and reliability of diagnostic data for Volvo Penta’s boats. The primary goal was to leverage the relationships and correlations between signals to identify deviations that traditional models may fail to detect. The model’s performance was assessed in terms of its ability to learn the structure of normal data, detect synthetic anomalies, and provide meaningful insights without relying on labeled datasets. The study highlights the limitations of traditional evaluation metrics, which are often unsuitable for unsupervised learning approaches like the model used. Instead, the model’s effectiveness was demonstrated through reconstruction error analysis and its ability to handle the complexities of multivariate time-series data. Challenges such as data dimensionality, sequence length optimization, and noise handling were addressed to enhance the model’s robustness. The findings suggest that while the model excels at identifying synthetic anomalies and capturing temporal relationships, further work is needed to generalize its capabilities to real-world scenarios. This research lays the groundwork for improving diagnostic processes and supports the development of more adaptive and reliable anomaly detection systems.