Multivariate anomaly detection with LSTM layered Variational Autoencoder

Sammanfattning

The aim of this thesis was to develop and evaluate the effectiveness of a recurrent neuralnetwork layered autoencoder model for detecting anomalies in multivariate time-seriesdata, with a focus on improving the accuracy and reliability of diagnostic data for VolvoPenta’s boats. The primary goal was to leverage the relationships and correlations betweensignals to identify deviations that traditional models may fail to detect. Themodel’s performance was assessed in terms of its ability to learn the structure of normaldata, detect synthetic anomalies, and provide meaningful insights without relying on labeleddatasets.The study highlights the limitations of traditional evaluation metrics, which are often unsuitablefor unsupervised learning approaches like the model used. Instead, the model’seffectiveness was demonstrated through reconstruction error analysis and its ability tohandle the complexities of multivariate time-series data. Challenges such as data dimensionality,sequence length optimization, and noise handling were addressed to enhancethe model’s robustness. The findings suggest that while the model excels at identifyingsynthetic anomalies and capturing temporal relationships, further work is neededto generalize its capabilities to real-world scenarios. This research lays the groundworkfor improving diagnostic processes and supports the development of more adaptive andreliable anomaly detection systems.