M.Sc. Outi Savolainen defends her PhD thesis “Anomaly Detection and Predictive Uncertainty Estimation with Applications to GNSS Signals” on Friday the 13th of November 2026 at 13 in the University of Helsinki Exactum Building, Auditorium CK112 (Pietari Kalmin katu 5, basement. Her opponent is Professor Jari Nurmi (Tampere University) and custos Professor Laura Ruotsalainen (University of Helsinki). The defence will be held in Finnish.
The thesis of Outi Savolainen is a part of research done in the Department of Computer Science and in the Spatiotemporal Data Analysis group at the University of Helsinki. Her supervisors have been Professor Laura Ruotsalainen (University of Helsinki) and Professor Simo Särkkä (Aalto University).
Anomaly Detection and Predictive Uncertainty Estimation with Applications to GNSS Signals
Many critical systems utilize machine learning for anomaly and similarity detection. However, to obtain reliable and robust results, the decision should be based on trust in the model's prediction. The data collected in real-world environments are often noisy and can contain characteristics added by the recording equipment deteriorating the model's performance. Often, the training data are imbalanced, meaning that some phenomena are rarely present while some occur more frequently. The existing training data are usually limited, and the representation of the data can be high-dimensional or complex-valued by nature. This dissertation specifically addresses key challenges related to data representation, uncertainty quantification, robustness, threshold determination, and generalization in critical real-world environments.
This doctoral dissertation develops novel machine learning methodologies for anomaly and similarity detection from challenging real-world data that are often high-dimensional, noisy, limited, or complex-valued. The models are evaluated with applications particularly demonstrated in radio frequency (RF) signal analysis and global navigation satellite system (GNSS) interference detection.
The first contribution is a deep learning method for complex-valued anomaly detection. By operating in the complex domain, the proposed method can exploit both magnitude and phase information present in the data. When the data is limited and all possible classes are not known beforehand, training with clean data samples, i.e., not including anomalies, is beneficial. However, clean data can have natural changes over time that may not be captured by the training data. A deep learning model may overfit the training data, and if such normal fluctuations are not considered during threshold determination, the model's performance may degrade when applied to real-time monitoring systems.
Our second contribution addresses the challenges of threshold determination and real-time monitoring in detection and classification using high-dimensional data. Often, the number of anomalous events is smaller compared to the normal events. Classifying all collected samples in the real-time monitoring systems can become excessive, especially if heavy data pre-processing is applied. Therefore, computational resources can be reduced by detecting possible anomalies and only classify these filtered samples.
Our third contribution addresses the challenges in predictive uncertainty estimation. We first analyze the classification performance of probabilistic and non-probabilistic networks using data collected from different environments where the sample classes are partially similar to each other. The observations are exploited in our novel model, which adapts to unseen data classes and predicts uncertainty in similarity detection. Specifically, we propose a Siamese network with an uncertainty estimation block for computing a similarity metric. Moreover, the external probabilistic block ensures that probabilities are not directly learned from the training data, increasing the model's adaptivity to unseen data classes.
Availability of the dissertation
An electronic version of the doctoral dissertation will be available in the University of Helsinki open repository Helda at
Printed copies will be available on request from Outi Savolainen: