GNSS Anomaly Detection with Complex-Valued LSTM Networks

Global Navigation Satellite Systems (GNSS) serve many critical systems. Unfortunately, the GNSS based services are threatened by interference causing anomalies to the acquired signals. To protect the critical infrastructure, navigation signal quality should be monitored, anomalies immediately detected, isolated, and back-up solutions used. Previous GNSS anomaly detectors concentrate on one interference type only. Although methods based on deep learning are emerging, most work use convolutional neural networks, which are transcendent in processing spatially correlated data, such as images. However, GNSS data has temporal correlation, which requires suitable models such as Long Short-Term Memory (LSTM) networks. Traditionally, deep learning models have been trained using supervised methods requiring laborious labelling and therefore slowing down the modelling of complicated real-world phenom-ena. This paper presents, as far as we know, the first unsupervised LSTM based autoencoder for GNSS anomaly detection. LSTM autoencoders used in other domains process data in real or semi-complex domains and we claim that processing the signal at fully complex domain will improve the detection. Thereby, we present here the first fully complex-valued detector and test it with both real and complex-valued GNSS data. Our model in the real domain provides results that are comparable with the equivalent supervised method's 95% accuracy, outperforming 92% with our complex domain model. We claim that this lower performance is due to the implementation challenges which will be carefully discussed to accelerate the future research.