Unsupervised Time Series Pattern Recognition for Purpose of Electronic Surveillance

conference paper

English

Signal classification is one of the main tasks of electronic surveillance. This paper focuses on extracting patterns from time series and testing robustness of pre-trained Neural Network (NN). A dataset of 10 different time series was created and used to train a neural network based on the Time-Series Representation Learning via Temporal and Contextual Contrasting (TS-TCC) model. The logits layer of the NN model was removed from this pre-trained model to obtain the feature vectors. A dataset containing 87 real signals acquired from passive surveillance sensors was passed to the NN to obtain embeddings that represent the features of the signals extracted from the NN. The dataset was then corrupted with missing pulses and spurious pulses and tested on pre-trained NN. This unsupervised learning method was able to recognize 76% of the signals even with 50% of the missing input data. The research showed that an important step to improve NN performance is to choose suitable data scaling method. The best results were achieved using the StandardScaler from scikit-learn preprocessing library.

pattern recognition, surveillance, neural network, pulse repetition interval

HORKÝ, P.; PROKEŠ, A.; HUBÁČEK, P.

12. 9. 2022

Institute of Electrical and Electronics Engineers Inc.

Polsko

978-83-956020-3-0

Proceedings of MIKON 2022

1

5

5

https://mrw2022.org/mikon/