Machine Learning Based Prediction of Frequency Hopping Spread Spectrum Signals

In an world shifting towards wireless communications, the already scarce electromagnetic spectrum within the unlicensed bands is becoming increasingly crowded. All wireless devices operating in those bands need to co-exist without interfering with each other. Frequency hopping spread spectrum (FHSS) is a communication technique especially resilient to interference due to its constant change of the carrier frequency and its narrowband transmission bandwidth. Furthermore, it produces minimal interference to other signals in the same frequency band using wider bandwidth. However, interference can also be harmful even for FHSS transmissions as a result of the loaded ISM bands. Intelligent spectrum sensing techniques can contribute to a more efficient spectral usage. In this paper, we propose a supervised learning algorithm which predicts the future time-frequency location of a FHSS signal. We design a convolutional neural network which is trained on a dataset, obtained from measurements of two FHSS sources. Based only on a small observation window of 50 ms, it predicts the signal appearance of the following 25 ms in a time-frequency representation. To show that we can accurately predict the signal, we introduce a special score measure. The mean score of about 0.9 with small standard deviation demonstrates the high fidelity prediction of the signal's evolution.