https://www.mdu.se/

Modulation type classification is a part of waveform estimation required to employ spectrum sharing scenarios like dynamic spectrum access that allow more efficient spectrum utilization. In this work multiple classification features, feature extraction, and classification algorithms for modulation type classification have been studied and compared in terms of classification speed and accuracy to suggest the optimal algorithm for deployment on our target application hardware. The training and validation of the machine learning classifiers have been performed using artificial data. The possibility to use instantaneous values of the time domain signal has shown acceptable performance for the binary classification between BPSK and 2FSK: Both ensemble boosted trees with 30 decision trees learners trained using AdaBoost sampling and fine decision trees have shown optimal performance in terms of both an average classification accuracy (86.3 % and 86.0 %) and classification speed (120 0000 objects per second) for additive white gaussian noise (AWGN) channel with signal-to-noise ratio (SNR) ranging between 1 and 30 dB. However, for the classification between five modulation classes demonstrated average classification accuracy has reached only 78.1 % in validation. Statistical features: Mean, Standard Deviation, Kurtosis, Skewness, Median Absolute Deviation, Root-Mean-Square level, Zero Crossing Rate, Interquartile Range and 75th Percentile derived from the wavelet transform of the received signal observed during 100 and 500 microseconds were studied using fractional factorial design to determine the features with the highest effect on the response variables: classification accuracy and speed for the additive white gaussian noise and Rician line of sight multipath channel. The highest classification speed of 170 000 objects/second and 100 % classification accuracy has been demonstrated by fine decision trees using as an input Kurtosis derived from the wavelet coefficients derived from signal observed during 100 microseconds for AWGN channel. For the line of sight fading Rician channel with AWGN demonstrated classification speed is slower 130 000 objects/s.

This paper is focused on the blind symbol rate estimation for the digital FSK modulated signals, based on the classification between three symbol rate classes: 10, 100, and 1000 KSymbol/second using the scalogram images obtained from continuous wavelet transform with Morse wavelet. Pretrained deep learning AlexNet has been transfer learned to classify between symbol rate classes. Training, testing, and validation data sets have been composed of the artificial data generated using Bernoulli binary random signal generator modulated into FSK signal corrupted by additive white Gaussian noise (AWGN) noise with SNR ranging from 1 to 30 dB. Training and validation data sets have been augmented to obtain twice more extensive data set i.e 1800 scalogram images, compared to the original size of 900 samples. The average classification accuracy during validation has reached 99.7 % and during testing 100 % and 96.3 % for the data sets with SNR 25–30 dB and 20–25 dB respectively. The proposed algorithm has been compared with cyclostationary and has shown improved classification accuracy, especially in conditions of low SNR.

This paper is focused on the blind vacant frequency channels detection in 56 MHz observation band divided into 56 channels (1 MHz each) implemented on in-house developed hardware based on the AD9364 transceiver operating in automatic gain control (AGC) mode. Vacant channel detection has been modeled as a binary hypothesis testing problem. Three signal detection algorithms widely used in the literature including energy detection, wavelets, and cyclostationary have been tested and evaluated for potential use in our target application. Primary, offline testing has been performed in the Matlab environment using the data samples captured on the target receiver's front end as a time-domain complex signal. Data samples containing one, two, or three signals generated by hardware signal generator and modulated into 2FSK, BPSK, or QPSK with symbol rate 10, 100, or 1000 kSymbol/s. The highest accuracy of 91.0 % has been observed in the offline detection for continuous wavelet transform, while energy detection has demonstrated 86.4 % accuracy. Cyclostationary detection has shown no distinguishable difference in the spectrum correlation values calculated for the AWGN noise sample and the sample containing BPSK and 2FSK modulated signals. Energy detection and discrete wavelet transform have been implemented on our target hardware and tested in the office environment in conditions that could be approximated by AWGN channel. Test sequences containing one or two signals have been generated by the signal generator and received and processed by our target radio node. Discrete wavelet transform has demonstrated 85.73 % and energy detection 85.25 % accuracy in real-time testing. © 2022 IEEE.

Modulation type classification is a part of waveform estimation required to employ spectrum sharing scenarios like dynamic spectrum access that allow more efficient spectrum utilization. In this work multiple classification features, feature extraction, and classification algorithms for modulation type classification have been studied and compared in terms of classification speed and accuracy to suggest the optimal algorithm for deployment on our target application hardware. The training and validation of the machine learning classifiers have been performed using artificial data. The possibility to use instantaneous values of the time domain signal has shown acceptable performance for the binary classification between BPSK and 2FSK: Both ensemble boosted trees with 30 decision trees learners trained using AdaBoost sampling and fine decision trees have shown optimal performance in terms of both an average classification accuracy (86.3 % and 86.0 %) and classification speed (120 0000 objects per second) for additive white gaussian noise (AWGN) channel with signal-to-noise ratio (SNR) ranging between 1 and 30 dB. However, for the classification between five modulation classes demonstrated average classification accuracy has reached only 78.1 % in validation. Statistical features: Mean, Standard Deviation, Kurtosis, Skewness, Median Absolute Deviation, Root-Mean-Square level, Zero Crossing Rate, Interquartile Range and 75th Percentile derived from the wavelet transform of the received signal observed during 100 and 500 microseconds were studied using fractional factorial design to determine the features with the highest effect on the response variables: classification accuracy and speed for the additive white gaussian noise and Rician line of sight multipath channel. The highest classification speed of 170 000 objects/second and 100 % classification accuracy has been demonstrated by fine decision trees using as an input Kurtosis derived from the wavelet coefficients derived from signal observed during 100 microseconds for AWGN channel. For the line of sight fading Rician channel with AWGN demonstrated classification speed is slower 130 000 objects/s.