Comparison of feature-based classifiers in Automatic Modulation Classification systems
In South Africa, the telecommunications regulator, ICASA, follows a ?xed spectrum allocation approach. The radio spectrum is assigned to incumbents (primary users) on a long-term basis. This practice follows a wholesale approach auctioning spectrum to the highest bidder. This approach leads to under-utilised spectrum which causes an arti?cial spectrum scarcity. Two spectrum allocation processes have been proposed in literature: dynamic spectrum market model, and a spectrum commons model. Cogni-tive radio (CR) is an enabling technology for either of these models. Spectrum sensing is a key element and should be performed ?rst before allowing sec-ondary user access. Energy detection, cyclostationary feature detection, matched ?l-tering and cooperative sensing has been proposed as spectrum sensing techniques. An Automatic Modulation Classi?cation (AMC) system detects the unknown modu-lation type of a received signal in preparation to demodulate the signal and retrieve its information content. AMC plays an important role in military and civilian applica-tions such as signal con?rmation, interference identi?cation, surveillance, monitoring, spectrum management, counter channel jamming and signal intelligence. Future Software-de?ned Radio (SDR) and CR systems must be able to sense the spec-trum for signals present in the pursuit of enabling Dynamic Spectrum Access (DSA). This interest in increasing spectrum access and improving spectrum ef?ciency, com-bined with SDR and new realisations that machine learning can be applied to radios, have created interesting possibilities, such as CR. The International Telecommunications Union for radio communications (ITU-R) gives guidelines for the technical identi?cation of digital signals. The signal's spectral form, frequency, bandwidth, instantaneous amplitude and phase can be used for this pur-pose. For any regulator, it is important to monitor signals of interest and to identify them accordingly. Doing this involves traditional software packages which follow a brute-force approach in demodulating the signals. Each demodulation approach gets tested against the signal of interest until a match is found. Modern digital signals are modulated using a variety of modulation techniques. In this dissertation, an investigative study is presented towards ?nding a simple approach to identifying and classifying M-PSK and M-QAM signals in the UHF fre-quency band. Two approaches can be followed when deciding on a classi?cation ap-proach: Likelihood-based (LB) approach or a Feature-based (FB) approach. A FB (also known as a pattern-recognition) classi?cation approach is followed in the dissertation. A combination of instantaneous time-domain and higher-order statistical features are extracted from the signal's instantaneous amplitude and phase. A Support Vector Ma-chine (SVM) is used to solve the classi?cation problem. The performance of the AMC is tested in an Additive White Gaussian Noise (AWGN) and multipath fading channel. Two use-cases for evaluating the performance of the classi?er is presented: with and without Signal-to-Noise Ratio (SNR) estimation. In-troducing SNR estimation as part of the feature set increased the classi?cation accu-racy for Quadrature Phase Shift Keying (QPSK) and 8-Phase Shift Keying (PSK) sig-nals at low SNR. A 2% classi?cation accuracy improvement was obtained at 4 dB for QPSK signals, while a 12% classi?cation accuracy improvement for 8-PSK signals was obtained for an SNR of 1 dB. Furthermore, the performance of the proposed classi?er was assessed for two multi-path channel conditions: for a stationary transmitter and receiver, and secondly for a moving receiver. Four randomly selected Doppler shifts were chosen and evaluated. An overall classi?cation accuracy of 90% was reported for the stationary case, while the accuracy of the different Doppler shifts were 85%, 86%, 77.5% and 78% respectively. Finally, the performance of the classi?er was evaluated using recorded In-phase and quadrature (I/Q) data of a TETRA signal. The proposed classi?er correctly identi?ed the TETRA signal to be part of the PSK modulation group. However, the classi?er was not able to determine the modulation order. The research done in the dissertation showed that following a simple FB classi?cation approach to classifying digital signals is possible. The work showed that higher-order statistics extracted form the instantaneous amplitude and phase of the received signal can be used as features.
- Engineering