Wei Su and John Kosinski U.S. ARMY CECOM RDEC Intelligence and Information Warfare Directorate Fort Monmouth, New Jersey

1. COMPARISON AND SIMULATION OF DIGITAL MODULATION RECOGNITION ALGORITHMS Wei Su and John Kosinski U.S. ARMY CECOM RDEC Intelligence and Information Warfare Directorate Fort Monmouth, New Jersey

2. Automated Signal Exploitation and Modulation Recognition signal Demodulated signal Cooperative Communication modulation type pulse shaping symbol rate center frequency signal Non-cooperative Communication ? ? ? ? Modulation recognition Signal processing Statistical estimation

3. Factors Affect the Modulation Recognition Results • Residual carrier frequency estimation • Carrier frequency and phase tracking • Baud rate estimation and pulse timing • Pulse shaping recovery • Channel distortion

4. Center frequency offset A PSK8 signal PSK8 signal with phase shift produced by center frequency offset

5. Pulse Timing • Re-sampling will be necessary if the symbol frequency divided by sample frequency is not an integer • Error is introduced in re-sampling 4 samples per symbol 10/3 samples per symbol Before re-sampling After re-sampling

6. Transmitter Radio Channel RF Receiver Equalizer Decision Maker Channel Distortion Before 16-QAM After An adaptive filter (lower figure) is used to eliminate the multiple channel distortion in wireless communication. A distorted 16-QAM signal (left figure) is blindly equalized to give a correct output (right figure) before recognition process

7. Literature Search • A web site search of “modulation recognition” gave 52,500 hits. • More than a hundred publications on modulation recognition. • Many GOTS and COTS products available for various applications.

8. Some Well-known Approaches Modulation feature extraction • I-Q Analysis • Zero-crossing • Power-law Modulation classification • Pattern recognition • Maximum likelihood

9. Problems in Algorithm Comparison • Deferent modulation type • Deferent assumption • Deferent signal environment • Deferent specification MSK FSK8 CW FSK4 16QAM FSK FSK2 32QAM PSK 64QAM PSK8 PSK4 UNKNOWN PSK PSK2 ASK8 ASK4 ASK2

10. thresholds CW PSK2PSK4ASK2ASK4FSK2FSK4 frequency phase amplitude recognition tree Base Band Signal s.t.d. abs Azzouz and Nandi’s Method Based on signal parameter variances and variance thresholds

11. Phase Variance is affected by the Center Frequency Offset phase tracking unwrapped phase tracking BPSK with residual carrier after phase tracking and compensation

12. Phase Variance is Affected by Pulse Shaping (p/4 DQPSK)

13. Phase Variance is Affected by Phase Wrap

14. Limitation in Azzouz and Nandi • Not robust to center frequency offset • Not robust to pulse shaping • Not robust to phase wrap • Thresholds depend on SNR level • Restricted in modulation types (2 bits)

15. baud rate detector BPF timing circuit templates Freq. s.t.d IF I CW PSK2 PSK4 PSK8 ASK2 FSK2 delta phase / tan-1 histograms recognition tree Q ( )2 sqrt + ( )2 amplitude Liedtke’s Method Based on delta-phase and histogram correlation

16. Limitation in Liedtke’s Method • Manual tuning of center frequency • Manual tuning of the time-recovery band pass filter • Limited in modulation types

17. frequency estimation s.t.d. PSK2 PSK4 PSK8 FSK2 FSK4 FSK8 recognition tree phase estimation delta-phase histogram IF zero-crossing counter histogram templates templates Zero-crossing Detection Method Based on zero-crossing phase, and delta phase histogram

18. Limitation in Zero-Crossing Method • Not robust to multiple frequencies • Need high SNR • Need high sampling rate

19. peak detection ( )2 FFT signal PSK2 PSK4 ASK2 FSK2 recognition tree ( )4 envelope s.t.d. Power-Law Classification Method Based on power spectrum analysis

20. Limitation to Power-Law Method • Need very high sampling rate • Affected by pulse shape • Limited in modulation types • Noise is amplified with high order

21. frequency frequency estimation s.t.d. abs phasex2 s.t.d. I abs X2 phase CW PSK2PSK4PSK8ASK2ASK4FSK2FSK4 track and correct phase / tan-1 s.t.d. down demodulation IF recognition tree abs ( )2 amplitude + sqrt s.t.d. timing recovery abs Q ( )2 adaptive thresholds Modified Azzouz and Nandi

22. Center Frequency Offset Correction Method 1: Estimated phase symbol x The unwanted phase rotation (left figure) is corrected adaptively (right figure) by a modified blind carrier phase tracker (upper figure) so that the feature variation methods can be used robustly for modulation recognition Before BPSK After Method 2: CFO = angle( S(yky*k-1)fs/2/p k = 1, 2, …, N yk is a sample at k fs is the sampling frequency

23. frequency frequency estimation templates peak detector timing circuit CW PSK2PSK4PSK8ASK2ASK4ASK8FSK2FSK4FSK8 down demodulation I delta phase / tan-1 variance test recognition tree Q ( )2 + sqrt ( )2 amplitude Modified Liedtke IF

24. Summary • There are many publications but most of them are related to base band symbol recognition • Center frequency offset (CFO), pulse shape, timing, and channel fading can be estimated but the estimates will not be perfect • Signal type UNKNOWN should be defined in all algorithms • A good modulation recognition algorithm should be robust to CFO, timing error, and fading