Modulation Classification Based Compressed Sensing For Communication Signals

The theory of compressed sensing (CS) has shown that compressible signals can be accurately reconstructed from a very small set of randomly projected measurements. Sparse representation of the signals plays an important role in the signal reconstruction of compressed sensing. In this paper, we propose to use signal modulation information to obtain a better sparse representation for communication signals in compressed sensing. In our approach, a tree-structured modulation classification system is used to classify five types of signal modulations: Amplitude Modulation (AM), Frequency Modulation (FM), Amplitude Shift Keying (ASK), Frequency Shift Keying (FSK) and Phase Shift Keying (PSK). The tree-structured classification system uses four signal features to classify the five modulation types, and all features are computable in the analog domain. To select a sparse transformation for the input signal, we propose a pre-trained Karhunen-Loeve transform (KLT) based CS, in which a set of KLT transformation matrices is obtained by an offline learning process for all modulation types. In an online real-time process, the modulation information of the input signal is classified and then used to select one of the pre-trained KLT matrices for providing a better sparse representation of the signal for CS-based signal reconstruction. Our experimental results show that our modulation classification technique is effective in identifying the five modulation types of noisy input signals, and our KLT based CS reconstruction has much better performances than Fourier and wavelet packet based CS for the communication signals we tested.