Recognition Of Subtle Features Of The Communication Signal Classification

The classification of communication signal was mainly used in individual communication transmitter identification and further provides an indication for viewing the architecture of the communication network. The Identification process was based on the research of the fine feature which was brought by the difference of the transmitter hardware. This thesis focuses on the problem of feature transformation as well as the feature dimensionality reduction to provide a practical method to classify the similar signal.(1) We clarify the concept of fine feature and signal identification, and analyze the underlying generation mechanisms of the subtle feature. We also give an overview of the pattern classification process and discuss both the feasibility and the framework of the classification based on subtle feature.(2) Traditionally, the objective of time-frequency research uses the standard time-frequency distribution as the analysis tool, consequently falling into the problem of optimizing on an predefined time-frequency kernels, thus the resulting time-frequency representation (TFR) failing to discriminate the difference. It may be advantageous to design a class-dependent time-frequency kernels method highlights the difference of class. Each point on ambiguity plane is considered independently when using fisher discriminate ratio in the original feature reduction method, so the feature space of reduced dimension contains the classification information redundancy. Karhunen-loeve expansion can help eliminate the correlation between feature elements but are computation complex. A joint fisher discriminate ratio and karhunen-loeve expansion method was used to efficiently reduce the feature dimension and decorrelate the feature while saving the computation power and preserving the classification performance. Simulation results suggest the method's computation simplicity and good classify performance.(3) We apply the local wavelet packet transformation and the local cosine transformation with the smoothed window and folding operation to assure the continuous and the orthogonality of the resulting coefficient. A best-basis operation and discrimination power analysis was implemented to choose a subset of the bases for classification. We find if the Fisher discrimination ratio was calculated on the chosen subset, a further feature reduction can be achieved. Moreover, Owing to the existence of fast algorithm for wavelet decompose and cosine transform, the classification operation can be speed up while retaining a satisfied performance.