The context of this thesis is the recognition of communication systems in a non-cooperative context. We are interested in the convolutional code reconstruction problem and in the constellation labeling reconstruction (the mapping used to associate a binary sequence to a modulated signal). We have defined a new statistical method for detecting if a given binary sequence is a noisy convolutional code-word obtained from an unknown convolutional code. It consists in forming blocks of sequence which are big enough to contain the support of a parity check equation and counting the number of blocks which are equal. It gives the length of the convolutional code without knowledge of the constellation labeling. This method can also be used to reconstruct the dual of a convolutional code when the constellation labeling is known. Moreover we propose a constellation labeling recognition algorithm using some equivalence classes. Two types of classes are defined: linear and affine. We observe a noisy signal which is partially demodulated (with a default labeling) and assume that the data are coded by a convolutional encoder. Thus we use the reconstruction of a code as a test and run through the classes which reveal a code structure. This classification improves the complexity of the search for small constellations (4-PSK and 8-PSK). In case of 16-QAM to 256-QAM constellations we apply the algorithm to Gray or quasi-Gray labelings. The algorithm does not give a unique result but it allows to find a small set of possible constellation labelings from noisy data.
