Error retro-propagation for a chain for the blind demodulation of a digital telecommunication signal

Abstract

The present invention concerns a real-time method for the blind demodulation of digital telecommunication signals, based on the observation of a sampled version of this signal. The method comprises the following steps: —acquisition, by a sampling, of a first plurality of signals in order to each constitute an input of a network of L processing blocks (G, F, H), also referred to here as 'specialized neurons', each neuron being simulated by the outputs of the preceding block, the first plurality of signals being input into the first block simulating a first neuron of the network in order to generate a plurality of outputs of the first block; each neuron F being simulated by the outputs of an upstream chain G and stimulating a downstream chain H; each set of samples passes through the same processing chain; —the outputs of the last blocks of the network ideally correspond to the demodulated symbols; —addition of a nonlinearity to each of the outputs of the last block of the network making it possible to calculate an error signal and propagation of this error in the reverse direction of the processing chain ('retropropagation'); —estimation, upon receipt of the error by each neuron (i), of a corrective term δθand updating, in each block, of the value of the parameter θaccording to θ+=δθ.