Device for linearising a power amplifier of a communication system by digital predistortion

A device for linearising, by digital predistortion, a power amplifier of a communication system comprising at least one radio frequency transmitter configured to transmit at least one frame of symbols, the linearising device comprising:

• a module for digitally predistorting the signal supplied by the radio frequency transmitter configured to infer a polar domain predistortion to be applied to said signal, the predistortion module having an indirect neural network learning architecture and comprising:
• a first neural network configured to correct an amplitude distortion produced by the power amplifier by receiving as input the amplitude of an output |y| of the power amplifier, the first neural network being configured to provide a reciprocal function of the amplitude distortion function produced by the power amplifier;
• a second neural network configured to correct a phase distortion produced by the power amplifier by receiving as input the amplitude of an input |x| of the power amplifier, the second neural network being configured to provide a phase function f−Φ, providing the opposite in phase of each phase;
• wherein the linearising device further comprises:
• a module for optimising the first neural network and optimising the second neural network configured to implement:
• a meta-learning of the first neural network, and
• a meta-learning of the second neural network,
• the optimisation module being configured to implement each meta-learning from a set of operating data of the power amplifier, each meta-learning comprising a plurality of distinct prediction tasks respectively associated with distinct groups of operating data and metadata from the data set, the distinct groups being representative of distinct use cases of the power amplifier,
• the optimisation module being configured to implement the meta-learning of the first neural network, and the meta-learning of the second neural network, using:
• from a first part of said previously obtained data set, a meta-initialisation of the first neural network, respectively of the second neural network, providing beforehand an initialisation of the initial weights of the first neural network, respectively of the second neural network;
• from said initial weights and a second part of said data set corresponding to the current use case of the power amplifier, a meta-matching of the initial weights into optimal weights of the first neural network, respectively of the second neural network;
• the optimisation module being configured to train the first neural network and the second neural network by optimising a cost function.