A Group Lasso-Based Neural Network Pruning Method for Power Amplifiers

Abstract

The invention discloses a Group Lasso-based neural network clipping method for a power amplifier, which belongs to the technical field of communication. First, for the original neural network structure, all the output weights connected to the same neuron in the other layers except the output layer are grouped into one group; after grouping, L is performed on each group of weights separately. 2 norm constraint; then L of the weights of all groups in the original neural network structure 2 The sum of the norms is added to the original loss function Loss as a Group Lasso penalty 1 After that, get a new loss function Loss 2 ; Through the BP algorithm to Loss 2 Minimize the training, find the weight groups whose convergence is close to 0, and remove the neurons connected by these weight groups to obtain a simplified neural network, and use the BP algorithm to analyze the original loss function Loss 1 Perform training to obtain a simplified neural network model after training, and use the model to model or predistort the power amplifier. The invention can suppress the overfitting in the training process, improve the pre-distortion effect, reduce the amount of calculation, and is convenient for engineering application.

Description

technical field [0001] The invention belongs to the technical field of communication, in particular to a Group Lasso-based neural network cutting method for power amplifiers. Background technique [0002] With the continuous development of communication technology, in order to use limited spectrum resources more efficiently, new modulation techniques with high spectrum efficiency are used in communication standards. However, these new modulation methods make the peak-to-average ratio of the signal higher and the envelope fluctuation larger, causing the signal to be seriously distorted after passing through the RF power amplifier; not only the EVM before and after the signal output is seriously deteriorated, but also serious out-of-band distortion . [0003] In order to correct the distortion produced by the signal passing through the power amplifier, the power amplifier needs to be linearized. Among the current linearization technologies, the predistortion technology is wi...

AI Technical Summary

Problems solved by technology

[0008] In order to overcome the disadvantages of the neural network in the prior art, such as difficulty in manual debugging, easy overfitting, weak robustness, difficulty in engineering implementation, and large hardware resource overhead, the present invention provides a neural network based on Group Lasso for power amplifiers The clipping method can clip or directly train the neural network structure of the real multi-layer neural network, thereby suppressing overfitting and easy engineering implementation

Images