Parameter self-learning interference suppression method based on expanded deep network

Abstract

The invention provides a parameter self-learning interference suppression method based on an expanded deep network. The method comprises the steps of carrying out the short-time Fourier transform of an original echo signal containing interference, obtaining a time-frequency spectrum data matrix, carrying out the sub-block segmentation of the time-frequency spectrum data matrix, building the expanded deep network, taking the training set as the input of the expanded deep network for training, and substituting the training set into a verification set at the same time for model evaluation to obtain an optimal training model; and inputting a test set into the model with optimal training to identify and analyze the interference echo matrix. According to the method, interference is effectively separated and subjected to signal reconstruction, an original unsupervised decomposition problem is converted into a supervised neural network learning problem, the number of iterations is greatly reduced, algorithm hyper-parameters can be obtained through self-adaptive solution, the network does not depend on a large number of data sets any more, and the possibility of over-fitting of the training network under a small data volume is also avoided.

Description

technical field [0001] The invention relates to the field of signal processing, in particular to a method for interference suppression. On the basis of the iterative optimization method based on the model, the present invention introduces the concept of deep expansion, combines the cyclic neural network to realize the equivalent replacement of the iterative steps, and completes the optimization self-learning of the hyperparameters of the algorithm, reduces the complexity of the algorithm, and can realize the Effective extraction and separation of radio frequency interference in the signal. Background technique [0002] In recent years, the problem of radio frequency interference in radar echoes has become increasingly serious, which has adversely affected imaging processes, image interpretation, and quantitative remote sensing applications. The implementation ideas of current interference separation and reconstruction methods are mostly based on model-driven, using specific...

AI Technical Summary

Problems solved by technology

However, such deep learning methods essentially rely on big data, and achieve interference suppression by approximating limited closed-form expressions by learning from training samples, which are closer to black-box mechanisms and lack interpretability.

At the same time, the model parameters are very large, and the quantity and quality of training data are relatively high, otherwise overfitting is prone to occur

Images