A Superimposed Signal Separation Algorithm for Satellite Navigation Receivers in Complex Environments Based on Capsule Networks

To improve the accuracy of satellite navigation in complex environments, it is crucial to develop advanced signal reception algorithms that mitigate the multipath effect and enhance signal clarity and reliability. This paper proposes a blind source separation algorithm based on capsule networks for mixed signals. In the encoder stage, the one-dimensional signal dataset is first transformed into a time-frequency image set using STFT to adapt to subsequent operations. In the separator stage, a structure combining capsule networks with convolutional blocks and deconvolutional blocks is adopted. Additionally, attention mechanisms, residual learning, and dilated convolutions are incorporated to further analyze and filter the output of the features by the input layer. This allows the separator to extract signal details layer by layer, separate different signals, increase the receptive field of convolution kernels, cover a larger signal range, and enhance the perception ability of the capsule layer. In the decoder stage, we reconstructs the time-frequency binary image of the separated signals by performing morphological methods on the separated signals output by the separator, thereby achieving the purpose of signal processing.