Self-supervised optical fiber sensing signal separation based on linear convolutive mixing process

This paper proposeds a self-supervised signal separation method based on a linear convolutive mixing process. The method comprises three components: a linear convolutive mixer, a semantic token extractor, and a query-based signal separator. During the training phase, source signals undergo convolutional mixing within the mixer, which is a better mimic of the realistic optical fiber sensing process when compared with the linear simultaneous mixing process, resulting in a mixed signal that could be used for the self-supervised learning of the separator. The source signals' embeddings are then produced by the semantic token extractor and used as query tokens in the separator. Finally, mixed signal and source embeddings are combined and fed into the separator to produce the target source signal. The proposed method could be even used in a zero-shot setting. And the number of training samples could be expanded with this random combination of mixed signal and source embedding. In an interior setting, experimental optical fiber sensor data are gathered, including cyclical vibrations and human motions like jogging. The results of the signal separation experiments demonstrate the effectiveness of the proposed method.