Signal Separation Based on Knowledge Representation

The separation of mixed signals typically requires appropriate prior assumptions, while traditional signal separation methods struggle to describe the differences in separation targets with significant features. This paper proposes a signal separation framework based on knowledge representation, where separation targets are represented with knowledge, guiding the branches of autoencoders for signal separation. Firstly, under the proposed knowledge representation framework, corresponding knowledge representations are obtained based on observed mixed signals. Secondly, the number of branches of the autoencoder is determined based on the number of separation target signals. Then, utilizing the results of knowledge representation, a branch autoencoder network is constructed, with branches guided by knowledge to achieve the separation of target sub-signals. Finally, a self-encoding network architecture is constructed with a combination of observation signal reconstruction error and knowledge-guided error constraints. Through numerical simulations on a layered velocity model, the Marmousi-II geological model, and the MNIST dataset, the proposed method is validated by comparing the numerical energy differences between predictions and ground truths, demonstrating its effectiveness under both limited and ample data conditions.