Unsupervised Source Separation for Multi-Label Classification

This paper exploits blind source separation for the purpose of multi-label classification (MLC). The proposed method addresses the multi-label classification problem in two stages, source separation of data consisting of two or more observed classes followed by a single-label classification. This method is evaluated on two data-sets: the new application to electromagnetic interference (EMI) signals that are used for high-voltage power asset condition monitoring, and overlapping handwritten digits (multi-MNIST) for further model validation. Our approach is compared to other deep neural networks such as variational auto-encoder (VAE) and capsule network (CapsNet). Results show high average precision on both data-sets and outperform in most metrics other deep auto-encoding networks that infer label decision from the latent variables distribution.