I am testing my supervised NMF algorithm to extract signal from observation that have only one source inside. I am new here and I wonder this is very weak model or not? Is it acceptable in signal processing academic paper to start experimental result with factorization of only one source separation from the environment?
Yes, that's a first sanity check. Basically, the idea of averaging multiple observations of a single signal buried in noise is a basic issue in signal processing. You can have several observations with multiple noise realizations and statistics, with different delays on the signal, with different data transformations (more or less linear).
If this does not work in your case, don't expect to much of NMF performance on multiple sources, and verify assumptions or test other source separation methods.