Information extraction is a very important problem nowadays. In diagnostics, it is particularly useful when one desires to isolate information about machine damage from a measured diagnostic signal. The method presented in this paper utilizes the idea that is based on a very important topic in numerical algebra, which is nonnegative matrix factorization. When applied to the matrix of multidimensional representation of the measured data, it can extract very useful information about the events which occur in the signal and are not recognizable otherwise. In the presented methodology, we use the algorithm called Semi-Binary Nonnegative Matrix Factorization (SB-NMF), and apply it to a time-frequency representation of the real-life vibration signal measured on faulty bearing operating in a belt conveyor driving station. Detected impulses of local damage are clearly identifiable. Performance of the algorithm is very satisfying in terms of time efficiency and output signal quality.