In order to discriminate and identify different industrial machine sounds corrupted with heavy non-stationary and non-Gaussian perturbations (high noise, speech, etc.), a new methodology is proposed in this article. From every sound signal a set of features is extracted based on its denoised frequency spectrum using Morlet wavelet transformation (CWT), and the distance between feature vectors is used to identify the signals and their noisy versions. This methodology has been tested with real sounds, and it has been validated with corrupted sounds with very low signal-noise ratio (SNR) values, demonstrating the method’s robustness.