Vibro-Acoustic is a relevant field used for non-destructive diagnosis. Indeed, sound and vibrations radiated by a machine are representative of its working order and ensure that a machine is in good working order by just using microphones and accelerometers. Therefore, end of assembly line controls in mass production industries are done using signal processing methods applied to acoustic and vibratory recordings collected during the finished product operating cycle. However, this diagnosis often needs to be done on signals devoid of noise in order to prevent from misinterpretations which requires to measure machines in healthy environment decoupled from noise and vibrations, but anechoic rooms are not compatible with mass production industries because of specific structures to design at each end of lines involving space congestion and high costs. Deep learning methods have been successfully applied in pattern recognition tasks such as computer vision and automatic speech recognition. They allow to design decision models based on prior learning without the need of any feature engineering. Strength of this, is the fact that models consider full input signal instead of extracted features. This work introduces results achieved with Deep Learning methods for the compliance classification of rotating machines based on sounds recorded in harsh environment.