Long-term performance of biological olfaction far exceeds that of the artificial counterpart (electronic noses or gas sensor array). This is achieved even if individual Olfactory Receptors Neurons (ORNs) are affected by long-term instabilities similar to those found in chemical sensors. Part of olfaction behavior is due to the properties of the olfactory bulbs, where ORNs signals are processed. Here, we present a preprocessing neural network based on a multi-compartment models of glomerulus units in olfactory bulbs. These models are characterized by a closed loop architecture where feedbacks provide robustness towards drift and sensor failure events. Finally, it allows classifying real time the signals coming from a sensor array showing unsupervised detection of chemical stimuli.