The natural identification of an odor, smelled once, highlights a new idea on the artificial intelligence for electronic noses to mimic as close as possible the biological olfactory system. For the first time, a natural on-line training with only one sample, to extract both eigen-weights and eigen-bias, for each odor is built herein to elaborate a natural identifier neural model in a real work environment. The proposed model efficiently reduces the maximum extent of traditional neural models complexities, namely generic work-laboratory, dimensional data learning (typically from 60 to 80%), model adaptability complication, time-consuming, heavy experiment materials and chemical products. The outcomes show that the invented model identifies correctly 100% the three hazardous contaminants of the test data (all data minus only three samples for learning) processed from gas sensors weakened in the harshest conditions. Therewith, this natural identifier model methodology is elaborated easily enough to be efficient in different environments conditions to accurately control the air quality safety in any local area.

The assessment sensors’ signals, noised in real complex condition, is a serious problem, which restrains the performance of the neural model application laboratory in its real application. Indeed, the Self-Organizing Maps (SOM) and neural multilayer perceptron (MLP) are fused to reinforce the olfactory devices power for optimum pollutants identification. As soon as the SOM is developed, with the whole key parameters, carefully selected (as Epanichnikov kernel function), to automatic elimination of uninformative signals, the few useful signals are MLP trained to optimally determine the model to improve the results performance. Fast precise identification, short execution duration, small local data, having strong signals selected, are all required to attenuate the sensors several criticalities emerged from the lowest content (1ppm) of interfering mixed gases at relative humidity presence. Moreover, the practical application of this hybrid tool on 1ppm concentration has another advantage to ensure the air quality, either in critical condition control or in safety.

The closely detailed search of sensed feature quality to ensure fault detection from weak sensors, which can skew the application model results signals, is required. As principal component analysis (PCA) is limited to correct isolate faulty signals’ impact, unlike the relative performance of neural Kohonen self-organizing map’s model to monitor air quality on any real complex condition. Indeed, this unsupervised method is enhanced by itself (2-SOM) and SOM hierarchical clustering (SOMHC) with both learning types; sequential and batch. These former models are improved also with Bubble, Gaussian, Gaussian Cut and Epanichnikov Kernel neighborhood functions, in graphical user interface form. Therefore, the study demonstrates more eective and complete results showing in quantization and topography errors including responses classification accuracy as well as in the KSOM-HCs dendograms view. Furthermore, this tool is relevant to provide the credible informations of pollutant detection, dedicated to Human Health safety, despite the conditions complexity.