Air Pollution monitoring and measurement are generally done using sampling techniques and analysis equipment often heavy, complex and expensive. Although these methods offer a high measurement precision which is essential to answer standards requirements, they are not adapted for quality oriented applications where simple information with low precision can be sufficient. The use of semiconductor gas sensors networks can provide the answer for a “low cost” system intended for such applications in air pollution detection fields. Three identical portable autonomous sensors arrays were built, each containing nine commercial semiconductor sensors especially chosen to detect a large range of pollutants usually encountered in ambient air and for a large part of them regulated. In order to overcome the temporal instability and the lack of reproducibility of these sensors, a calibration and normalisation procedure was developed. The obtained systems were used for on-site pollution monitoring in association with the French National Network of Accredited Associations for Air Quality Monitoring (AASQA). Gathered data from sensors systems and network data (NO, NO₂, CO, PM2,5, …) were treated using nonlinear regression algorithms like Neural Networks with an original “fuzzy logic” type pre-treatment in order to compute a model able to predict the membership degree for three predefined pollution categories: traffic, urban and photochemical pollution, along with a pollution index describing the severity of the predominant pollution. The prediction rate was estimated system per system, and site per site for six sites. It has been shown that it was possible to obtain a quasi-universal model with a success rate over 80%.