The use of artificial neural networks-based pattern recognition techniques is now frequent and efficient in the gas sensor signal processing domain. Neural network data sets are generally built with steadystate sensor responses. Nevertheless, when the detection speed is an essential parameter, they must be monitored in a real time mode. In this paper, we present a new dynamic approach and illustrate it with surface acoustic wave sensor NO 2 responses. A filter and a detector constitute the system. They are both implemented with neural networks and both use shifting temporal windows. The filter is based on a recirculation first neural network whereas a backpropagation second neural network is used for the detector whose output is compared to a threshold to turn on an alarm. The aim is to realize a smart sensor. Since no theoretical results exist yet to find the optimal size, initialization and parametrization of backpropagation neural networks, we have studied the influences of weight initialization, temporal window width, hidden neuron number and learning rule parameters. The results show that the best convergence speeds are obtained for weight initial values that are very dependent on the network topology. It was also found that the learning quality and generalization properties were independent of the weight initialization. Other results on the network architecture and performances are presented and discussed.