Predict the indoor air quality becomes a global public health issue. That's why Airbox lab® company develops a smart connected object able to measure different physical parameters including concentration of pollutants (volatile organic compounds, carbon dioxide and fine particles). This smart object must embed prediction capacities in order to avoid the exceedance of an air quality threshold. This task is performed by neural network models. However, when some events occur (change of people's behaviors, change of place of the smart connected object as example), the embedded neural models become less accurate. So a relearning step is needed in order to refit the models. This relearning must be performed by the smart connected object, and therefore, it must use the less computing time as possible. To do that, this paper propose to combine a control chart in order to limit the frequency of relearning, and to compare three learning algorithms (backpropagation, Levenberg-Marquardt, neural network with random weights) in order to choose the more adapted to this situation.