On nuclear sites, such as nuclear power plants, instruments for measuring atmospheric radioactivity (CAM - Continuous Air Monitor) are deployed to ensure the radiation protection of workers. This type of instrument continuously samples ambient air aerosols on a filter, measures as an energy spectrum the radioactivity accumulated on the filter in real time, and alerts the operator if an atmospheric contamination is detected. To detect any artificial alpha emitters (e.g., 239Pu), the CAM estimates the contribution of Radon progenies (218Po & 214Po) that interfere in this region. Under standard conditions, this estimation is done very accurately, but in the specific case of nuclear facilities dismantling, sudden variations in aerosol ambiance, both in terms of size distribution and concentration, lead to a modification of the nuclear measurement, hence an incorrect estimation of the background noise and ultimately a false increase in artificial alpha activity. This can lead to a false alarm (Hoarau et al., 2022). In this work, we are therefore interested in developing a deep learning algorithm to automatically detect the presence, or absence, of artificial alpha emitters based on the knowledge of the nuclear measurement. It is a promising new strategy to correctly assess airborne contamination even in high radon atmosphere and with coarse aerosols interferences. However, the “black-box” aspect of decision-making of neural networks in an issue as sensitive as nuclear safety represents a major obstacle to the use of these techniques in practice. It is hence essential to implement tools to visualise the classification or regression process of the model. Theoretical guaranties and constraints can also be imposed on the model reasoning, even though it can lead to the deterioration of pure performances. In particular, an isotonic neural network was implemented in order to match prior physical knowledge on the problem.