Architecture neuromorphique pour l'identification de menaces radiologiques dans le cadre des applications de sécurité
This 42 months collaborative public-private partnership project aims to develop a low-cost, unified nuclear threat detection platform, capable of real-time neutron-gamma discrimination and real-time radionuclide identification with plastic scintillators usage. An original signal processing strategy is proposed, based on artificial spiking neural networks, to enable fast radionuclide identification at low count rate, even for mixtures. This project relies on low-cost and robust solid organic scintillators previously designed in the Neutromania ANR project and on recent work in neuroscience that shows that spike-based learning and recognition can lead to optimal decoding of population of codes. The project ambition is to enable low-cost and large-scale deployment of radiological threats detection and identification systems by states, in order to curb illegal traffic of radioactive material from mafia and terrorist groups, which constitutes a serious threat to population and dramatically endangers surrounding people health. The signal processing approach that we propose for this platform takes inspiration on the working principle of the biological cochlea, which temporally encodes the sound spectrum through band-pass filterbanks. Spiking neural networks have gained momentum over the last decade, not only as a neuroscience modeling tool, but also as an efficient and computationally powerful processing model, for learning and recognition of temporally coded data. In the cochlea, each channel responds to a small frequency band and the global output spiking activity corresponds to a probabilistic coding of the input stimuli. This is very similar to nuclear disintegration detection signals, where events can be filtered according to their energy. Jointly to traditional signal preprocessing and novel spike-based processing, an efficient neural network based signal processing approach must also be developed to discriminate neutron and gamma signals. On the basis of neural network technology, an alternative to traditional spectroscopy will be developed in order to give quick and accurate radionuclide identification. For this purpose, a signal acquisition device will be jointly designed in association with a digital signal processing hardware architecture leading to the complete prototype of a low-cost threat detection and identification platform. A real world experimental database will be built in order to qualify its radionuclides identification and neutron/gamma discrimination capacities of the prototype and to prepare to real world tests. While organic scintillators are a favorable study case, the proposed approach must be able to work using scintillators on-the-shelf. In terms of industrialization, plastic scintillators are less expensive, less temperature sensitive, robust and are manufacturable in large volumes, which make them good candidates for low-cost, large-scale, deployments. Finally, designing a new generation of signal processing for such detectors can makes the difference that enables an easy industrialization of our solution in comparison with costly and harmful state-of-the-art approaches.