Single event upsets simulations on neural networks

International audienceThis paper investigates the tolerance of Artificial Neural Networks with respect to a particular type of fault, so called Single Event Upset (SEU), induced by harsh environments. SEU faults result in the inversion of the content of a memory cell as a consequence of the impact of a charged particle with sensitive regions of an integrated circuit, and can potentially affect any of the registers, flip-flops or memory cells of the circuit. SEU like faults were injected in a software model of an Artificial Neural Network devoted to the detection of a particular phenomenon, called "protonic or electronic whistlers", perturbing the operation of electronic equipments on board spacecrafts. Two versions of the tested ANN have been experimented as well as two different data formats. Obtained results showed that ANNs tolerate a high percent of SEU-like faults, making the ANN approach and the corresponding dedicated circuits suitable to space applications