Airix accelerator diagnosis using pattern recognition

The AIRIX accelerator is used to realize flash radiography during hydro-test experiments. We have developed a tool to monitor the functioning state of this facility. This procedure is based first on signal processing and data mining. It allows to select the most important parameters to characterize the functioning state. A pattern recognition strategy realizes the identification of the functioning state. We propose an original algorithm to construct and initialize a radial basis function neural network with experimental data, and to take into account new states which could appear. The objective is to detect transient states before failure. (authors)