The AIRIX facility is a high-current linear accelerator (2^3.5 kA) used for flash-radiography at the CEA of Moronvilliers (France). The general background of this study is the predictive maintenance of the AIRIX components by smart diagnosis. We are interested in analyzing the performances of the high voltage (HV) generators, which furnish the energy for accelerating the beam, and the fault detection on the whole machine. We present a tool for fault diagnosis based on pattern recognition using a neural network. We use, among other things, statistical models to define an error vector, which must be classified. To reduce the redundancy of this information and the computation time, we study two algorithms, the principal components analysis and the curvilinear components analysis. A three-layer Radial Basis Function (RBF) neural network realizes the decision rule. We propose an original algorithm to construct the network with unsupervised data. In order to take into account new states or unknown failures, we propose a strategy to adapt the network in a continuous learning. Finally, we present the first results obtained with this tool on the installation.