A fault diagnosis scheme for unknown nonlinear dynamic systems with modules of residual generation and residual evaluation is considered. Main emphasis is placed upon designing a bank of neural networks with dynamic neurons that model a system diagnosed at normal and faulty operating points. ; To improve the quality of neural modelling, two optimization problems are included in the construction of such dynamic networks: searching for an optimal network architecture and the network training algorithm. To find a good solution, the effective well-known cascade-correlation algorithm is adapted here. ; The residuals generated by a bank of neural models are then evaluated by means of pattern classification. To illustrate the effectiveness of our approach, two applications are presented: a neural model of Narendra's system and a fault detection and identification system for the two-tank process.