A new fault classification system for analog circuits is presented. The proposed system utilises the pattern recognition potential of neural networks and the population-based search strategy of genetic algorithms in detecting and isolating faults in analog circuits. Features that characterise the circuit behaviour under fault-free and fault situations are first simulated or measured. An unsupervised fault-grouping algorithm that estimates the overlaps between different faults in the features space is then introduced. Accordingly, a suitable training set is constructed and employed to train a population of genetically evolved neural networks to recognise circuit faults. A two-phase analog fault classification strategy is also developed. Experimental results demonstrate the high classification accuracy of the proposed system.