The existing fault diagnosis system focuses on data pre-processing to extract fault features, the fault modeling process is complex, and it is difficult to discover and learn new fault modes during system operation. An analog circuit fault automatic diagnosis system based on CFPNN (Constructive Forward Propagation Neural Networks) was analyzed in this paper. The CFPNN's computation is very small and it can capture all kinds of transient-state fault. The system fault-diagnosis precision is 100% when the circuit fault-feature deviation within ± 4%. This system can set a reject pattern to avoid risk from false-recognition with big disturbance and discover new knowledge (new fault-sample). The system not only learn new testing-sample(new knowledge) and finish information fusion by adding new-sample in the original CFPNN but also without retraining the history samples. The above conclusions are validated by theory and experiments.