An approach for fault diagnosis based on pattern match of characteristic waveforms in frequency domain was proposed. It used the principle of global similarity on the characteristic waveforms of FFT-based amplitude spectrum of the vibration observation samples belonging to the identical pattern class. It was then applied to fault diagnosis of rolling element bearings, and compared with some typical diagnosis approaches based on pattern recognition. It was found that the cosine similarity, correlation similarity or mutual information similarity from pattern match of the characteristic waveforms in frequency domain reached the maximum respectively, when the test sample belonged to the identical pattern class as the training samples. According to this, the characteristic similarity threshold for pattern classification was determined, which contributed to 100% classification accuracy. The results show that the proposed approach for fault diagnosis needs neither sophisticated feature extraction nor complicated classifier design. It can accurately classify multiple complex patterns only by simple pattern match on characteristic waveforms in frequency domain and feature threshold comparison. In addition, it is suitable for solving the problem of small -sample classification with high classification efficiency and strong self -learning ability. It is obviously superior to some typical diagnosis approaches based on pattern recognition, andhas great application potential in constructing on-line automatic fault classification system.