This paper presents a novel method for early detection of fatigue crack anomaly in complex mechanical structures, which is built upon the concepts of symbolic dynamics and finite state machines. The experimental apparatus, on which the crack detection method is tested, is a multi-degree of freedom mass-beam structure excited by oscillatory motion of two vibrators. The evolution of fatigue crack at one or more of the three failure sites causes slow variations in the natural frequencies of the mechanical structure, which are detected at an early stage from the time series data of displacement sensor signals. The proposed anomaly detection method has been validated by comparison with existing pattern recognition techniques.