A rotary machine abnormality detection device capable of determining abnormality with higher accuracy, a method thereof, and a rotary machine provided therewith. A rotary machine abnormality detection device detects abnormality in a rotary machine M having at least first and second rotating bodies using a specific feature amount based on a frequency spectrum of measurement data of vibration. The device obtains, from the frequency spectrum, a fundamental wave component F1 of a fundamental wave lobe mesh frequency f1 and n-th harmonic wave components Fn of n-th harmonic wave lobe mesh frequencies fn up to XAB, and obtains the specific feature amount based on these components. The f1 is obtained by dividing a rotational frequency by a teeth number of the first rotating body and the XAB represents the least common multiple of the teeth numbers of the two rotating bodies.