The invention discloses a method for diagnosing vibration faults of a rotary machine. According to the method for diagnosing the vibration faults of the rotary machine, vibration data in various states are collected, process amount data in various states are correspondingly acquired at the same time, and therefore the vibration data and the process amount data in various typical vibration fault states of the rotary machine can be obtained. Characteristic values are extracted from the vibration data and the process amount data respectively to form characteristic value matrices, the characteristic value matrices completely describe the process varying rules of various typical faults, namely a sample matrix database of the various typical faults. On the basis, according to the method for diagnosing the vibration faults of the rotary machine, the image recognition technology is applied to the field of fault diagnosis, the defects of manual fault recognition and single characteristic quantity fault recognition in the present field of fault diagnosis are overcome through matching calculation of the characteristic value matrices, and high-accuracy automatic fault diagnosis is achieved.