Aiming at the shortcomings of neural network recognition in static pattern recognition, a dynamic pattern recognition technology developed in recent years—hidden Markov model is used to analyze gearbox vibration signals. First, the statistical characteristics of the gearbox vibration signals in the time domain, frequency domain and time-frequency domain are extracted to form a 34-dimensional full feature vector. Trained a set of full feature-hidden Markov model libraries;then, through the principal component analysis technology, the full feature vector is reduced in dimension, and the first 7 principal components whose absorption information is more than 98% constitute the principal component feature vector. Another set of principal component-hidden Markov model library was trained. Two sets of independent model libraries are used for gearbox fault identification. The results show that the full feature-hidden Markov model library has 97. 9% accuracy for the identification of normal gears and gear broken tooth and 100% for gear pitting. The program takes 22. 328 s. The recognition accuracy of component-hidden Markov model library for gear pitting and gear tooth failure is 100%. The program takes 4. 879 s. Therefore, the dimensionality reduction processing of the vibration signal feature does not reduce accuracy of fault identification, but improves the accuracy of fault recognition, and greatly increases the speed of the program. This is of great significance for fault diagnosis of mechanical systems.