Abstract:In random working condition, the characteristic of amplitudefrequency modulation will change with different rotor speeds and load, which makes the weak fault diagnosis of Wind turbine gearbox extremely difficult. In view of this, we proposed to use different constant conditions instead of random conditions, and the normalized full vector band energy was used to diagnose the weak failure of gearbox under random conditions. In this algorithm, random conditions are decomposed into several constant conditions first to reduce its dimensions. Then, homologous signals of each constant condition are melded by full vector theory to ensure the integrity of the weak source information. In order to eliminate the modal aliasing effect caused by difference conditions, full vector signal is broken by using FIR filter. Given that band energy can quantitatively distinguish different operating mode, the rate of band energy can distinguish the working state and the information entropy reflects the difference between the motivational source and the motivation way. The sum of each band energy entropy, rotating frequency band energy and the rate of change of band energy are extracted as the distinct feature of gear working state. After the above steps, 150 groups of vibration signal from wind turbine gearboxes under random conditions were identified by Bayesian classification after GMM describe the signal character, and the accuracy rate shows that early local weak fault in random condition can be identified accurately.