In the existing early fault diagnosis research of gear, most proposed methods focus on the vibration signal of tooth surface pitting fault set artificially and then the fault characteristic frequency is extracted. However, it is not clear when fault happened. In the running state of gear, how to characterize the change of performance state and the detection of health state is an important issue in the health management and fault diagnosis of gear. Accelerated life tests of gear from the 7 speeds double clutch transmission (7DCT) of plug-in hybrid electric vehicle (PHEV) is carried out for life-cycle vibration signals data acquisition. Typical time domain and frequency features are used to characterize gear performance degradation. The early weak fault stage is determined from the life cycle curve. In order to detect gear pitting fault, a parameter optimized variation mode decomposition (VMD) approach by particle swarm algorithm (PSO) is used to extract fault frequency. Then, the weighted kurtosis (WK) of envelope spectrum as the fault sensitive index is used to detect the mode with the most gear pitting fault information. This index has the most sensitivity and relativity to gear faults in the original signal than other entropy. The proposed method is also validated by life-cycle data of rolling element bearings. This method is suitable for weak fault feature detection of rotating machinery at the early fault stage.