The order tracking is an effective method to solve the problem of variable speed fault diagnosis. The key premise is that there is a speed signal as a reference. However, due to the influence of strong background noise and weak harmonic relations, the accuracy and adaptability of the existing speed estimation methods need to be further improved. Therefore, an improved multi-order probability approach (MOPA) based on multi-sensor signals is proposed to estimate the instantaneous speed. Firstly, according to the unity of fundamental frequency and the difference of dominant component of different sensor signals, the time-frequency diagram with a strong harmonic relationship is constructed through the normalization and fusion of instantaneous slices of the time-frequency diagram. Secondly, to eliminate the intermittent constant frequency and short-time broadband background noise in the transverse and longitudinal direction of the time-frequency diagram under time-varying conditions, a sliding noise reduction method is proposed. Finally, MOPA is implemented based on the processed time-frequency diagram to realize automatic estimation of instantaneous speed, and the fault diagnosis problem under variable speed of wind power gearbox is solved by combining the order tracking method. The measured data evaluate that the accuracy and adaptability of the instantaneous frequency estimated by the improved MOPA are better than those of the opposite methods. The mean absolute percentage error is 0. 56% , which is lower than 15. 73% , 13. 99% , and 1. 21% of the comparison methods. Combined with the order analysis, the abnormality of the wind turbine gearbox under variable speed is diagnosed.