Abstract: Aiming at the problem that it is difficult to identify the early fault symptoms of transformer on-load tap changer, considering the speed of signal calculation and the accuracy of fault diagnosis, we propose an improved random forest fault diagnosis algorithm based on image texture feature. The preprocessed vibration signal is converted into a two-dimensional time-frequency grayscale diagram which reflects different operating states through wavelet packet decomposition. The gray level co-occurrence matrix describing the relationship between pixels in the region is used to quantitatively characterize the original signal, and six-dimensional eigenvectors are extracted and input into the random forest algorithm. The traditional voting rules ignore the individual strengths and weaknesses of the classifier, thus, the area under the curve (AUC) of operational characteristics of testee is used as a criterion to construct an improved random forest classifier so as to realize the accurate recognition of the abnormal state of the on-load tap changer. The experimental results show that GLCM texture features and improved random forest classifier increases the recognition accuracy to 97.5%, with zero false alarm rate and better online diagnosis efficiency. This method has high application value in the field abnormal state identification of on-load tap changers.