The Auto Regression Moving Average(ARMA) model was established based on online oil monitoring data when the equipment was running stably,and the model residuals was used to implement fault detection and early warning.The model residuals for a period of time before the fault was classified into the stationary and fault latency period by Kmeans clustering.The mean value of the center of the two clusters was set to the residual threshold value.The system will alarm once crossing the border of residual threshold value.The performance of the method was validated with the real online oil monitoring data.The results show that the ARMA model has a good fitting effect on the online monitoring data,and the system can effectively alarm in advance by setting the residual limit value when the equipment is running in the fault latency period before failure.