Abstract: The number of industrial and commercial users is huge, and it is difficult to perceive the abnormal reactive power compensation fault in time due to the lack of necessary professional knowledge, which is easy to cause operating with defect for a long time, reducing the power factor of electricity consumption and increasing up the cost of energy consumption. Taking advantages of the characteristics of controlling the power factor in the target range and the active load and power factor having roughly fixed joint distribution during the normal operation of the user reactive power compensation device, a perception method of abnormal compensation for user reactive power based on the joint distribution of measurement data is proposed. First, the frequency distribution matrix of active power and power factor is generated based on the user metering data, so as to characterize the operation status of the reactive power compensation device for the user. Then, the two-dimensional convolutional neural network（CNN） is applied to train and test the sample frequency distribution matrix sets in normal and failure conditions. When the frequency distribution matrix of the sample to be inspected is diagnosed as an abnormal label by the model, it is considered that the reactive power compensation device on that day is in an abnormal fault state. The test analysis based on the field fault data shows that the proposed method has good performance in accuracy and false detection rate, which can prompt users to repair the equipment and improve the power efficiency in time.