The monitoring of transformer oil temperature through online data is crucial for assessing the reliability of transformer operations and the condition of internal insulation. However, the intricate environment where the oil temperature sensor operates can lead to frequent failures, resulting in potentially erroneous monitoring data. These anomalies can mislead the assessment of the transformer’s state. To address this issue, this paper begins by examining the various types of sensor data abnormalities. It then presents an innovative approach to abnormal detection based on a neural-symbolic framework specifically for transformer oil temperature data. The method utilizes an autoformer model for the neural component and incorporates symbolic reasoning to enhance the detection speed. The proposed technique effectively identifies multiple abnormalities within long-term monitoring data. Experiments conducted on a variety of transformer oil temperature datasets demonstrate the method’s effectiveness and high accuracy, with the speed of detection significantly surpassing that of existing techniques.