Abstract: Traditional methods for detecting anomalies in temperature observation data by human or AI (Artificial Intelligence) face limitations in practical applications. To overcome these limitations, this paper proposes a temperature anomaly detection method based on a reconstruction-based AI model trained with a Transformer-based AutoEncoder on temperature observation data from stations of the China Meteorological Administration. The experimental results show that, with an appropriate threshold, the reconstruction-based AI model can effectively identify typical anomalies, maintaining high detection capability while significantly reducing the false detection rate. In some cases, the reconstruction-based AI model did not achieve the detection performance of traditional operational methods, especially when the observation data drifted systematically. However, the high error scores of missed samples indicated strong practical value. Future research could improve the model by introducing additional signals or combining it with other methods. The proposed reconstruction-based AI model provides a new technical means for data quality control and demonstrates its potential in this field.