Abstract: Efficiently detecting abnormal data holds crucial importance in dam safety monitoring. Traditional methods exhibit instability in quantifying deviation sizes. This paper introduces a deep learning-based algorithm for identifying abnormal data in dam safety monitoring. It mimics the manual identification process, employing classification and identification stages for detecting abnormal data. This approach is adept at handling data with uncertain change trends, generating labeled datasets through a feedback mechanism. Experimental results demonstrate the algorithm’s commendable performance, with an average accuracy rate exceeding 0.97 and a full rate of 0.97 or higher. Notably, it achieves an accuracy rate exceeding 0.99 for various types of anomalous addition patterns, particularly excelling in identifying small value anomalies. Compared to traditional methods, this algorithm exhibits superior detection stability, robustness, and practicality in anomaly data detection.