Intelligent perception of Yellow River levee operational status based on transformer model: anomaly detection and pattern recognition

The intelligent perception of the safe operation of Yellow River levees is crucial for national flood control and disaster reduction. Traditional monitoring and assessment methods often face challenges in early warning capabilities when dealing with massive and complex data. To address this issue, this paper proposes an intelligent perception framework for the operational status of Yellow River levees based on the Transformer model. This framework integrates two core functions: first, it utilizes a Transformer Autoencoder (TAE) to learn data reconstruction patterns under normal operating conditions and performs high-precision anomaly detection based on reconstruction errors; second, it employs a Transformer encoder to extract deep temporal features of levee operational status, combined with unsupervised clustering algorithms, to effectively identify various operational patterns (such as stable periods, high water level impact periods, and potential risk precursor periods). To validate the effectiveness of the method, a semi-synthetic dataset simulating various operating conditions and anomalies of a certain levee section of the Yellow River was constructed. Experimental results show that the proposed TAE model performs excellently in anomaly detection tasks, with F1 scores and AUC values reaching 0.86 and 0.92, respectively, significantly outperforming Isolation Forest (F1 score approx. 0.61) and LSTM Autoencoder (F1 score approx. 0.76). In terms of pattern recognition, the clustering effect based on Transformer features (Silhouette Coefficient 0.62, Adjusted Rand Index 0.78) also surpassed comparative methods. The research indicates that the Transformer model provides an effective new approach and technical support for the intelligent perception of Yellow River levee operational status, early risk warning, and scientific operation and maintenance management.