Identification of deformation outliers in concrete dams considering local anomaly features

Concrete dams constitute critical infrastructure for water resource management, flood control, and power generation, and their safe operation is paramount to public safety, property protection, and socioeconomic stability. Deformation monitoring serves as a fundamental means to assess structural health and detect early signs of instability. However, due to complex environmental loads, material degradation, and structural evolution, deformation monitoring data frequently contain outliers—measurements that significantly deviate from normal behavioral patterns. Traditional detection methods, such as statistical approaches or conventional machine learning models, often suffer from poor adaptability to local anomalous features, limited feature extraction capability, and low classification accuracy. Statistical methods typically rely on distributional assumptions that rarely hold for complex dam data, while single machine learning models often fail to capture subtle, transient anomalies embedded in noisy sequences. Specifically, existing methods struggle to effectively distinguish among four critical anomaly types: outlier, step, oscillation, and trend anomalies, thereby leading to potential safety risks. To address these challenges, this study proposes a novel multi-type outlier identification method that integrates Shapelet Transform (ST) with advanced deep learning frameworks. The proposed methodology specifically combines a hybrid one-dimensional Convolutional Neural Network (1D-CNN) and Long Short-Term Memory (LSTM) network to enhance both the accuracy and interpretability of anomaly detection. The framework begins with ST-based local feature extraction, where candidate shapelets—discriminative subsequences representing key temporal patterns—are generated using sliding window techniques across monitoring sequences. The minimum Euclidean distance between each candidate and all time series is calculated to construct a distance matrix. Quality evaluation using Information Gain (IG) selects high-quality shapelets that best distinguish normal behavior from various anomaly types, ensuring that only the most discriminative features are retained. These selected shapelets are then used to construct a feature matrix, where each row corresponds to a sequence’s minimum distances to all shapelets, augmented with class labels.

Subsequently, this transformed feature matrix is input into the hybrid 1D-CNN-LSTM model. The 1D-CNN layer extracts local spatial features through convolutional kernels, effectively processing the shapelet-derived distance features, while the LSTM layer captures long-term temporal dependencies, enabling the model to characterize both abrupt shifts and gradual trends within the deformation data. Hyperparameters were optimized through grid search and cross-validation to ensure robust performance. Finally, a classifier with optimized parameters is employed to identify and classify unlabeled data. The method was validated using engineering data from the Laxiwa Arch Dam, focusing on monitoring points at the dam crest, shoulder, and middle body. The monitoring data spanned from November 2013 to July 2021 and were processed using a 14-day sliding window to balance feature coverage and computational complexity. Due to the scarcity of real anomalies, synthetic anomalies were generated according to physical patterns to augment the training set. Experimental results demonstrate that the proposed method achieves precision and a weighted evaluation metric exceeding 93.00% across different structural positions. Specifically, the model achieved a precision of 93.34% and an F1 score of 95.88% on the test set. Compared with traditional machine learning models such as ST-SVM, ST-XGBoost, and ST-RF, as well as the conventional CNN-LSTM model, the weighted evaluation metrics improved by 5.01% to 6.34%. The model successfully identified various anomaly types, including outlier, step, oscillation, and trend anomalies. At specific monitoring points such as PL5-1 and PL3-3, all anomalies were detected, with only one missed detection at PL7-1. These findings indicate that the extracted shapelets correspond to physical anomaly patterns, thereby providing interpretability. The proposed method effectively distinguishes among different types of outliers, exhibits strong adaptability to structural position differences, and offers significant improvements in detection accuracy and model interpretability for concrete dam safety monitoring.