Dam safety monitoring is of vital importance, due to the high number of fatalities and large economic damage that a failure might imply. This, along with the evolution of artificial intelligence, has led to machine learning techniques being increasingly applied in this field. Many researchers have successfully trained models to predict dam behavior, but errors vary depending on the method used, meaning that the optimal model is not always the same over time. The main goal of this paper is to improve model precision by combining different models. Our research focuses on the comparison of two successful integration strategies in other areas: Stacking and Blending. The methodology was applied to the prediction of radial movements of an arch-gravity dam and was divided into two parts. First, we compared the usual method of estimating model errors and their hyperparameters, i.e., Random Cross Validation and Blocked Cross Validation. This aspect is relevant not only for the importance of robust estimates, but also because it is the source of the data sets used to train meta-learners. The second and main research topic of this paper was the comparison of combination strategies, for which two different types of tests were performed. The results obtained suggest that Blocked CV outperforms the random approach in robustness and that Stacking provides better predictions than Blending. The generalized linear meta-learners trained by the Stacking strategy achieved higher accuracy than the individual models in most cases.

Dam safety monitoring is critical due to its social, environmental, and economic implications. Although conventional statistical approaches have been used for surveillance, advancements in technology, particularly in Artificial Intelligence (AI) and Machine Learning (ML), offer promising avenues for enhancing predictive capabilities. We investigate the application of ML algorithms, including Boosted Regression Trees (BRT), Random Forest (RF), and Neural Networks (NN), focussing on their combination by Stacking to improve prediction accuracy on concrete dam deformation using radial displacement data from three dams. The methodology involves training first-level models (experts) using those algorithms, and a second-level meta-learner that combines their predictions using BRT, a Linear Model (LM) and the Greedy Weighted Algorithm (GWA). A comparative analysis demonstrates the superiority of Stacking over traditional methods. The GWA emerged as the most suitable meta-learner, enhancing the optimal expert in all cases, with improvement rates reaching up to 16.12% over the optimal expert. Our study addresses critical questions regarding the GWA’s expert weighting and its impact on prediction precision. The results indicate that the combination of accurate experts using the GWA improves model reliability by reducing error dispersion. However, variations in optimal weights over time necessitate robust error estimation using cross-validation by blocks. Furthermore, the assignment of weights to experts closely correlates with their precision: the more accurate a model is, the more weight that is assigned to it. The GWA improves on the optimal expert in most cases, including at extreme values of error, with improvement rates up to 41.74%. Our findings suggest that the proposed methodology significantly advances AI applications in infrastructure monitoring, with implications for dam safety.

Dams are complex systems that involve both the structure itself and its foundation. Rheological phenomena, expansive reactions, or alterations in the geotechnical parameters of the foundation, among others, result in non-reversible and cumulative modifications in the dam response, leading to trends in the monitoring data series. The accurate identification and definition of these trends to study their evolution are key aspects of dam safety. This manuscript proposes a methodology to identify trends in dam behavioural data series by identifying the influence of the time variable on the predictions provided by the ML models. Initially, ICE curves and SHAP values are employed to extract temporal dependence, and the ICE curves are found to be more precise and efficient in terms of computational cost. The temporal dependencies found are adjusted using a GWO algorithm to different function characteristics of irreversible processes in dams. The function that provides the best fit is selected as the most plausible. The results obtained allow us to conclude that the proposed methodology is capable of obtaining estimates of the most common trends that affect movements in concrete dams with greater precision than the statistical models most commonly used to predict the behaviour of these types of variables. These results are promising for its general application to other types of dam monitoring data series, given the versatility demonstrated for the unsupervised identification of temporal dependencies.

Dam-safety monitoring flags pendulum readings outside a prediction band, but autocorrelated residuals make these transgressions cluster into streaks of redundant warnings causing alarm fatigue. This study aimed to reduce it by learning the normal behaviour of the transgression series and warning only when their persistence departs from it. Three detectors are compared: a Gaussian-threshold baseline (M0), a sliding-window detector with a 99th-percentile threshold (P99), and a two-state hidden Markov model (HMM) that learns streak persistence. They were applied to 28 pendulums of three dams (1, 23, and 4 sensors), and an identifiability criterion, met by 20, decides when the HMM can be trusted. In Dam2, the most representative, the HMM reduces flagged records and episodes by 26.0% and 33.3% with respect to P99 among identifiable sensors and the proposed configuration (HMM where identifiable, P99 otherwise) by 19.2% and 25.8%, with variation across sensors. On a synthetic series with habitual streaks and a known drift, the HMM flagged only the drift, after 12 records, whereas the other detectors, including rules of two, three, and five consecutive transgressions, also flagged the habitual streaks. Operationally, the HMM replaces such hand-set rules with a persistence learned from each sensor’s data.

Automated dam-safety monitoring compares each pendulum reading with a prediction band from data-driven predictive models, but its operational deployment is undermined by alarm fatigue: because prediction residuals are autocorrelated, band transgressions might arrive in clustered streaks and a per-record threshold overwhelms the operator with redundant warnings. This work treats the transgressions as a binary time series whose normal behaviour is learned during training and compares three detectors, each addressing a limitation of the previous one: a Gaussian-threshold baseline (M0), a sliding-window detector with a 99th-percentile threshold whose window is sized from the residual autocorrelation (P99), and a two-state hidden Markov model (HMM) that learns the persistence of the streaks. Applied to 28 pendulums of three dams, M0 can even increase the number of flagged records, whereas the HMM flags the fewest: in the largest case study it reduces them by 58.960 % with respect to the raw transgression series and by 45.594 % with respect to P99. An identifiability criterion based on the emission separation indicates when the HMM can be trusted (20 of 28 sensors) and when to fall back to P99, and a sensitivity analysis shows that the estimates are robust in a neighbourhood of the reference initialisation, with the self-transition probability of the Normal state as the critical parameter.