A robust SVR-based outlier detection algorithm for monitoring data with data weighting and temporal dependency

Abnormal observations in historical dam-monitoring records can distort conventional support vector regression (SVR), while fixed residual thresholds cannot follow local temporal variations. To address these two problems, this study develops a robust SVR-based outlier-detection framework with observation weighting and temporal adaptation. The prediction stage does not directly employ conventional SVR. Instead, a segmented Huber-Danish (HD) reliability function is embedded in a weighted iterative SVR fitting procedure. Observation weights are updated from robust standardized residuals so that samples consistent with the prevailing relationship retain their influence, whereas observations with moderate and extreme residuals are progressively suppressed. Particle swarm optimization (PSO) is used as a supporting parameter-selection tool to coordinate the SVR hyperparameters and the segmentation parameter, yielding the HD-PSO-weighted iterative support vector regression predictor. The resulting residuals are subsequently screened using a time-weighted (TW) criterion. The criterion-level development integrates a sliding time window, temporal-distance weighting, and a hybrid weighted-mean-unweighted-standard-deviation strategy to construct locally adaptive detection intervals. Validation on contaminated dam-monitoring sequences shows that the proposed predictor maintains stable fitting performance as the contamination ratio increases and outperforms Bayesian regression, multilayer perceptron, recurrent neural network, and conventional SVR in resistance to abnormal observations. Comparisons with the interquartile range, Pauta, rolling Z -score, and rolling median absolute deviation criteria further demonstrate that the TW stage achieves a better balance between precision, recall, and false-alarm control. The framework is therefore suitable for robust data-quality screening in non-stationary dam-monitoring records.