关键词: 水工结构, 异常识别, 局部异常系数, 野值, 长周期监测数据

Abstract: A density-based clustering outlier detection algorithm using improved local outlier factors is presented for analysing long-term hydraulic structure monitoring data. It is aimed at the problems that the distribution assumptions are difficult to meet, the number of outliers to be processed is limited, and the outliers are difficult to effectively be quantified. It divides the long-term data set into extreme clusters, outlier clusters, and normal clusters; in each cluster, anomalous possibilities are assigned in different ways, and it obtains an anomalous possibility that considers independent variables and effect sizes comprehensively. Its sequence diagram realizes the identification and quantitative analysis of long-term data sets of hydraulic structures. The core algorithm requires no distribution assumptions. This method can improve the definition of the reachable distance for the local outlier factor algorithm, expanding the difference between high and low anomaly coefficients. Thus, it can easily distinguish the outliers from other data points. Based on the long-term monitoring data from a water transfer project, their credibility is calculated using such a sequence diagram for cases where the number and locations of outliers are unknown. Using the credibility as the weight of the regression model, the predictions are greatly improved in comparison to the unweighted model, verifying the effectiveness of our new method.

Key words: hydraulic structure, anomaly recognition, local outlier factor, outlier, long-term monitoring data