基于孤立森林算法的高拱坝施工期变形异常值检测模型

A model for detecting deformation outliers during construction of high arch dams based on isolated forest algorithm

【目的】高拱坝建设过程中的倒悬变形是导致施工期裂缝的原因之一,采用连续测斜设备开展施工期变形监测时,因分层浇筑等施工过程造成的监测异常值需进行检测和判断。【方法】结合工程实际所需精度,设计了变形监测试验以模拟拱坝工作状态和“突变”过程。针对试验结果,提出了一种基于孤立森林算法的异常值检测模型,通过小波变换预处理数据,降噪后的数据采用改进孤立森林算法检测异常值,将结果通过窗函数回归分析,并和实际调节数据进行对比。【结果】结果表明,对比传统孤立森林算法,改进孤立森林算法在各个调节值下的识别率均达到95%。当调节值为0.1 mm时,本模型的预测变形为0.105 mm,误差为5.0%;当调节值为0.6 mm时,本模型的预测变形为0.610 mm,误差为5.0%;当调节值为1.0 mm时,本模型的预测变形为1.010 mm,误差为1%;当调节值为1.4 mm时,本模型的预测变形为1.424 mm,误差为1.7%。【结论】根据模型预测结果和实际调节值对比,本模型的异常值识别率在各加载位移下均大于96%,误差率在各加载位移下均小于5%,能有效识别测斜仪在高拱坝施工期位移监测中的“突变”异常值。

[Objective]The inversion deformation during the construction of high arch dams is one of the causes of cracks during the construction period.When continuous inclined measuring equipment is used to monitor the deformation during the construc-tion period,the abnormal values caused by the construction process such as layered pouring need to be detected and judged.[Methods]Combining with the required accuracy,this paper designs a deformation monitoring test to simulate the working condi-tion and“sudden change”process of arch dams.For the test result,an outlier detection model based on the isolated forest algo-rithm is proposed.The data are pre-processed by wavelet transform,and the outliers are detected by the improved isolated forest algorithm after noise reduction,and the result are analyzed by window function regression and compared with the actual adjust-ment data.[Results]The result show that,compared with the traditional isolated forest algorithm,the recognition rate of the improved isolated forest algorithm reaches 95%at each adjustment value.When the adjustment value is 0.1 mm,the prediction deformation of this model is 0.105 mm with an error of 5.0%;when the adjustment value is 0.6 mm,the prediction deformation of this model is 0.610 mm with an error of 5.0%;when the adjustment value is 1.0 mm,the prediction deformation of this model is 1.010 mm with an error of 1.0%;when the adjustment value is 1.4 mm,the prediction deformation of this model is 1.424 mm with an error of 1.7%.[Conclusion]According to the comparison between the prediction result of the model and the actual ad-justment value,the outlier recognition rate of the model is greater than 96%under each loading displacement,and the error rate is less than 5%under each loading displacement,which can effectively identify the“sudden change”outliers of the inclinometer in the displacement monitoring.