为了减轻大坝安全监测数据异常识别的数据处理压力,解决传统方法难以辨别非最值异常点的问题,提出利用卷积神经网络(CNN)识别大坝安全监测数据异常模式。监测数据过程线的周期性及异常值的显著差别使CNN得以发挥图像分类功能,分别将存在单个突跳点、无异常、存在震荡段、台阶、多个突跳点、台坎的监测数据过程线作为6类图像,人工生成65 000张训练数据及6 500张测试数据,6类图像的数量比为1∶1.5∶1∶1∶1∶1。利用CNN对混合6种过程线图像的测试数据集进行图像分类,总体准确率为0.973 1,且6种图像的准确率都至少为0.93。进一步对CNN进行改进,构建CNN监测数据异常识别模型,增加数据异常位置搜索功能;模型输入为监测数据过程线图像,输出为图像编号、图像类别及异常位置。研究成果有助于实现大坝自动、及时预警,及时了解大坝安全状况。

Traditional methods have difficulties in identifying the non-extreme value outliers in monitoring data of dam safety. To alleviate the pressure of data processing, we propose to use convolutional neural network (CNN) to identify the anomalies. The periodicity of process lines of monitoring data and the significant difference in outliers allow CNN to classify the process lines of monitoring data as six categories of images: process lines with single abrupt jump point, with no anomaly, with multiple abrupt jump points, with oscillating segments, with steps, and with berms. A total of 65,000 training data images and 6,500 testing data images are artificially generated. The ratio of the number of six types of images is 1∶1.5∶1∶1∶1∶1. The overall accuracy of CNN in classifying the mixed six process line images is 0.973 1, and the accuracy for each category is above 0.93. Moreover, we further improve the CNN and build an anomaly identification model by adding a function of searching the position of data anomalies. The input of the model is the process line image, while the output is the image number, image category and anomaly position.