基于MetaOD模型选择的岩土工程数据异常检测方法

岩土工程现场及室内参数测试数据是工程施工、设计、评价的基础。异常数据的存在往往会误导施工、设计等参数的确定, 数据异常检测是确保工程安全可靠的最基本但极为重要的工作。针对传统异常检测算法没有模型选择这一过程而导致检测的盲目性, 提出了基于元学习的异常检测算法(meta-learning outlier detection, MetaOD)和数据挖掘算法相结合的异常检测模型体系。该体系首先根据数据的特点选择适合不同数据类型的初始模型类型及其参数, 并对选择出的同类型算法的参数进行求均值处理; 然后再采用遴选出的算法进行数据异常诊断, 进而提高异常检测的准确性。为了评估模型的有效性, 采用加州大学欧文分校提出的机器学习检验数据集(glass数据集)进行检验分析。结果显示, 采用该模型体系进行异常检测时查准率达到96.41%, 远高于其他检测算法。最后, 应用该模型体系对澳门花岗岩单轴抗压强度数据集和均昌隧道的地下水位监测数据进行了异常检测分析, 并分别识别出9个和10个异常点。 

Outlier detection method for geotechnical engineering based on MetaOD model selection

The geotechnical engineering field and indoor parameter test data are the foundation of engineering construction, design and evaluation. The existence of abnormal data often misleads the determination of parameters such as construction and design. Data anomaly detection is the most basic but extremely important task to ensure the safety and reliability of a project. Aiming at the blindness of detection due to the lack of model selection in traditional anomaly detection algorithms, this paper proposes an anomaly detection model system based on a combination of meta-learning outlier detection (MetaOD) and data mining algorithms. The system first selects the initial model class and its parameters suitable for different data types according to the characteristics of the data, averages the selected parameters of the same type of algorithm, and then uses the selected algorithm to diagnose data anomalies, thereby improving the anomaly accuracy of detection. To evaluate the effectiveness of the model, the machine learning test dataset (glass dataset) proposed by the University of California Irvine, is used for test analysis. The results show that the accuracy rate of anomaly detection using this model system reaches 96.41%, which is much higher than that of other detection algorithms. Finally, the model system is applied to the uniaxial compressive strength dataset of the Macau granite and the groundwater monitoring data of the Junchang Tunnel to carry out anomaly detection and analysis and to identify 9 and 10 abnormal points, respectively.