Performance Evaluation of One-Class Classifiers (OCC) for Damage Detection in Structural Health Monitoring

Structural Health Monitoring (SHM) has become an area of continuous research with the ever-increasing demand for the safety of civil structures. The damage in civil structures can be detected using multimodal data from sensors, which presents instances of both damaged and undamaged data. The availability of damaged data in real life is difficult to obtain from a healthy structure and hence the problem of damage detection needs to be attempted using normal healthy data, and it becomes synonymous with the anomaly or novelty detection. One-Class classifiers work on the principle that the abundance of healthy data can be used to model an envelope of conditions which, if violated by any data instance, can be termed as damage or outlier detection. We have attempted an array of classifiers on a benchmark structure dataset (IASC-ASCE) both from supervised and unsupervised machine learning domain and propose a comparison between their success rates in determining damage in civil structures. We used classical techniques such as One-Class Support Vector Machines (OC-SVM), One-Class Isolation Forest (OC-IF), One-Class K-means clustering (OC-KMC), One-Class K-nearest neighbors (OC-KNN), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), One-Class Principal Component Analysis (OC-PCA), Local Outlier Factor (LOF) and One-Class Gaussian Distribution (OC-GD). These techniques were tested on the IASC (International Association for Structural Control)–ASCE (American Society of Civil Engineers) SHM benchmark problem for a range of noise levels and range of force intensities to cover wide variations in the generated dataset using MATLAB based simulation. Our study helps us conclude that OC-SVM, Isolation Forest, and OC-PCA are the most robust algorithms for the anomaly detection task.