Structural state classification of 3D frame – a machine learning approach

Abstract

Machine Learning (ML) approach is one of the finest methods in Structural Health Monitoring (SHM). Its capacity in handling huge data and accuracy in predictions has made it a powerful tool in SHM. The data from the sensors monitoring the structure is used for damage detection and identifying the sources of disturbances in the structural system. This study examines the algorithms of machine learning for categorising the various structural states in a three-dimensional frame. Three structural states are taken in the present study for the dynamic analysis: the normal case (UD), the added mass case 1 (DC1), where an extra mass is added at level 1 of the frame, and the added mass case 2 (DC2), where an extra mass is added at level 2. Three machine learning algorithms, viz., the K-Nearest Neighbor (K-NN) algorithm, Support Vector Machines (SVM) and decision trees are used for classification. During dynamic tests, the model is excited with a uniaxial shake table which has a 40 kg payload capacity, and the responses are obtained using accelerometers. To categorise the structural states, the extracted data is split into three sets (set I, set II, and set III) with variations in training and testing data samples and fed through the aforementioned algorithms. It is concluded from the study that the linear SVM is highly suited for set III with 98% accuracy, while fine tree and fine kNN are best suited for set I data with 98% and 97% accuracy, respectively. Also, ML technique has the ability to handle massive data with utmost prediction.