Damage Identification Method Based On Spatial-Temporal Window Principal Component Analysis And AutoGluon

In recent years,structural failures have occurred from time to time and are extremely harmful.To ensure safety and property as well as to reduce economic losses,most large civil engineering structures have been installed with structural health monitoring systems.However,the structural data collected by the health monitoring system not only can hardly reflect the health status of the structure intuitively but also have problems such as missing data and serious noise,which bring more challenges to the condition assessment of the structure.Therefore,how to use long-term health monitoring data for damage identification is still one of the main difficulties in the field of structural health monitoring.In this paper,we use feature engineering to perform principal component analysis on long-term health monitoring data,from which damage-sensitive features——combined eigenvectors are selected and combined with a new machine learning algorithm framework Auto Gluon to perform damage identification.The results show that the combined features and the ensemble learning algorithm proposed in this paper have higher damage recognition accuracy and stronger robustness than the general PCA methods and traditional machine learning algorithms.The main research contents of this paper are as follows:(1)Based on the finite element model of the planar beam and frame structure,a long-term health monitoring simulated strain data set containing a large number of damage conditions under temperature load is generated.Moving Principal Component Analysis(MPCA)and Double-Window Principal Component Analysis(DWPCA)were used to obtain the first and second eigenvector of the associated damage and to study the influence law of the damage on these eigenvectors.(2)Traditional PCA methods usually consider only the first principal component or eigenvector for damage identification,which has low sensitivity and is difficult to meet the requirements of high precision damage identification.To solve the problems of poor noise resistance and low accuracy of damage recognition in PCA feature engineering away from the sensing location,multiple features,including the first and second feature vectors of MPCA,and the first and second feature vectors of DWPCA,are considered as the input of three machine learning algorithms including Decision Tree,Random Forest,and k-Nearest Neighbor to verify the combined features for different noise environments.The results demonstrate that the combined features are more effective and robust in identifying damage at locations far from the sensor.(3)The input or features of traditional machine learning algorithms have a significant impact on damage recognition.To address this problem,this paper introduces a novel Auto ML framework,Auto Gluon,which is capable of achieving highly accurate and robust algorithms without picking inputs or features by integrating a combination of multiple machine learning models.Therefore,this paper combines the combined features of DWPCA with the novel machine learning framework Auto Gluon to explore the damage localization and quantitative recognition capability of the new method on different structures based on the damage dataset of the plane beam and frame structures,and analyze the reasons for optimizing the recognition effect.The results demonstrate that the method outperforms both the machine learning algorithm with traditional features as input and the classical deep learning algorithm for damage recognition of different structures,and has certain structural migration capability for damage recognition of different structures.