Machine learning-driven advancements in structural health monitoring: comprehensive multi-state classification for three-dimensional structures

Applying Machine Learning (ML) in Structural Health Monitoring (SHM) has proven to be highly effective. ML’s ability to handle large datasets and provide accurate predictions has made it a powerful tool in SHM. This study utilizes ML algorithms to categorize structural states within a three-dimensional frame. Nine structural states are examined in this research for dynamic analysis and classification. Three ML classifiers - Decision Tree, K-Nearest Neighbor (KNN), and Support Vector Machine (SVM) algorithms - were employed for the analysis. Python libraries are used to train and test the data with these algorithms. Dynamic tests were performed, exciting the model using a uniaxial Shake table with a 40 kg payload capacity and recording accelerometer responses. The classifiers’ performance was compared based on various classification metrics, such as accuracy, precision, recall, F1 score, and specificity, using confusion matrices and Receiver operating characteristic(ROC) curves. The study’s primary objective is to classify the structural states of a three-storied building frame through ML algorithms. The decision tree algorithm exhibits exceptional performance, achieving an impressive 94% accuracy rate with a specific 0.80 train-test split for set 1D data. Meanwhile, KNN impressively achieves a 92% accuracy, even with a 0.66 split for set 1 C data, maintaining a consistently high 90% accuracy level at lower splits of 0.50 and 0.33. The transition to three-channel data significantly enhances the decision tree’s accuracy by an impressive 23%, reaching 94%. A consistent 0.67 train-test split consistently yields reliable accuracy across all three algorithms. While the F1 score favours the decision tree (94%) for set 1 data and KNN (93%) for set 2 data, it’s important to note that ROC, Area under the curve(AUC) values, may vary due to class imbalances. This study provides valuable insights into algorithm selection, optimal split ratios, and relevant metrics for an efficient and robust approach to structural health monitoring.