Publication: Application of machine learning in predicting the critical vibration of mechanical structure
Loading...
Date
2024-07-01
Authors
Tan Chu Heng
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This project successfully executed various processes to extract dynamic parameters and developed a predictive model for structural health monitoring using machine learning techniques. Experimental Modal Analysis (EMA) was utilized to extract the dynamic parameters of the target structure, derived from the mobility graph at a dominant excitation frequency of 24 Hz, yielding dynamic mass md = 6.697490265 kg, dynamic damping cd = 1188.225173 kg/s and dynamic stiffness kd = 299777.9544. These parameters enabled the construction of a MATLAB Simulink model to log sufficient data for machine learning applications aimed at predicting the structural response under excitation. Operational Deflection Shape (ODS) analysis was performed by exciting the structure using an imbalance motor, recording its actual response, resulting in 201 time-domain samples from the excitation domain and a maximum vibration amplitude of 4.34 m/s². The health status of each sample was benchmarked at 1 dB above the mean of the maximum amplitude at 2.19 m/s²,. Benchmarking allowed classifying samples’ status as healthy or unhealthy based on this threshold. Machine learning models using Support Vector Machine (SVM) and k-Nearest Neighbors (k-NN) classifiers were developed and optimized using Bayesian Optimization, with Principal Component Analysis (PCA) employed to mitigate multicollinearity among the features. Optimal hyperparameters for the SVM model were 𝑪𝑷𝑪𝑨 = 12.9334 and γPCA = 2.9543, while for the k-NN model they were kPCA = 1 and dPCA =Minkowski distance. Cross-validation results demonstrated that the SVM classifier outperformed the k-NN classifier, with the SVM model achieving a mean accuracy of 98.57% after 5-fold cross-validation, compared to 97.86% for the k-NN model. The smaller kPCA value indicated potential overfitting in the k-NN model. Thus, the SVM model is concluded to be more robust, exhibiting consistent performance across varying datasets.