Condition monitoring framework for damage identification in CFRP rotating shafts using Model-Driven Machine learning techniques

Abstract

Real-time condition monitoring (CM) through vibration measurements is instrumental in detecting faults and enabling predictive maintenance for mechanical systems. The accuracy and robustness of a CM application depends among others on the availability of data for different health states which typically requires complete experimental measurements. In this work, a novel CM framework for damage detection and localization is presented and applied to a lab-scale CFRP transmission shaft, using Support Vector Machine models (SVM) trained by data generated through numerical simulations of an Optimal Multibody Dynamics (MBD) model of the physical system under different rotating speeds. First, the MBD model corresponding to the real shaft structure, including support bearings, is developed and optimized using a small number of initial healthy state measurements throughout the operating speed range. The goal is to perform damage identification on different health states by exploiting simulated instead of experimental responses for supervised health state classification. The study specifically addresses damage scenarios like cracks and delamination at specific locations on the shaft. Simulated damaged states, represented by a local reduction of stiffness on the optimal MBD model, enable the generation of labeled training datasets for SVMs. The trained SVMs’ robustness and accuracy are then validated by accurately classifying faults on the physical system, proving the proposed damage detection method’s generalization capabilities and highlighting its potential.