Machine learning based condition monitoring for gear transmission systems using data generated by optimal multibody dynamics models

Condition monitoring (CM) of mechanical systems such as gear transmissions can be performed with vibration measurements and processing of the recorded signals for identification of possible faults. 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. The presented work contains a novel CM scheme using a Convolutional Neural Network (CNN) trained by data generated through numerical simulations of a Multibody Dynamics (MBD) system. The goal is to perform damage identification on different health states by exploiting simulated instead or experimental responses for supervised health state classification. First, a MBD model corresponding to the real system is developed and optimized using experimental measurements of the healthy mechanical structure. Then, data is generated by the MBD model using an uncertainty simulation repetitive load case algorithm to account for various model parameters inaccuracies. The simulated data is used after to train a supervised CNN classifier which is finally validated on vibration measurements of the physical system. The classifier is shown to be capable of generalizing to the experimental damages, proving therefore the potential of the model-based proposed framework for CM. The presented methodology may find application in cases where experimental measurements are difficult or nearly impossible to acquire. CM was performed on the benchmark gear transmission systems for different rotation speeds and the limitations are discussed.