Hybrid Modeling of an Induction Machine to Support Bearing Diagnostics

In this article, a novel hybrid model of an induction machine is proposed that can emulate the response of a machine with a faulty bearing. The idea behind developing such a topology is to have the response quite close to that from a real asset while keeping it computationally efficient. The aim is to develop an accurate and efficient model, akin to digital twins, which have the potential for real-time operation. Therefore, the model is divided into two parts. One is a physics-based model that takes fundamental equations and motor construction parameters to yield an intermediate response. All the major parameters are taken into account such that the fundamental component comes quite close to that of the real asset and the bearing fault signature comes in the same order. These signatures are quite small and some small parasitic effects or the assumptions taken while simplifying the model might not impact the fundamental component that much but they alter the signature's amplitude quite significantly. One way is to model all the parasitic effects, which might increase the computation effort significantly. Another way is to take all the parasitic effects altogether and bridge the difference using a statistical approach which is developed using experimental data. Therefore, the current measurements were taken for several bearings with different fault severity. These measurements are processed and quantified such that the net outcome can express the evolution of the signature with increasing fault severity. The same is done for the data generated using the physics-based model. Finally, the difference in the responses is reduced using the neural network such that it can mimic real-world machine behavior closely. The analytical model followed by statistical adjustment overall is considered a hybrid model. The ultimate goal of this methodology is to generate extensive datasets encompassing diverse operating conditions that can be used further to estimate the health of the bearing and possibly be used for training predictive algorithms to estimate bearing RUL in motors. The proposed methodology is developed for the machine operating at 1000 and 1500 RPM and is validated for three different operating speeds.