Yusuf Altintas

The effects of wear, preload loss, and missing balls on the dynamics of ball screw drives in machine tools are modeled and incorporated into the finite element model of the drive assembly for condition monitoring. The contacts between the ball–nut and ball–screw are modeled using Hertzian springs, whose stiffnesses vary as a function of the worn contact area. These contact stiffnesses are then transformed to the finite element nodes on the nut and screw. The frequency response functions at the motor shaft and table, which can be measured by commercial computer numerical control (CNC), are predicted for various faults at different positions of the table. The experimentally validated model demonstrates that the faults primarily affect the first coupled torsional-axial mode of the ball screw drive and can be utilized for automated condition monitoring of ball screw drives.

The primary goal of the paper is to monitor the health of the spindle in machine tools to ensure optimal performance and reduce costly downtimes. Spindle health monitoring is essential to detect wear and cracks in spindle bearings, which can be challenging due to their gradual development and hidden locations. The proposed approach combines physics-based modeling and data-driven techniques to monitor spindle health effectively. In Part I and Part II of the paper, mathematical models of bearing faults and spindle imbalance are integrated into the digital model of the spindle. This allows for simulating the operation of the spindle both with and without faults. The integration of fault models enables the generation of vibrations at sensor locations along the spindle shaft. The generated vibration data from the physics-based model are used to train a recurrent neural network-based (RNN) fault detection algorithm. The RNN learns from the labeled vibration spectra to identify different fault conditions. Bayesian optimization is used to automatically tune the hyperparameters governing the accuracy and efficiency of the learning models during the training process. The RNN classifiers are further fine-tuned using a small set of experimentally collected data for the generalization of the model on real-world data. Once the RNN classifier is trained, it can distinguish between different types of damage and identify their specific locations on the spindle assembly. The proposed algorithms achieved an accuracy of 98.43% on experimental data sets that were not used in training the network.