Ball screws play a critical role in high-quality precision manufacturing. The use of machine learning and artificial intelligence for the diagnosis of machines’ health status is increasingly pertinent. The processing of big data originating from machine sensors is crucial. However, installing multiple sensors on the object requiring diagnosis may be costly. A sensorless strategy using built-in signals to determine the conditions of a hollow ball screw was deployed. Moreover, we evaluated the most discriminative parameters among fusion sensor signals by using Fisher’ criteria. A support vector machine (SVM) as diagnostic tool was used. In the absence of prominent characteristic features in data, the conventional SVM cannot classify the data well. To address this concern, we constructed a feature engineering for distinguishing features from the raw data to facilitate the SVM classification process well. In addition, we validated the physical phenomenon in realistic ball screw conditions through feature extraction. Experimental results demonstrated the average diagnostic accuracy levels for the ball screw preload, pretension, cooling system, and table payload were 98.91%, 94.08%, 91.69%, and 93.5%, respectively, after feature engineering was applied successfully.
Lame problem for a multilayer viscoelastic hollow ball with regard to inhomogeneous aging
