A revisit for the diagnosis of the hollow ball screw conditions based classification using deep learning

• Published in: Measurement and control (London) Vol. 55; no. 9-10; pp. 908 - 926
• Main Authors: Huang, Yi-Cheng, Chuang, Ting-Hsueh, Lin, Chia-Jung
• Format: Journal Article
• Language: English
• Published: London, England SAGE Publications 01.11.2022; Sage Publications Ltd; SAGE Publishing
• Subjects: Accuracy; Artificial intelligence; Artificial neural networks; Ball screws; Classification; Datasets; Deep learning; Feature extraction; Graphics processing units; Machine learning; Mathematical models; Multidimensional data; Object recognition; Parameters; Sensors; Support vector machines; Three dimensional models; support vector machine; long- and short-term memory; convolutional neural network; hybrid neural network; Hollow ball screw
• ISSN: 0020-2940; 2051-8730
• DOI: 10.1177/00202940221092040

Hollow ball screws play a vital role in high-quality precision manufacturing, in which sensors are used to obtain useful data. The use of artificial intelligence to determine the condition of machines is a major trend, and installing multiple sensors on relevant objects may be prohibitively expensive. This study compares a machine learning method based on a support vector machine (SVM) with deep learning methods. In the machine learning strategy, built-in signals from internal parameters are used to determine the condition of the hollow ball screw. Rather than using a graphics processing unit (GPU), feature engineering and then Fisher’s score were applied to determine the most representative parameters from fusion sensor signals to perform SVM classification. For the deep learning method, a GPU was employed instead of data cleaning; a long- and short-term memory (LSTM) network and hybrid convolutional neural network (CNN) and an LSTM network were used to automatically extract effective features and then perform classification. Thus, the deep learning approach enabled classification of one-dimensional (1D) and multidimensional data sets, through which users can obtain suitable configurations for various sensors according to cost and other requirements for classification accuracy. We compared the accuracies of the LSTM, CNN–LSTM, and SVM models by using multiple datasets. The 3D LSTM model with current, rpm, and additional linear scale signals provided good synergy for feature extraction and resulted in the maximum accuracy close to the highest accuracy of the supervised fine-tuned SVM model by using current signal. Datasets that are unrepresentative according to FE analysis may be useful in deep learning. The advantages of using a deep learning model without feature extraction relative to those of SVM learning was investigated.