Early fault diagnosis of ball screws based on 1-D convolution neural network and orthogonal design

Abstract

Ball screws, the most frequently used mechanical components to transform rotary motion into linear motion, can directly affect the precision and service life of engineering machines. Once the efficiency and accuracy of ball screws degrades, the performance and safety of machines are hard to guarantee. Conventional fault diagnosis researches of ball screws are mainly focused on ordinary faults such as preload loss and wear, and lack of the researches on early faults such as lubrication degradation which may progress into the ordinary faults. Additionally, the fault diagnosis models proposed in previous studies divide the fault diagnosis into two separated stages: feature extraction and fault classification, which prevents the usage for real-time applications. The specifically designed algorithm in features extraction stage may be also not workable on other objects. To tackle these drawbacks, this paper proposes a highly accurate early fault diagnosis model of ball screws based on a state-of-the-art deep learning technique, called One-Dimensional Convolutional Neural Network (1-D CNN). Experiments simulating the lubrication degradation of ball screws are specially designed for the early fault diagnosis of the ball screws. Moreover, a concise and efficient approach based on orthogonal design is exploited to scientifically obtain the optimal parameters of the 1-D CNN. The results of a case study verify the superiority of the proposed method in establishing a highly accurate 1-D CNN based fault diagnosis model.