| dc.description.abstract | As an important component of precision machines and machin tools, it is important to monitor the preloads of ball screws, and its preload degradation or backlash exists will affect the positioning accuracy. In order to address this issue, This study propose a small sample size augmentation technique and Deep Transfer Learning (DTL) method for ball screw preload condition monitoring, the Generative Adversarial Network (GAN) model to deal with ball screw preload data shortage problem, a deep learning-based domain adaptation method to improve the pre-training model accuracy for different loading conditions of the ball screw feeding system. The scope of this study is divided into four parts–(1) Experimental planning and data collection: The ball screw feeding system is operated at constant speed, and the vibration signals of two different load are collected(30 kgw weight and ball screw load module)for three types of preload states (4%, 1% preload and 12 µm backlash), the 30 kgw weight load collects 2500 data for individual preload states and ball screw load module collects 500 data, as source domain and target domain data respectively; (2) Analysis of vibration signals: The relationship between individual preload states and ball screw vibration was investigated by spectrum and envelope analysis, analyze the characteristic frequencies and causes of vibration signals,when the preload degradation or backlash exists, in addition to the ball pass frequency caused by ball impact there are several sideband components, i.e., the amplitude modulation phenomenon caused by ball impact when the ball screw is running; (3) Data augmentation model building: To investigate the performance of the GAN framework, the model is trained using 500 datas the 30 kgw weights load of 4% ball screw preload vibration signal,and created the generated dataset.We find original the 30 kgw weights load of 4% preload vibration signal features and used to create artificial simulated signals, and then three classification models were used to compare the performance of the data augmentation method ,These models are Convolutional Neural Network (CNN), Multilayer Perceptron (MLP) and eXtreme Gradient Boosting (XGBoost); (4) Deep transfer learning model building: Propose a domain adaptation transfer method to solve the problem of old and new data bias, use the CNN model as the transfer object to fine-tune and add an adaptation layer, calculate the distance between two distributions. Compare the effects of polynomial and Gaussian kernel functions on the stability and accuracy of the model, and then the validity of the transfer learning model is verified by using the dimension reduction method. From the results, the data generated by the proposed GAN model does not contain all signal features. However, the accuracy of the model is improved to 99.5% after the dataset be augmented by blending the generated data; the accuracy of the domain adaptation method reaches 98.6% for source domain and 94.7% for target domain, which can effectively classify the ball screw preload states in the source and target domains. | en_US |