Unsupervised Condition Diagnosis of Linear Motion Guide Using Generative Model Based on Images

Recently, linear motion (LM) guides have been widely used in industrial processes, especially for precise positioning applications. A LM guide typically requires a custom design for specific characteristics of several industrial fields, which is time consuming for manufacturing process; additionally, and the production line with failed LM guides cannot be fixed rapidly. Therefore, to reduce production loss during such periods, it is important to prepare for maintenance and management in case of a fault through a real-time diagnosis of LM guide conditions. Currently, studies on condition diagnosis applying deep learning algorithms are actively being conducted, and actual measured signals in the industrial fields are being used as training data. In this study, the condition diagnosis of LM guides is conducted through a variational auto-encoder (VAE) with convolutional layers. Normal and fault data are measured using an LM guide unit that consists of four LM blocks and one ball screw. The measured signals are converted to spectrogram images through short-time Fourier transform, and only normal data are used to perform network learning. The trained model is applied to classify the normal and fault states, and the reconstruction error is utilized as the evaluation metrics for classification performance. In order to validate the performance, the results are compared to those of a restricted Boltzmann Machine (RBM), and a stacked auto-encoder and a VAE, which does not consist of a convolutional network. Through this study, it is shown that the LM guide diagnosis is possible with a network (VAE with convolution layers) learned with only normal states and the performance is shown to be superior to other networks.