Deep generative model with time series-image encoding for manufacturing fault detection in die casting process

作     者：Song, Jiyoung Lee, Young Chul Lee, Jeongsu 

作者机构：Korea Inst Ind Technol Smart Liquid Proc R&D Dept 113-58 Seohaean Ro Siheung Si Gyeonggi Do South Korea Gachon Univ Dept Mech Smart & Ind Engn 1342 Seongnamdaero Seongnam Si Gyeonggi Do South Korea 

出 版 物：《JOURNAL OF INTELLIGENT MANUFACTURING》 (智能制造业杂志)

年 卷 期：2023年第34卷第7期

页      面：3001-3014页

核心收录：

学科分类：08[工学] 0802[工学-机械工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Korea Institute of Industrial Technology as "Development of intelligent root technology with add-on modules'' [KITECH EO-22-0005] 

主　　题：Fault detection Generative adversarial network Variational autoencoder Time series data Image encoding Semi-supervised learning 

摘      要：The increasing demand for advanced fault detection in manufacturing processes has encouraged the application of industrial intelligence based on deep learning. However, implementing deep learning technology at actual manufacturing sites remains challenging because the data acquired during the manufacturing process are not only unlabeled but also imbalanced time series data. In this study, we constructed semi-supervised manufacturing fault detection methods to deal with the imbalanced time series data obtained from manufacturing applications, based on recently proposed deep generative models: variational autoencoder-reconstruction along projection pathway (VAE-RaPP) and Fence generative adversarial network (Fence GAN). To apply a semi-supervised learning algorithm, 1000 labeled samples of good product were prepared. The deep generative models learned the features of good product from these labeled samples during training. Consequently, the model was sufficiently trained to distinguish good and defective product in unlabeled samples. Additionally, we converted the time series data acquired during the manufacturing process into images to improve the feature extraction capability of deep neural networks based on three encoding methods: Gramian angular difference field (GADF), Markov transition field (MTF), and recurrence plot (RP). The performance of these methods was then compared using four evaluation indicators: area under the receiver operating characteristic (AUROC), average precision (AP) score, precision-recall (PR) curve, and accuracy. The VAE-RaPP exhibited outstanding performance in all types of encoding methods when compared with the Fence GAN. This research provides a novel approach that combines the encoding of time series into images and deep generative models for manufacturing fault detection.