计算机科学
自编码
加权
过程(计算)
质量(理念)
断层(地质)
相关性
人工智能
故障检测与隔离
机器学习
模式识别(心理学)
数据挖掘
算法
人工神经网络
程序设计语言
数学
医学
哲学
几何学
认识论
地震学
执行机构
放射科
地质学
作者
Ziyuan Wang,Chengzhu Wang,Yonggang Li
标识
DOI:10.1016/j.engappai.2024.108051
摘要
The severity of faults and the corresponding solutions in the complex and large-scale modern manufacturing industry are determined based on their impact on product quality. Detecting faults that affect product quality is crucial for minimizing downtime and reducing maintenance costs. However, obtaining real-time indicators of product quality is often challenging. Therefore, in practical production settings, it becomes essential to develop methods for detecting process-quality faults without relying on real-time quality data. Existing process-quality concurrent fault detection methods establish process-variable-quality-variable relationships to design monitoring statistics. Variational autoencoder (VAE) is extensively utilized to disentangle quality-related information from quality-unrelated information by ensuring the independence of the extracted features. However, the current VAE-based model structures lack theoretical support from probability models, and the feature extraction solely from process variables may not fully capture quality indicators. Therefore, we propose a novel VAE based on knowledge sharing and correlation weighting (KSCW-VAE) with theoretical interpretability to realize process-quality concurrent fault detection. First, to extract information from quality variables, a novel VAE model with two input branches is constructed. The EncoderNet branch includes both process variables and quality variables, while the PriorNet branch comprises only process variables. Then, Kullback–Leibler (KL) divergence constraint is introduced to facilitate effective extraction of combined information from process variables and quality variables by leveraging knowledge sharing between EncoderNet and PriorNet. To differentiate quality-related and quality-unrelated information, separate regression networks for process variables and quality variables are constructed by segregating latent variables. In addition, a network weight regularization based on canonical correlation analysis (CCA) is designed to reduce the importance of weak quality-related process variables (WQRPVs). Finally, a monitoring procedure and two statistics are developed for implementing process-quality concurrent fault detection. Extensive experiments are conducted to demonstrate the superiority of the proposed algorithm compared to advanced alternatives.Compared to the baseline model, the proposed method demonstrates an average accuracy improvement of 16.4% to 1% and 32.1% to 1.225% in two respective cases.
科研通智能强力驱动
Strongly Powered by AbleSci AI