残余物
计算机科学
人工智能
自编码
机器学习
特征提取
深度学习
特征学习
可靠性(半导体)
特征(语言学)
数据挖掘
质量(理念)
人工神经网络
过程(计算)
模式识别(心理学)
算法
功率(物理)
语言学
物理
哲学
认识论
量子力学
操作系统
作者
Yalin Wang,Jiang Luo,Chenliang Liu,Xiaofeng Yuan,Kai Wang,Chunhua Yang
出处
期刊:IEEE Transactions on Instrumentation and Measurement
[Institute of Electrical and Electronics Engineers]
日期:2022-01-01
卷期号:71: 1-11
被引量:15
标识
DOI:10.1109/tim.2022.3214611
摘要
Deep learning has been widely used in quality prediction of industrial process data due to its powerful feature extraction capability. However, the limitation of deep learning hierarchical feature extraction manner will discard the valuable information related to quality variables in the original data, seriously impairing the reliability and usability of deep learning applications in the industry. The residuals, as the deviations of the actual values from the predicted values of the quality variables, could indirectly reflect this important information. To this end, residual information is introduced into the deep neural networks to effectively guide the feature learning process of each layer. In this paper, a novel layer-wise residual prediction network based on stacked autoencoder (LR-SAE) is developed to obtain better feature representation from raw data and residual information related to quality variables. Based on this, the learned features are more reliable and representative, which could improve the performance of quality prediction. Finally, two industrial examples are applied to verify the effectiveness of the proposed method. Besides, the effects of the residual prediction of each network layer and the final quality prediction are carefully discussed on the proposed method. In two industrial applications, extensive experiments show that the prediction accuracy of the proposed method outperforms the traditional methods.
科研通智能强力驱动
Strongly Powered by AbleSci AI