A mothed of improving identification accuracy via deep learning algorithm under condition of deficient labeled data

• Published in: Chinese Control Conference pp. 2281 - 2286
• Main Authors: Jian-Guo Wang, Zhi-Duo Cao, Bang-Hua Yang, Shi-Wei Ma, Min-Rui Fei, Hao Wang, Yuan Yao, Tao Chen, Xiao-Fei Wang
• Format: Conference Proceeding
• Language: English
• Published: Technical Committee on Control Theory, CAA 01.07.2017
• Subjects: Automobiles; Biological neural networks; Controlled Auto Regressive; Data models; Deep Learning; Finite Impulse Response; Machine learning; Mathematical model; System identification; Training
• ISSN: 1934-1768
• DOI: 10.23919/ChiCC.2017.8027697

In industrial process, some important variables such as quality index, efficiency index and concentration of product components are difficult or even impossible to be measured directly due to the limitation of technology. This phenomenon leads to few labeled data and plenty of unlabeled data. Traditional identification method for controlled auto regressive (CAR) model usually cannot deal with unlabeled training data. As a result, these traditional identification methods may receive poor identification precision or even cannot work entirely. To solve the problems above, this paper proposes a new identification method based on deep learning (DL). Firstly, the CAR model is transformed into finite impulse response (FIR) model solve the problem of lack of autoregressive part; Secondly, autoencoder of deep learning make full use of unlabeled data to pretrain the model; Thirdly, small amount of label data is used for fine-tuning. As a semi-supervised learning method, deep learning can be able to extract more information from unlabeled data than traditional supervised learning method. The results show that the proposed method can acquire higher identification accuracy than BP neural network.