Font Size: a A A

Semi-Supervised Soft Sensor Based On Deep Learning

Posted on:2020-06-09Degree:MasterType:Thesis
Country:ChinaCandidate:H LiFull Text:PDF
GTID:2392330572969969Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
In order to track the quality of products,monitor the state of the process and control the process stably and reliably,it is very important to detect the key quality variables in the industrial process.However,due to the complex and harsh actual production environment and limited technical level,some key quality variables can not be real-time detected in the actual production process.Therefore,soft-sensing technology can establish the mathematical model between the key quality variables and the process variables which are easy to measure,so that the real-time prediction and estimation of the key quality variables can be realized.The traditional soft sensor modeling process often assumes that the input and output samples collected are one-to-one correspondence.However,in the actual industrial process,the samples of key quality variables are often obtained through experiments or analysis of precision instruments which are costly.Conventional soft sensor models can only use the samples containing key quality variables and abandon a large number of unlabeled samples(samples containing only process variables).In this way,not only we build an inaccurate soft sensor model,but also waste some useful information contained in unlabeled samples.A semi-supervised soft sensor modeling method based on deep learning is proposed to solve the problems of limited labeled samples,non-linearity and dynamics in industrial process.The main research includes:(1)A semi-supervised soft sensor modeling method based on semi-supervised Auto-encoder is proposed to solve the problems of limited labeled samples and non-linearity.This method is improved on the basis of conventional soft sensor modeling method which is applied deep learning.The traditional methods usually adopt unsupervised learning using the unlabeled samples and then conduct supervised learning using the labeled samples.Our new method combines the Auto-encoder and the neural network together.Finally,the validity of the method is verified on the debutanizer column data set.(2)A semi-supervised soft sensor modeling method based on long-short-time-memory(LSTM)is proposed to solve the problems of limited labeled samples and dynamic.By LSTM model,not only a large number of unlabeled samples can be used,but also this model carn discover the correlation characteristics between samples Attention mechanism is added to the original LSTM to promote the bad performance when the length of LSTM sequence becomes long.On this basis,a convolution layer is added to further extract the local features between samples.So the LSTM-CNN-Attention model is proposed.Finally,each model is validated on the data set of carbon dioxide absorption column.(3)A semi-supervised soft sensor modeling method based on active neural network is proposed to solve the problems of non-linearity and scarce labeled samples xwhich makes it impossible to build accurate model.Adding labeled samples is the most direct way to solve this problem.Active neural network method can select the most valuable labeled samples from a large number of unlabeled samples set and then these unlabeled samples are labeled by experts or laboratory.Then,we can update labeled samples set by the selected new samples.Using more labeled samples to model can lead to more accurate model.Active learning neural network can achieve maximum effect improvement of soft sensor model through the minimum unlabeled samples labeling Finally,the validity of this method is verified on the data set of debutanizer.
Keywords/Search Tags:Soft Sensor, Semi-supervised Learning, Deep Learning, Data Imbalance, Non-linearity, Dynamic Nature
PDF Full Text Request
Related items