| In modern industrial processes,real-time access to quality variables(key process variables)helps companies control product quality,adjust production strategies and optimize production processes.However,sometimes due to factors such as sensor cost,detection time lag and offline maintenance,some quality variables cannot be efficiently obtained in real time,so soft sensor models of quality variables and auxiliary variables(related process variables)need to be established to supplement them.In recent years,deep learning technology has made achievements in many fields,so there are more and more soft sensor modeling research based on deep learning methods.Recurrent neural networks are commonly used deep learning methods for processing time series data,and recurrent neural networks including classical recurrent neural network unit(Vanilla RNN),long short-term memory network unit(LSTM),and gated recurrent unit(GRU)can be used to build soft sensor models.Generally,deep learning uses an offline learning strategy to train neural networks,and no adjustments are made after training is completed on the training set.If process data representing different working conditions can be collected from industrial processes,soft sensor models can be trained using offline learning strategy commonly used in deep learning.However,in some cases,due to factors such as switching of working conditions and environmental changes,the training set data similar to the real situation distribution cannot be collected,and the just-in-time learning strategy needs to be used for soft sensor modeling.Using the recurrent neural network as the basic network structure,this paper improves the unit structure of the long short-term memory network and similarity measurement in just-in-time learning strategy.The main research contents are as follows:(1)Aiming at the problem that the long short-term memory network cannot generate the corresponding hidden state of the previous time step when the current time step input changes,a soft sensor modeling method based on IF-LSTM is designed.Firstly,the information filtering unit(IFU)is constructed,and the hidden state of the LSTM unit of the previous time step is predicted by the input of the current time step,and the hidden state component is scaled according to the prediction error to achieve information filtering.Secondly,an IF-LSTM unit linked by the information filtering unit(IFU)and the long short-term memory network(LSTM)unit is constructed,and the IFU generates the filtered hidden state according to the current input,and enters it to the LSTM unit to replace its original hidden state,thereby further improving the performance of LSTM.Finally,the IF-LSTM network is constructed by the IF-LSTM unit,and a soft sensor modeling method based on the IF-LSTM network is proposed.Experiments in the sulfur recovery unit and debutanizer column dataset show that the soft sensor model based on IF-LSTM has the best predictive performance compared with other recurrent neural network-based soft measurement models(LSTM and GRU).In the debutanizer column dataset,the IF-LSTM soft sensor model is also the best compared to the experimental results documented in other literature.(2)Aiming at the problem that the preset similarity measurement method in just-in-time learning cannot select effective historical set subsets in different soft sensor modeling tasks at the same time,a just-in-time learning strategy based on square error similarity soft sensor modeling method(SE-JITL)is designed.Firstly,the squared error matrix is calculated in the history set,and its elements are the squared error when training different soft sensor local models on the history set samples and testing them in turn,when the squared error is small,indicating that the corresponding two samples are more similar.Secondly,the similarity measurement model is constructed,any two historical set samples are stitched as input,and the corresponding elements of the square error matrix are used as labels for gradient descent training,and the predicted output of the similarity measurement model after training is square error similarity.Finally,when making predictions,the similarity measurement model input stitched samples of samples from history sets with testing samples,output the square error similarity,and follow the just-in-time learning soft sensor modeling process to realize the prediction output of the test samples.Experiments show that this method can effectively select the subset of history sets in the soft sensor tasks of both sulfur recovery unit and debutanizer column dataset.Compared with the just-in-time learning soft measurement modeling method based on Euclidean distance and cosine similarity,the similarity measurement model in SE-JITL can be trained by gradient descent and has stronger adaptability. |
