Study On Muitl-modality And Time Delay Industrial Process Modeling Based On Deep Learning Methods

The intelligentization of industrial process modeling,control,optimization and decision-making is the key of realizing the intelligence of industrial process.Among them,process modeling is the basis and prerequisite for intelligent control,optimization and decision-making.For increasingly complex industrial processes,it is hard to obtain accurate mathematical models to describe the energy and material balance relationships of them.In recent years,with the widespread application of supervisory control and data acquisition(SCADA)systems in the industrial process,a large amount of process data has been saved,raising the research interest of data-driven industrial process modeling methods.Multi-modality,time delay,and existing abnormal conditions are common characteristics of complex industrial processes.Traditional data-driven modeling methods are difficult to achieve satisfactory modeling results on these complex industrial processes.This thesis proposes three improved deep learning model based on Recurrent Neural Network(RNN)and Temporal Convolutional Network(TCN)to model the industrial process with one of the three characteristics separately.Then,the synthesis of above three models are proposed to model the industrial process with the above three characteristics.The main work of the thesis is summarized as follows:(1)To solve the problem that the key data points of the muitl-modality industrial process is hard to capture during the model-swtiching phase,an encoder-decoder model based on the attention mechanism is proposed.Firstly,we point out the limitation between the input and output length of the traditional RNN model,and adopt the encoder-decoder architecture.Secondly,considering the dynamic characteristics of the industrial process and the long-term dependency of data,the gated recurrent unit(GRU)is applied as the basic unit of the encoder-decoder.Thirdly,for the key data points during the model-swtiching phase,the attention mechanism is introduced to dynamically capture the key information in the data sequence,and pay more attention to the key information.In the training process,to reduce the impact of the cumulative error of the data sequence,a teaching forcing method is used to ensure that the model converges faster.The effectiveness of the method is verified by a real industrial dataset.(2)Aiming at the modeling of and time delay identification of the time delay industrial process,a controlled recurrent neural network model based on discrete Bayesian optimization is proposed.Firstly,to identify the values of time delay,a controlled recurrent layer is proposed by improving the update form of the hidden state in the original recurrent layer.Then the values of time delay are modeled as parameters in the mask layer and regarded as the hyper-parameters of the model.During the training process,the values of time delay are obtained by discrete Bayesian optimization method instead of back-propagation algorithm.Finally,to ensure that the discrete Bayesian optimization method can obtain the global optimal solution(i.e.,obtain the correct values of time delay)at the minimum loss value,the 2regularization term of the mask layer is added to the original mean square error loss function.The proposed method can not only establish an accurate system model,but also accurately identify the values of time delay.The effectiveness and robustness of the method is proved by the simulation datasets of two industrial systems.(3)Aiming at the problem that the abnormal data in the industrial process is much fewer than the normal data,and the degraded data has no label,a data labeling method based on Pearson correlation coefficient and a class-balance strategy based on large sample undersampling is put forward.On the basis of these two methods,an abnormaly identification network is established using Temporal Convolutional Network(TCN).At first,a method for measuring the similarity of time-series data based on Pearson correlation coefficient is proposed,which calculates the degree of similarity between the unlabeled degraded data and the adjacent normal data and abnormal data respectively.Then a class-balance strategy based on undersampling method is proposed to establish multiple class-balanced subsets,where each subset contains part of normal data and all degraded data as well as abnormal data.We train one TCN on each subset,and use an ensembling method to consider the results of all TCNs as the final classification result during inference process.Finally,an improved focal loss function is proposed since the original focus loss function cannot be applied in our model directly.The reliability of the method is verified by a public industrial dataset.(4)Comprehensively consider the above three problems(the key data points during modal-switching is difficult to capture;there exists time delay;there exists abnormal working conditions,and the abnormal data is much fewer than the normal data),an encoder-decoder model with attention mechanism and dynamic compensation network for abnormal working condition combined with discrete Bayesian optimization is proposed.GRU is chosen as the basic unit of the encoder-decoder model.First,to solve the problem that the key data points is hard to capture during the model-swtiching phase,an attention mechanism is introduced between the encoder and the decoder to dynamically capture and put more emphasis on the key information in the data sequence.To cope with the time delay,a mask layer is introduced before the encoder-decoder model,and the values of time delay are modeled as parameters of the mask layer and regarded as hyperparameters of the model.The values of time delay are not obtained by the back-propagation algorithm,but by discrete Bayesian optimization method.The mask layer corresponding to the correct values of time delay can effectively simulate the input time delay and improve the modeling accuracy.Finally,considering the existing of abnormal condition and the abnormal data is much fewer than the normal data,we first use a time-series data similarity measurement method based on Pearson correlation coefficient to calculate the abnormality degree,and then propose a dynamic compensation network for abnormal working condition based on GRU.This model uses the mean square error loss function with a L2 regular term of the mask layer.The effectiveness of the proposed model is verified by an real industrial dataset.