The Research On Data-driven Batch Process Online Monitoring And Quality Prediction Method

As one of the most important processing methods in modern industrial production,batch process has the advantages of small batch production,rapid cutting of multiple varieties and high added value.It has been widely used in many fields such as fine chemical industry,biopharmaceutical,polymer reaction,semiconductor manufacturing and food processing.With the advent of Industry4.0 era,the modern industrial system is developing towards more complex and larger scale.Once the process fails,it is likely to cause incalculable social harm and economic losses.In addition,with the improvement of national economic level,people’s requirements for product quality are also rising.Therefore,the study of reliable process monitoring and quality prediction model is of great significance for ensuring the safe and stable operation of batch process and improving product quality,which has become an urgent problem to be solved in modern industrial process.The complex and varied production mode of batch process makes it difficult to establish accurate mechanism model.However,the extensive deployment of intelligent sensing devices and the rapid development of communication technology make the data-driven approach show unique advantages in batch process monitoring and quality prediction.The existing methods have made some idealized assumptions and constraints on batch processes,such as the process variables do not have autocorrelation,the production process runs in a single condition and the process has rich production data.However,the actual batch process has the problems of strong dynamic,nonlinear,multi-stage and lack of quality feedback,which does not conform to these assumptions,which makes the monitoring performance of the existing methods decline and even cannot be applied.Therefore,on the basis of in-depth analysis for the complex characteristics of batch process,this dissertation carries out systematic research around the dynamic nonlinear feature extraction,fine process monitoring,multistage quality prediction,batch end quality prediction,small batch quality prediction and other issues.The main research contents and contributions are as follows:(1)Aiming at the problem of batch process monitoring with the coexistence of nonlinear,dynamic and non-Gaussian characteristics,a process monitoring method based on the global enhanced multiple neighborhoods preserving embedding algorithm is proposed.By selecting distance neighbor,time neighbor and angle neighbor for process data,GEMNPE algorithm fully represents the similarity of data in time and space,and takes into account reconstruction errors,order information of three kinds of neighbors and global structure based on variance to construct an enhanced local-global objective function.It can solve the dynamic problem of process data while extracting the local and global structural characteristics of process data.Considering the assumption that the low-dimensional features obtained by GEMNPE do not satisfy the Gaussian distribution,a monitoring statistic is constructed based on support vector data description to eliminate the adverse effects of non-Gaussian characteristics on the monitoring performance.The monitoring performance of the proposed method is verified by numerical examples and penicillin fermentation process.The results show that the proposed method can fully explore the dynamic nonlinear characteristics of the process,so that the process faults can be detected timely and accurately.(2)Aiming at the problem of fine monitoring of multi-stage batch process,a quality-related fault detection method based on multiway weighted elastic network is proposed.First,by introducing the angle information of process data,an improved affinity propagation clustering algorithm is proposed to identify multiple stages of batch processes without relying on prior knowledge.Second,the MWEN regression model is established at each stage,and each stage is further divided into quality-unrelated subspace and quality-related subspace.Kernel density estimation is used to measure the contribution degree of each element in each subspace to the fault,and different weights are assigned to each element according to the contribution size,so as to enhance fault features and eliminate irrelevant features such as noise.Finally,SVDD-based monitoring indexes are established in the quality-unrelated subspace and quality-related subspace respectively to realize the dual monitoring of production safety and quality anomalies in batch processes.The effectiveness of the proposed method is verified by the penicillin fermentation process and hot strip rolling process.The results show that the IAP algorithm can accurately divide the batch process into multiple stages,and the MWEN method can effectively enhance the fault features,so as to quickly detect the occurrence of fault and further diagnose whether the fault will affect product quality.(3)Aiming at the quality prediction problem of multi-stage batch process,a quality prediction method based on multi-stage fusion regression network is proposed.First,IAP algorithm is used to divide batch process into multiple stages.Second,considering the input reconstruction errors and quality prediction errors,a stacked input-isomorphic and quality-driven autoencoder is developed for each stage,which can fully extract the nonlinear quality-related features of each stage while reducing the cumulative input loss.Finally,the self-attention mechanism is used to fuse the quality correlation features of each stage to obtain the global quality correlation features including the features of each stage and the potential correlations among the stages.The global features are input into a fully connected regression layer to realize the quality prediction of the multi-stage batch process.The effectiveness of the proposed method is verified by penicillin fermentation process.The results show that the proposed method can obviously improve the quality prediction performance of multi-stage batch process.(4)Aiming at batch end quality prediction of batch process,a batch end quality prediction method based on multi-resolution feature selection is proposed.First,a multi-resolution feature selection method based on grey wolf optimization algorithm is proposed.The low resolution value of each variable is optimized by grey wolf algorithm to reduce the redundancy of data in sampling time,and the influence of irrelevant variables is eliminated by setting a reasonable resolution threshold.Second,considering the accurate extraction of nonlinear dynamic quality-related features,an attention-weighted supervised LSTM network is proposed.By using the attention mechanism,different weights are assigned to each input feature to enhance the quality-related features and eliminate the quality-unrelated features.In addition,quality variables are input to LSTM units to guide the learning process of dynamic nonlinear quality-related features.Finally,the predicted value of batch end quality is obtained by an additional full connection layer.The effectiveness of the proposed method is verified by the injection molding process and penicillin fermentation process.The results show that the proposed method can greatly reduce the redundancy of input features on the time scale,and improve the accuracy and robustness of the batch end quality prediction model.(5)Aiming at the problem of batch end quality prediction for batch process with lack of data,a new quality prediction method based on transfer learning is proposed.First,the multi-stage characteristic of batch process is considered in feature selection,and the multi-stage and multi-resolution feature selection based on grey wolf optimization is implemented to obtain more simplified input features.Second,in order to improve the quality related feature extraction ability of the model,a dual attention and bidirectional supervised LSTM network is proposed.The network designs an attention module at the input end to pay more attention to the quality-related input,and establishes a supervised bidirectional LSTM network to consider the history and future dynamic information of process data to more accurately describe the dynamics of batch data.In addition,the output features of BSLSTM at each moment are weighted with attention to solve the feature loss caused by using only the output feature of the last moment for quality prediction.Finally,the modeling knowledge of similar old process is transferred to the new process by transfer learning,so as to realize the quality prediction modeling of process with small batch.The effectiveness of the proposed method is verified by penicillin fermentation process.The results show that the proposed method can establish a more accurate quality prediction model for the new production process with small batch data.