Research On Fault Diagnosis And Prediction For Multi-stage Batch Process

Batch process is one of the most important processes in modern industrial field.With the development of advanced sensors and measurement technologies,a large number of batch production data have been accumulated.Effective fault diagnosis and prediction methods can guarantee the safety and improve the quality of products.Batch process have strong multi-stage characteristic,the relationship between process variables does not change linearly with time variables.It is closely related to the change of process reaction principles and process characteristics.The multi-stage characteristic is an important feature that makes the batch process complex and difficult to construct fault diagnosis and prediction models.Therefore,reasonable and effective stage division is an important prerequisite for fault diagnosis and prediction.According to the strong non-Gaussian,nonlinearity,time sequence and dynamic properties of the data,this paper proposes a multi-stage batch process fault diagnosis and prediction model,the main contents are as follows:(1)Multi-stage division and data processing of batch processAccording to the different mechanism characteristics of each stage in batch process,the K-means method is used to divide the data into stages,and the stage data is used as the input of each model.The division of stages lays the foundation for fault diagnosis and prediction.In view of the non-Gaussian and high-dimensional characteristics of batch data,KECA method is used to reduce the dimensionality of different batch data to improve the efficiency of fault diagnosis and prediction.(2)Multi-stage fault monitoring of batch process based on MBA-MH-SVDDIn view of the strong nonlinearity and non-Gaussian of batch process data,considering the difficulty of constructing stable control limits during fault monitoring,a fault monitoring method based on support vector data description is proposed.Based on the stage division,a multi-boundary hypersphere control limit MH is defined for fault monitoring,and the modified bat algorithm is used to optimize the parameters of SVDD,and the MBA-MH-SVDD is constructed in each stage.Substitute random batches into the monitoring model to determine whether its status is normal and realize online fault monitoring.(3)Multi-stage fault diagnosis of batch process based on Ada Boost-DSIn order to analyze the type of fault batches,considering the differences in the data mechanism characteristics of each stage and the lack of diagnostic performance by a single classifier,a multi-stage ensemble learning fault diagnosis model based on Ada Boost-DS is proposed.After stage dividing and dimensionality reducing,the ensemble learning method is used to integrate the weak classifier decision stump.Select the best number of weak classifiers and construct the Ada Boost-DS diagnostic model in each stage.The random batch is preprocessed and substituted into the model for fault diagnosis.This method combines the advantages of the ensemble learning algorithm and the classification algorithm,and it can achieve fault diagnosis effectively.(4)Multi-stage fault prediction of batch process based on MBA-NARXAccording to the time series and dynamic properties of batch process data,considering the difficulty of prediction through stage data,a NARX neural network with high adaptability to time series and dynamic properties is introduced to conduct multi-stage prediction research.First,construct a prediction model in each stage,and use the intelligent algorithm MBA to optimize the number of hidden nodes in the network.This method can build a prediction model based on historical data and realize online fault prediction.To verify the validity and reliability of the above algorithm model,the penicillin fermentation process and semiconductor etching process are used as examples.And the case study is conducted from simulation data and actual production data,which enrich the research results of batch process fault monitoring,diagnosis and prediction,and reveal the necessity and possibility of further research.