Industrial Process Modeling And Monitoring With Limited Data

With the advent of the Industry 4.0,a new industrial revolution is happening in manufacturing,creating more challenges and opportunities for modern industry.In order to effectively promote the productivity and competitiveness,factories are achieving larger production scale,more complex production process and higher levels of automation.As a result,the possibilities of fault occurrence and quality fluctuation are greatly increased.Data-driven process modeling method,which covers fault detection,fault diagnosis,key indicator prediction,etc.,is of great significance in guaranteeing the safe operation status,ensuring the product quality and reducing power consumption.It requires neither physical model nor expert experience,the model is established by mining useful information from the adequate process data.In the era of big data,a large amount of process data is stored in the industrial field,however,data insufficiency can not be avoided in some specific scenario we concern about.Faced with the problem of local data insufficiency,how to model the process timely and accurately is an urgent problem.This paper attempts to solve the above problem from two perspectives of synthetic data generation and process image mining,which makes up for the lack of process data and realizes accurate modeling and analysis.The main research contents of this paper are organized as follows(1)Aiming at the insufficiency of unlabeled data,a variational autoencoder based data synthesis method is proposed.At the stage of generative model training,there are too few data in the target region to fully train the variational autoencoder,samples from other source regions are consequently added into the training set,effectively utilizing the similar data characteristics between the target region and the source regions.At the stage of data generation,synthetic data are generated for the target region and further screened with the criterion of similarity to avoid unexpected data far away from the target region.Subsequently,a model correction mechanism is proposed to avoid model biases caused by the imbalance of the data in target region and other source regions Finally,the synthetic data are combined with the original data to establish fault detection model for the target region.(2)Aiming at the insufficiency of labeled data,a regressor-embedded semi-supervised variational autoencoder based data synthesis method is proposed.In this generative model,when the inputs are labeled data,the model is similar with the original variational autoencoder;when the inputs are unlabeled data,the quality variables are firstly predicted by the embedded regressor,then the unlabeled data and their corresponding quality variables are fed into the model.Because there are too few labeled data in the target region to fully train the generative model,the unlabeled data in the same region and the labeled data in other source regions are used to provide information lacking in the target region;that is,the former ones compensate for the deficiency of data characteristics in the target region,and the latter share some common information with the scarce labeled data.And then,the procedures of data synthesis,data selection and model correction are iteratively implemented to avoid model biases caused by the data imbalance between the target region and the source regions.Finally,the synthetic labeled data and the original scarce labeled data are used together to establish a regression model for the target region and the quality variables of the unlabeled data can be predicted on the basis of this model(3)Aiming at the data insufficiency in dynamic process,a corrected weighted recurrent variational autoencoder based data synthesis method is proposed.In this generative model,the encoder and the decoder are composed of long short time memory units,the dynamic characteristics of input data can be extracted effectively and the data with similar dynamic characteristics are generated.Considering that there are many parameters in the generative model while the original data in the target region are insufficient,data from other source regions are also added into the training set,effectively utilizing the similar autocorrelation and cross-correlation relationship between the target region and the source regions.Besides,a double safeguard mechanism is proposed.Firstly,weight coefficients are added to the loss function,ensuring the original data of the target region hold a greater proportion in training phase Secondly,the procedures of data synthesis,data selection and model correction are iteratively implemented to further reduce the model biases.Finally,the synthetic data are combined with the original data to establish dynamic model for monitoring the target region(4)In view of the severe insufficiency of process data,it is difficult to build a generative model for data generation,an image-based fault detection method is consequently proposed.A deep belief network based deep learning framework is designed to extract useful features from the image data.In this framework,each original image is segmented into small sub-images,which are subsequently inputted into corresponding sub-networks to extract local features.Afterwards,all the local features are fused by a global network to extract global features.This framework not only reduces the model complexity to a great extent,but also remarkably improve the training efficiency without deteriorating the fault detection accuracy.Meanwhile,a new statistic is specially developed for the proposed deep learning framework,making it possible to achieve feature extraction and fault detection in one composite model.