Industrial Process Fault Classification Based On Semi-Supervised Deep Learning

Process monitoring monitors the production process,detects abnormal process faults and restores it in time,which is of great significance to the industrial process.Fault classification is part of the process monitoring method that classifies faults by establishing a model between variables and fault categories.With the large amount of industrial process data collected by the DCS system,data-driven fault classification methods have been widely used.However,in the actual industrial production,it is time-consuming and expensive to label the category of the fault data,and requires expertise in the process domain.Therefore,there are often only a small number of labeled samples,and most of the data are unlabeled,which makes it necessary to establish a semi-supervised learning method.Considering the powerful feature extraction ability of deep learning,especially the learning ability on unlabeled data,this research studies the fault classification method based on semi-supervised deep learning.The main contents are as follows(1)The traditional autoencoder model have the problems of lack of generalization ability and easy overfitting caused by only using labeled data when fine-tuning.To solve that,a semi-supervised fault classification model based on feature-aligned autoencoder is proposed.The feature alignment autoencoder aligns the features extracted by the autoencoder on the labeled and unlabeled data during fine-tuning.Experiments show that the proposed model has better generalization ability than the traditional autoencoder method,and effectively reduces the over-fitting phenomenon caused by the lack of labeled samples(2)Many semi-supervised learning algorithms assume that labeled samples and unlabeled samples belong to the same distribution.However,unlabeled fault samples has not been filtered and labeled by experts,and industrial processes sometimes have problems such as drift and noise.The above assumption is difficult to guarantee,and this will lead to a decrease in the performance of the semi-supervised algorithm.This paper proposes a fault classification method based on robust feature-aligned autoencoder.The unlabeled samples that are inconsistent with the labeled sample distribution are filtered or given lower weights,and then the subsequent modeling is performed.This method effectively improves the robustness of the model and reduces the problem of model performance degradation caused by inconsistent labeled and unlabeled sample distribution(3)Pre-training can make fully use of unlabeled samples and help semi-supervised learning.However,there are few existing pre-training methods for industrial processes,and a single pre-training method has its limitations.This paper proposes a semi-supervised fault classification method based on an ensemble pre-training model,which contains a variety of different dynamic pre-training methods,and then uses ensemble learning to combine them as base learners.Pre-training can not only improve the performance of the base learner,but also make sure that there are enough differences between base learners to use the advantages of ensemble learning.The proposed method make base learners use each other as teachers to learn from unlabeled samples,and the rationality of this strategy is proved...