Fault Classification With Weakly Supervised Learning

Process monitoring is an important technical means to ensure the reliability and stability of the production state of modern industry.It can reduce the cost of industrial production,improve the quality,ensure the continuous,stable and sustainable operation of industrial production process,and reduce the occurrence of faults and industrial accidents.It has become an indispensable part of modem industrial process.With the development of industry,the collection,storage and transmission of industrial process information become more and more popular.A large number of industrial process data can depict the scene of industrial generation process,and the data-driven process monitoring methods become more feasible and effective,so they have become a research hotspot in the field of process monitoring.As a popular branch of machine learning,deep learning has been widely used in computer vision,natural language processing,speech recognition and other fields.Therefore,supervised models such as deep neural network are widely used in industrial process fault classification with superior performance.But in the actual industry,it is high cost to obtain the strong supervision information such as accurate label in industrial process monitoring.Some weak supervision information such as unlabeled sample information,inaccurate label sample information is low cost and easier to obtain.This paper proposes fault classification methods based on weakly supervised learning to solve the problem of lack of considering the modeling of weak supervision information in industrial process.It focuses on solving the data utilization problem of two kinds of weak supervision information:inaccurate supervised information and incomplete supervised information.The main research work includes:(1)Aiming at the problem of inaccurate label and no label in the sample dataset obtained in the industrial process,on the premise that the label noise is only related to the class of the sample,this paper proposes that the cross entropy loss function of the traditional neural network model can be modified by introducing a label probability transition matrix to describe the relationship between the inaccurate labels and unknown true labels of process data.In order to get label probability transition matrix,this paper proposes a method to estimate the label probability transition matrix by using the two-components Gaussian mixture model and the maximum posteriori value of each class of samples.In addition,the proposed weakly supervised multi-layer perceptron model can be combined with the pseudo-label method to utilize the unlabeled samples,which improves the classification performance of the model.(2)The proposed weakly supervised multi-layer perceptron model can only be applied to the specific situation of label noise related to the label class,and the final classification effect of the weakly supervised multi-layer perceptron is limited to the pre-estimated label probability transition matrix.A probability graph is introduced to describe the label probability transition matrix,and the probability graph is combined with the neural network,that is,probability graph embedded in neural network.The neural network model uses the label probability transition matrix and other parameters learned from the probability graph model to modify the loss function of the neural network model.The learning process of the probability graph parameters and the training process of the neural network model are iterated,and finally the classifier that can be robust to the inaccurate label sample data set is obtained.In addition,for the problem of using unlabeled samples,by replacing the fully connected neural network in the model with Stacked Autoencoder,the model parameters of Stacked Autoencoder are initialized by unlabeled samples,which further improves the performance of the model.(3)In order to solve the problem that the label probability transistion matrix can only describe the label noise that is only related to the true label of the sample,and can not be used in the case that the label noise is also related to the sample itself.A weakly supervised learning fault classification method based on class prototype learning is proposed,and in each class,the feature representation of fault samples with accurate labels is selected as class prototype.By comparing the feature vectors of each fault sample with each class of class prototypes,the true label of the sample is deduced.Then,the loss calculated by inferred label and the loss calculated by original label are weighted as the loss function of the classification model.This loss adjustment method reduces the overfitting degree of the model to the inaccurate label samples.In addition,combined with pseudo-label method and class prototype learning model,the proposed model can further learn unlabeled sample information,which also improves the performance of the model.