Research On Deep Learning Based Fault Diagnosis Algorithms

The data-driven intelligent fault detection and diagnosis(FDD)system has become an important means to ensure the smooth operation of the chemical process,but the mainstream fault detection and diagnosis is limited by the shallow structure,often difficult to achieve satisfactory accuracy.In recent years,deep learning method has shown better performance than traditional methods in many fields.Therefore,the introduction of deep learning method into the field of fault detection and diagnosis is also one of the current research hotspots.However,the chemical process data has the characteristics of limited data scale,dependence among samples,complex nonlinear and coupling relationship among variables,so it is impossible to simply transfer the deep learning method to the field of chemical fault detection and diagnosis.Therefore,the work of this thesis is just based on this background.The research of this thesis is supported by the Natural Science Foundation of Zhejiang Province.The main research work and achievements are as follows(1)In the actual industrial system,the complex chemical environment often leads to the inability to obtain enough fault detection data,which limits the application and development of deep learning method in fault detection.Based on the above problems,this thesis proposes a sample space reconstruction strategy.The strategy is based on random sampling to construct the same or different sample pairs to expand the data scale.At the same time,it transforms the complex classification problem into the similarity comparison problem between the samples,reduces the complexity of the task and reduces the demand of the model for the amount of data.Based on the reconstruction strategy,Multi-scale Siamese Convolutional Neural Networks(Multi-scale Siamese CNN)is proposed by improving the structure of the Siamese CNN.In this algorithm,the samples are randomly selected according to the order of the normal sample pairs and the sample pairs with a fault sample pair.The convolution neural network is used for multi-scale feature extraction to make the similarity between the model learning samples and classify them.In the test stage,the test samples are compared with many groups of samples without fault,so as to achieve the effect of accurate detection.Finally,the algorithm proposed in this thesis is applied to TE process fault detection for performance verification,and compared with the conventional data-driven method.The comparison results verify the superiority of the algorithm proposed in this thesis.(2)In view of the problem that the current mainstream fault diagnosis model is difficult to effectively deal with chemical process data with strong nonlinear relationship,this thesis introduces three attention mechanisms to represent the importance of features and the relationship between features under the current input,and proposes a multi attention deep neural network(MA-DNN)algorithm.The algorithm not only uses deep neural network to mine deep chemical process information,but also uses three attention mechanisms to better express features and avoid the impact of insensitive features,so as to make full use of the information between features and enhance the fault diagnosis ability of the network.Finally,the performance of the model is also verified on the TEP data set.Experimental results show that the algorithm can effectively improve the effect of chemical process fault diagnosis(3)Aiming at the problem that the mainstream fault diagnosis model can not effectively utilize the dynamic information of chemical process data,this thesis proposes a convolutional neural networks based on self-attention(SA-CNN).In order to make use of the dynamic information of chemical process data,the algorithm constructs a self-attention mechanism structure based on the time sequence information.By using the moving window method to segment the data and introducing the local time sequence information,the self-attention mechanism can effectively represent the dynamic information of chemical process data using two-dimensional matrix.Finally,the convolution neural network is used to further extract the features and enhance the fault diagnosis performance of the model.By comparing with other mainstream fault diagnosis models with dynamic information in TEP data set,the superiority of SA-CNN algorithm is fully demonstrated.