Research On Fault Early Warning And Diagnosis Of Chemical Equipment Based On Deep Learning

With the rapid development of modern chemical industry,the operation principle and failure mechanism of the equipment are more and more complex,which brings about a significant increase in production energy efficiency,but also makes it more difficult to accurately predict the failure.And the equipment failure in the chemical industry is often very dangerous,even causing a large number of casualties.Therefore,how to early-warning the abnormal operation of the equipment in time and accurately diagnose the fault is an urgent problem in the field of chemical fault diagnosis.In recent years,with the development of high-precision sensors,computer technology and artificial intelligence,intelligent fault early warning and diagnosis method based on data-driven has gradually become a research hotspot.As an important branch of deep learning,because of its excellent feature extraction ability and advantages in processing high-dimensional,non-linear equipment operation data,it has made great achievements in the field of chemical fault early warning and diagnosis,providing a new paradigm.Based on deep learning,this paper analyzes and studies three kinds of data imperfections in the early warning and diagnosis of chemical equipment,which are the lack of fault samples,the imbalance of the number of fault samples and the change of working conditions.The specific research contents are as follows:(1)In order to solve the problem of mechanical equipment failure samples missing,an early warning method without failure samples is proposed based on the feature self-learning ability of denoising Auto-encoder.Firstly,the noise reduction self encoder is used to extract the abstract feature expression of the normal data of the equipment in the way of unsupervised learning,calculate the feature benchmark and set the alarm threshold;in the monitoring and early warning stage,calculate the distance between the abstract feature and the benchmark extracted by the model from the data to be tested,compare with the alarm threshold,find out the weak signs in the early stage of equipment failure in time,and achieve a large-scale improvement Pre-warning.The application of unsupervised deep learning model in mechanical equipment fault early warning solves the dependence of traditional machine learning methods on fault samples in model training,and avoids the complex process of manually selecting feature parameters.In the bearing fault experiment and the reciprocating compressor practical engineering,the proposed method shows a good fault early warning ability,and has a certain recognition ability for the weak symptoms in the early stage of fault occurrence,which greatly advances the alarm time point,and has a high application value.(2)In view of the situation that unmarked samples account for the majority of mechanical equipment data,a deep neural network model combined with active learning is proposed.Through the targeted selection of samples and manual marking in the active learning stage,the marked training sample set is expanded step by step,so as to improve the diagnosis performance as much as possible under the premise of using as few samples as possible.In the experiment of bearing fault diagnosis in Case Western Reserve University,the data are divided by imitating the fact that most of the fault samples are unmarked in practical engineering application.Firstly,a large number of unmarked sample sets are used to pre-training the stack denoising Auto-encoder,and the deep neural network is initialized.Secondly,in the active learning stage,the unmarked samples are selected according to certain standards in each iteration Manual marking is carried out to participate in the subsequent supervised training fine-tuning,and the iterative process of "sample screening manual marking supervised fine-tuning" is repeated until the model converges.The experimental results show that this method can improve the sample utilization rate,and improve the performance of the model far more than the traditional sample screening method when using the same number of samples,and more importantly,it can effectively inhibit the occurrence of false-positive results,which is of great significance for the fault diagnosis of mechanical equipment.(3)In view of the problem that the distribution of the training data and the data to be measured is inconsistent,which leads to the decrease of the diagnostic accuracy under the condition of the mechanical equipment under the off design condition,a off design diagnostic model based on the domain confrontation network is proposed.The model includes feature extractor based on convolution neural network,fault classifier based on full connection neural network and domain discriminator,and makes full use of the idea of generative adversary network to make the feature extractor and the domain classifier used to distinguish working conditions to conduct adversary training,so that the features learned by the model have domain invariance and become shared features under different working conditions,At the same time,under the constraint of fault classifier,the features learned by the model can maintain enough fault sensitivity.The experimental results show that the proposed diagnosis method can effectively overcome the influence of different working conditions on the diagnosis performance and ensure sufficient accuracy.