Research On Fault Prediction Method Of Centrifugal Compressor Based On Data Driven

With the development of industrial technology,industrial equipment is becoming larger and more complex.At the same time,ensuring the safe and stable operation of the equipment has received more and more attention.Centrifugal compressors are indispensable key equipments in the fields of metallurgy,chemical engineering,power generation,and refrigeration.If they fail,they will inevitably lead to the stagnation of the entire production system.If the failure is serious,it may even lead to a major accident,which may cause a lot of economic losses and may also cause safety problems.If some symptoms of the failure can be discovered in advance,measures can be taken in advance to reduce the loss and ensure production safety.Therefore,it is necessary to study the method of failure prediction.In the context of the era of big data,a large amount of data can be generated for analysis in the industrial field which provides the conditions for data-driven methods such as data mining and machine learning.In the field of artificial intelligence,especially in recent years,deep learning algorithms have achieved remarkable results in the fields of image processing,machine vision,speech recognition,and natural language processing.However,there are few research results in the field of industrial fault diagnosis and prediction.In this thesis,it is an attempt to study industrial failure prediction methods using deep learning algorithm long short term memory recurrent neural network(LSTM-RNN),and achieved good results.Details are as follows:(1)In order to improve the training efficiency and generalization ability of neural network,a feature selection method based on cluster analysis and information gain is proposed for multidimensional time series.In this method,each variable of the candidate sequences subset of different time step is divided into categories by a clustering algorithm,and then each variable of each subset is evaluated by information gain,thereby determining the number of time steps and the variable of the multi-dimensional time series.(2)A long-term memory recurrent neural network was constructed and trained as a time-series prediction model that is built using the normal operating data to predict the future vibration signal of the compressor.(3)Using the idea of anomaly detection,the prediction errors of the time series were collected and modeled with a multivariate Gaussian distribution.The logarithm of the probability density function is used as the evaluation value,on which a threshold is set to evaluate the status of the device.The experimental results show that the proposed method in this thesis can find the signs of device in a state of degradation before the failure occurs,and achieve the purpose of fault prediction.