Research On Fault Prediction Method Of Centrifugal Compressor Based On Deep Neural Network

Centrifugal compressor is a kind of rotating machinery,widely used in petrochemical industry,and it is an important equipment in industrial process.In the process of high-speed operation of the compressor,if the abnormality of a certain component evolves into a fault,the compressor will often stop and even damage,resulting in production stagnation and huge economic loss.Predicting key variables and identifying faults before compressor failures will be important to ensure equipment safety and reduce maintenance costs and losses.The purpose of fault prediction is to model historical data to obtain changes in key features before future failures occur.Long-term and short-term memory network(LSTM)has a good performance for long-term information memory of time series.Therefore,this paper uses the deep learning method LSTM to predict compressor failure.Firstly,according to the multiple measuring point data of the same variable in the process variable,the Kalman filter method is used to perform multi-sensor data fusion to obtain more accurate variable data.Secondly,the feature selection of process variables and vibration characteristics is carried out,and the highly linear correlation variables are eliminated by correlation analysis.By using the Relieff algorithm,the corresponding weights are calculated according to the correlation degree between different characteristics and faults of the compressor,the characteristics of low weight are eliminated,and the characteristics that contribute to the fault are retained for subsequent modeling.Finally,the selected variables are interpolated with abnormal outliers and standardized to eliminate the effects of different dimensions.On this basis,using the long short-term memory network(LSTM),the model of the normal training set is established to fit the normal state of the compressor,the vibration characteristics are predicted by the normal test set data,and the prediction accuracy of the model is evaluated by RMSE.By comparing different time steps,the prediction accuracy of the corresponding network is evaluated,and the final network time step is established.By using multivariate Gaussian distribution to model the prediction error of normal training set,a comparison between threshold value and pre-fault data error is set to identify the fault of the system.The results show that this method can effectively detect the abnormal change of the key characteristics of the compressor before the fault occurs,and realize the fault prediction of the equipment..