Research On Early Warning Model For Correlative Fault Of Centrifugal Compressor Based On DNN-HMM

As a fundamental power equipment in the gas transmission pipeline,the compressor unit is also a weak link vulnerable to faults in the long distance transmissio n system.Due to the interaction among subsystems,general methods of mechanical fault fail to satisfy the actual demand on the compressor unit,without taking other devices into account and regarding the compressor unit as a system.Through recognizing and early warning of possible faults in the compressor unit beforehand,fault disposal costs could be significantly lowered while the impact could also be reduced.Therefore,the improvement of unit safety and reliability will bring considerable economic and social benefits.In this paper,four themes concerning the compressor unit are highlighted,namely the correlation mechanism of different subsystems,relevance of operating parameters,clustering of state parameters,and the security early warning model.1.Process composition and operating characteristics of the compressor unit are systematically analyzed for researches on correlation mechanisms of gas state in the compressor unit combined with application of BWRS equation and N-S equation.The correlations of rotate speed,outlet pressure and outlet temperature are studied under different inlet flow rate.Three aspects are expounded,including the dry gas seal system and the compressor body,the lubricating system and the compressor body,as well as gas hydrate influencing laws at the inlet pipeline of the compressor.2.DNN-HMM early warning algorithm of the compressor unit is proposed based on the deep neural networks and hidden Markov model.The algorithm can be classified into three stages,namely preprocessing,training and testing stages.Deep neural networks are stacked by multi-layers of restricted Boltzmann machines,and output the characteristic value with the input of unit operation states,which are inputs of the hidden Markov model.Finally,the hidden state of compressor can be obtain by HMM.3.Erdem correlation coefficients are adopted and neighbor extremum method is proposed for the calculation of the maximum correlation coefficient,which is verified to be effective by synthetic control chart database.Based on empirical mode decomposition method,the compressor unit operation signals are denoised,and then operating parameters are conducted with correlation analysis,which provides the analysis conditions for subsequent clustering analysis works.4.In accordance with the clustering method based on density peaks,operating parameters are conducted with clustering analysis.Then the information entropy increment is introduced and the characteristics of the clustering method based on density peaks are analyzed,which comes to the conclusion that the density peak information entropy increment should be the selection basis of clustering center numbers.K-means clustering algorithms are compared to demonstrate the rationality of the cluster.Field parameters are clustered according to the method,which brings a set of operating state parameters with relative independence.5.The field data is applied for testing the model based on DNN-HMM.The result shows that this model can effectively solve the early warning problem of correlative faults among subsystems of the compressor unit.Compared with SVM algorithm,it is proved that the proposed DNN-HMM model behaves better in terms of accuracy and false alarm rate.