Research And Application Of Empirical Mode Decomposition For Fault Diagnosis Of Petrochemical Units

Because of the variability of the operating environment of rotating machinery,the complexity of vibration,as well as the hybridisation of noise and original signals,the vibration signals are nonlinear and non-stationary.The above problems bring challenging tasks to fault signal analysis,so it is of great significance to study bearing fault prediction and diagnosis.Many bearing fault diagnosis methods based on time-frequency have been proposed.However,the insensitivity of features and the need for a large amount of prior knowledge are still major disadvantages,resulting in the methods are not suitable for big data.With the rapid development of artificial intelligence,deep learning technology has become a bridge connecting mechanical big data.Compared with traditional machine learning,deep learning has better nonlinear recognition performance and better effect in fault diagnosis decision-making.Besides,transfer learning,as a new deep learning method,uses existing knowledge to solve different but related domain problems and reduces the requirement on data characteristics.In the traditional time domain analysis method,there are many limitations such as dependence on expert knowledge and large amount of feature calculation.In order to solve the above problems,this paper utilizes the advantages of deep learning in dealing with two-dimensional sparse problems and the advantages of transfer learning in reducing the complexity of deep learning.Application of CEEMD and data-driven in intelligent fault diagnosis,and a bearing compound fault diagnosis of F-CEEMDAN and transfer learning are proposed.The main research contents are as follows:(1)Application of CEEMD and data-driven in intelligent fault diagnosis is proposed,which preprocesses fault vibration signals by signal conversion method and combines the advantages of deep learning in image processing.Firstly,the fault diagnosis signals are decomposed by empirical mode,and several inherent mode functions are obtained.Secondly,one-dimensional vibration data is converted into two-dimensional images by signal conversion,and different deep learning models are used to identify and classify fault images.The proposed method studies the combination of CEEMD and signal conversion,which overcomes the problem of noise and modal confusion caused by the traditional method.The proposed method further improves the feature integrity and makes full use of the advantages of deep learning in machine vision recognition.(2)A bearing compound fault diagnosis of F-CEEMDAN and transfer learning are proposed.Firstly,on the basis of content(1),this section improves the method of CEEMD.The intrinsic modal functions decomposed by complete ensemble empirical mode decomposition with an adaptive noise were filtered by correlation,and the next step is to reconstruct intrinsic modal functions.Secondly,the reconstructed vibration data is converted into two-dimensional representation by signal conversion method.Transfer learning and deep learning model are used to identify the health status of compound faults.Compared with the traditional empirical mode decomposition fault diagnosis method,the superiority of the proposed algorithm is verified.The method overcomes defects such as insufficient use of data and the large time complexity.(3)The proposed method is tested on the data sets collected by the multi-stage centrifugal fans of Guangdong Key Laboratory of Petrochemical Equipment Fault Diagnosis and the Laboratory of Case Western Reserve University.In the content of(1),the average prediction accuracies of the proposed method are 98.75% and 99.74%,respectively,showing remarkable performance.The content(2)is to identify the health condition in the compound fault,the average accuracies of the experiment are 99.71% and 98.75%,the complexity of training times are greatly decreased and the stability is remarkable.