Research On Fault Diagnosis Of Pumping Unit Based On KPNMF-FNN

The oilfield production efficiency is closely linked to the fault diagnosis technology of pumping units.Most pumping units are located in remote and undeveloped areas,and the work strength is intensive with high temperature,the downhole production conditions are very complex,and parameter coupling problem is serious.When a fault occurs,if fault diagnosis can not be carried out timely and accurately,it may cause property loss or even serious personal safety accident.Therefore,establishing an efficient and accurate fault diagnosis model for pumping units has extremely important application value and development prospects.Fault diagnosis algorithm based on Artificial Intelligence(AI)has become a research focus at present.Nonnegative Matrix Factorization(NMF)is a new method of artificial intelligence.Its basic idea is reductionist though which means that the whole is constituted by parts NMF restore the original data matrix into the product of base matrix and coefficient matrix.It is noted that the matrices here are nonnegative,satisfy pure additivity,and explanability increase,have clear physical meaning.The algorithm is high efficiency and low consumption.It has become a powerful tool for processing non-negative data.Aiming at the problem of fault diagnosis of pumping units,a fault diagnosis model of pumping units based on kernel projection non-negative matrix factorization-false neighborhood algorithm is proposed.The main contents are as follows:Firstly,the development of non-negative matrix decomposition is briefly introduced.The non-negative matrix decomposition is deduced rigorously and meticulously from the angle of blind signal separation under Gauss distribution and Poisson distribution.This corresponds to the two objective functions of Euclidean distance and KL divergence,respectively.The iterative formula of alternating one-step gradient descent is given,and the calculation ideas of alternating least squares method and projection gradient method are briefly introduced.The existing non-negative matrix factorization algorithms are briefly summarized.Secondly,for the reason that the classical non-negative matrix decomposition is not very effective in dealing with linear non-separable data,the kernel method is introduced,and the projective non-negative matrix decomposition(PNMF)has the advantages of less computation and more stable algorithm than classical non-negative matrix decomposition,moreover,the sparsity of decomposition results can often be deduced from orthogonality,the orthogonal constrained kernel projective non-negative matrix decomposition(KPNMF)is proposed.Provide the iteration formula,and proof its convergence strictly.A feature extraction algorithm based on KPNMF is proposed.Finally,in order to reduce the computational burden of kernel projection non-negative matrix decomposition,the idea of False Nearest Neighbours(FNN)come from phase space reconstruction is combined,and a KPNMF-FNN algorithm is proposed.Based on this algorithm,dimension reduction and feature extraction methods are established.Finally,a fault diagnosis model based on this algorithm is proposed,and fault traceability and location by contribution graph method.Experimental on two non-linear complex processes of Tennessee-Eastman simulation platform and pumping unit,the experimental results show that the algorithm can quickly and accurately diagnose the fault status of the two systems,and the expected effect can be achieved.