| Submersible plunger pump lifting system is a new oil recovery equipment used in the oil field. It utilizes a linear motor as the power system, which avoids the energy loss caused by the energy exchange device in the traditional rotating motor and improves the oil pumping efficiency. However, submersible plunger pump is installed in a deep well, which makes it difficult to monitor its working condition using existing fault diagnosis methods and increases the difficulty on equipment repairment and maintenance and subsequently the business operating costs. Hence, it is important to develop an applicable fault diagnosis method for the submersible plunger pump to monitor its running state and give useful information when the equipment fails, upon which the equipment can be timely repaired to extend its life.Fault diagnosis belongs to the category of pattern recognition, which mainly deals with signal preprocessing, feature extraction and fault recognition, while feature extraction is the core step of fault diagnosis technology, which has a direct impact on the complex-ity and accuracy of the subsequent fault recognition algorithm. How to dig out the low dimensional features hidden in the high-dimensional data space to reduce the difficulty of subsequent fault diagnosis is still a hot topic in the field of fault diagnosis. In this paper, based on the locally linear embedding algorithm (LLE), a thorough study on the problem-s of the LLE algorithm in dimensionality reduction was conducted, and the appropriate solutions were put forward. The main work can be summarized as follows.The signal with non-stationary and low signal-to-noise ratio may reduce the dimen-sionality reduction effect of the LLE algorithm, which will affect the final recognition accuracy. A fault diagnosis method of multi algorithm fusion based on wavelet transfor-m, SVD decomposition and LLE (WSLLE) was proposed. In this method, the wavelet transform was used to obtain the frequency domain space of the signal, in which the sig-nificant features can be easily extracted. SVD is used to analyze the characteristics of the signal in different frequency bands, by which a new feature space was established. Then LLE was used to carry out the dimension reduction in the new feature space, and SVM was used for the fault recognition based on the results of the dimension reduction. At last the feasibility and accuracy of the proposed fault diagnosis method were verified by using the data set of the submersible pump and bearing fault.In order to resolve the high complexity of WSVLLE, the effect of the local structure on the embedded results in LLE was analyzed in depth. A new reconstruction objective function was built in high dimensional space based on the full consideration of the charac-teristics of the reconstruction objective function and combined with sparse regularization theory, and a new method for the fast solution of the objective function was explored. Based on these, the iterative shrink LLE (ISLLE) and the corresponding fault diagnosis method were proposed. During the mining process of the high dimensional local structure of LLE, the sparse regularization method was incorporated into the algorithm to improve the robustness of the algorithm. By introducing the surrogate function, the solution of the reconstruction objective function was converted into solving the series of linear equation, which can reduced the computational complexity of the algorithm. The adaptability and robustness of the ISLLE algorithm were verified by the artificial data set and the real data set.ISLEE belongs to an unsupervised learning algorithm, so the category information cannot be taken full advantage to guide the dimensionality reduction in ISLLE. Com-bined with the ISLLE algorithm in the third chapter and the semi-supervised learning techniques, the semi-supervised ISLLE (SS-ISLLE) algorithm was proposed, which can make full use of the priori knowledge of the data, and the separability of the embedding results was improved. And an fault diagnosis method was proposed based on SS-ISLLE and SVM. SS-ISLLE was applied to the original high dimensional data space for the low dimensional embedding coordinate of the data set, and the SVM recognizer was used in the low dimensional space to complete the identification of fault types. The practicability and generality of the proposed method were verified by using the Iris data set, bearing data set and the data set of the submersible plunger pump.SS-ISLLE operates in a batch mode, so the out-of-sample data cannot be efficiently processed. For resolving this problem, each step of SS-ISLLE was deeply researched to get the nature of the high complexity of processing the out-of-sample. And the incremental SS-ISLLE (ISS-ISLLE) was proposed to process the out-of-sample quickly and accurately. Because the local structure of dataset is linear, the embedding results of the out-of-sample data and the old samples whose structure were changed can be obtained by their neighbors, upon which gradient iterative algorithm is used to improve the accuracy of the whole embedding results. The feasibility, accuracy and computational speed of the algorithm were verified by the standard manual data set.Because the recognition accuracy of the proposed fault diagnosis methods are sen-sitive to the embedding dimension. An weighted correlation dimensionality (WCD) was proposed on the basis of in-depth study of fractal dimension estimation method, especially the correlation dimension estimation method, and the influence of the probability distri-bution of the data on the dimension estimation. In WCD, the length of the traversal of the vertex in the graph to its local neighborhood was calculated to measure the local proba-bility distribution. The weight of local "volume" of each sample to total "volume" was calculated by the Gauss radial basis function to realize the globalization of local "volume" The estimation accuracy of the WCD method was verified by the artificial data set and the real data set, and the complexity of WCD was also analysed. |
PDF Full Text Request