An evolutional deep learning method based on multi-feature fusion for fault diagnosis in sucker rod pumping system

Abstract

As the smart oilfield has grown, various deep learning technologies are being utilized to recognize the graphic feature of the indicator diagram in order to detect the fault type of rod pump system, so as to maintain the oilfield's regular production. However, the original data for the indicator diagrams from various oil fields are influenced by different geographic conditions, sensor equipment, acquisition software, etc. and exhibit specific environmental characteristics. This poses difficulties for indicator diagram-based fault diagnosis methodologies and necessitates the use of a more generalizable diagnosis model. To address the issue, a multi-feature fusion fault diagnostic model is proposed. The model fuses the Fourier descriptor of the indicator diagram as a feature with the graphic feature to enhance the robustness of the feature. Firstly, the two backbone networks perform feature extraction on the single-modal input data of their own networks. Secondly, the information from the indicator diagram and the Fourier descriptor are learned together as features through the interactive fusion module. And finally the integrated features are used for feature classification to obtain the output of the network. The accuracy of the diagnostic model when using only a single feature is respectively: 0.8233(the graphic feature), 0.9422(the Fourier descriptor feature) according to the findings of the validation experiment, and when using the fusion of the two features is 0.9724. The results demonstrate that the suggested multi-input feature fusion model performs better than the single-input model. The technique makes use of the correlation between the features to realize their complementary benefits and enhance the effectiveness of the diagnostic model.