Application of CNN Deep Learning to Well Pump Troubleshooting via Power Cards

Abstract

Recent years have seen extensive applications of deep learning, especially in identification and analysis of images, audios and texts, but incipient applications in petroleum industry. Shapes of loops in power cards of a pumping unit are valuable indicators for pump troubles. These troubles may cause engineering accidents, increase operation costs and reduce operation efficiency. This paper applies image recognition technique based on Convolution Neural Network (CNN) to well pump troubleshooting via power cards. Recent years have seen extensive applications of deep learning, especially in identification and analysis of images, audios and texts, but incipient applications in petroleum industry. Shapes of loops in power cards of a pumping unit are valuable indicators for pump troubles. These troubles may cause engineering accidents, increase operation costs and reduce operation efficiency. This paper applies image recognition technique based on Convolution Neural Network (CNN) to well pump troubleshooting via power cards. Firstly, we establish mathematical models both for displacements of the polished rod clamp of a pump and for loads of the polished rod during a reciprocating movement, and preset input parameters corresponding to pump trouble types and severity levels. Ideal benchmarking power cards as the media for pump troubleshooting are generated by simulating complete pumping processes via running the mathematical models with the preset pumping parameters. Secondly, we establish a power card classification model with the AlexNet method. Then we train it with the ideal benchmarking power cards to develop its function of pump troubleshooting and increase the classification accuracy. This model gains robustness and universality from manually presetting parameters for and full coverage of trouble types and severity levels. Thirdly, we train the classification model with real power cards and obtain the preliminary classification results. A further training makes it more practical and applicable to local operations of pump troubleshooting. In the further training, we localize the ideal benchmarking power cards via manual inspection and local expertiseby adjusting the preliminary classification results honoring field expertise. Finally, we randomly divide the localized benchmarking power cards into one training set and one testing set, and then train the classification model with the training set and then apply it to the testing set. The final classification results revealthe high accuracy and practicability of the classification model. It is recommended that GPU should be used for calculation with the classification model to satisfy clients' requirements for higher speeds and efficiency. It provides a feasible method to exploit the potential value of oilfield data assets. The work in this paper will function as a stepping stone in applying ideas, algorithms and models of artificial intelligence to more extensive and thorough aims.