Machine learning based sucker rod pump fault diagnosis using motor power curve

Relevance. The complexity of monitoring and diagnosing the condition of underground structural elements of sucker rod pumping units and large economic losses when operating this equipment with defects not identified in a timely manner. Aim. Development of methods for detecting faults in a sucker rod pump that do not require the involvement of highly qualified personnel for diagnosis, using information that is easily available on the surface. Methods. Machine learning methods (Decision tree method, K-nearest neighbors method, Support vector machine, Naive Bayes classifier) using motor power curves. Results and conclusions. The paper demonstrates the possibility of detecting faults in a sucker rod pump based on machine learning methods. The study was carried out on the basis of a developed simulation model of a sucker rod pump, used to reproduce motor power curves, taking into account the impact of the features of various equipment operation scenarios. Being the fundamental energy source for the oil production, motor power is directly related to the real-time operating condition of the oil well, and the motor power curve is a reliable source with the ability to increase the efficiency of sucker rod pump diagnostic. To train the machine learning classifiers and evaluate their performance accuracy, a number of characteristics were used, obtained from motor power curves for six different pump operating states. Namely, operating coefficients were calculated, representing the ratio of the power integral at each of the four stages of the installation operating cycle to the power integral for the entire cycle. The results show that the considered approach allows for high accuracy in diagnosing the operating conditions of a sucker rod pump. The classifier based on the decision tree method showed the highest efficiency among the four studied classifiers in identifying all six types of faults (95.8%), and the support vector machine method showed as well very high efficiency (90.3%).