Classification of Technical Condition of Pumping Unit Using Intelligent Fault Identification

The purpose of this study is to analyze methods for intelligent fault identification when determining the performance of a centrifugal pump, taking into account unstable operating conditions. Using the Matlab Simulink software environment, modeling of an automated diagnostics and control system was carried out using the example of a 12NDs-Nm pumping unit. Based on retrospective vibration diagnostic data of 221 centrifugal pumps, forecasting models for the technological state of the system were created. Data processing and model training were carried out using the Python programming language and the Loginom analytical platform. It was determined that vibration diagnostic data alone is not enough to determine the cause of failure, since the accuracy of the models for multi-class prediction (5 causes of failure) was less than 60%. When solving the problem of binary classification (good/failure) of the system, the best accuracy was shown by an optimized model based on the “Random Forest” algorithm, using the principal component method and selection of hyperparameters. Its accuracy is 97.4%, precision is 99%, and recall is 98%. The accuracy of the multilayer perceptron model is 91%. Creating a comprehensive intelligent analysis system to improve the operational reliability of oil pumping station equipment is possible based on the use of machine learning algorithms and their combination, but requires taking into account both the operating conditions and key parameters of specific systems and units, as well as the development of a database of expert solutions for certain cases for further creating and maintaining an expert system.