Development of a CNN-based fault detection system for a real water injection centrifugal pump

Large-sized centrifugal pumps play a major role in produced water injection systems in oil and gas production. Monitoring this equipment operation is vital to guarantee its efficiency and to reduce the occurrence of unplanned downtimes. The main goal of this work was to develop a fault detection system based on artificial intelligence (AI) algorithms for a water injection centrifugal pump located at an oil and gas company's offshore platform. The convolutional neural network (CNN) was the main algorithm investigated in this work. However, other machine learning techniques (i.e., support vector machines, random forest, and multilayer perceptrons) were used to compare against the results achieved by the CNN. A data-driven methodology was proposed for the stages of exploratory data analysis (EDA), data labeling, automatic hyperparameter optimization, training and testing of the chosen models. The results showed that the proposed methodology was effective and applicable. The CNN and the support vector machine (SVM) models presented interesting performances. The CNN model returned 93.7% precision and 59.8% recall, whereas the SVM model returned 84.6% precision and 66.7% recall. The challenges related to the use of a real dataset were also discussed, emphasizing the data labeling step. Applying the k-means clustering technique proved to be useful for labeling the data instances.