The capital-intensive oil & gas industry invests billions of dollars in equipment annually and it is important to keep the equipment in top operating condition to help maintain efficient process operations and improve the rate of return by predicting failures before incidents. Digitalization has taken over the world with advances in sensor technology, wireless communication and computational capabilities, however oil & gas industry has not taken full advantage of this despite being technology centric.
Dynamic seals are a vital part of reciprocating and rotary equipment such as compressor, pumps, engines, etc. and are considered most frequently failing component. Polymeric seals are increasingly complex and non-linear in behavior and have been the research of interest since 1950s. Most of the prognostic studies on seals are physics-based and requires direct estimation of different physical parameters to assess the degradation of seals, which are often difficult to obtain during operation. Another feasible approach to predict the failure is from performance related sensor data and is termed as data-driven prognostics. The offline phase of this approach is where the performance related data from the component of interest are acquired, pre-processed and artificial intelligence tools or statistical methods are used to model the degradation of a system. The developed models are then deployed online for a real-time condition monitoring. There is a lack of research on the data-driven based tools and methods for dynamic seal prognosis. The primary goal in this dissertation is to develop offline data-driven intelligent condition monitoring and prognostic methods for two types of dynamic seals used in the oil & gas industry, to avoid fatal breakdown of rotary and reciprocating equipment. Accordingly, the interest in this dissertation lies in developing models to effectively evaluate and classify the running condition of rotary seals; assess the progression of degradation from its incipient to failure and to estimate the remaining useful life (RUL) of reciprocating seals. First, a data-driven prognostic framework is developed to classify the running condition of rotary seals. An accelerated aging and testing procedure simulating rotary seal operation in oil field is developed to capture the behavior of seals through their cycle of operation until failure. The diagnostic capability of torque, leakage and vibration signal in differentiating the health states of rotary seals using experiments are compared. Since the key features that differentiate the health condition of rotary seals are unknown, an extensive feature extraction in time and frequency domain is carried out and a wrapper-based feature selection approach is used to select relevant features, with Multilayer Perceptron neural network utilized as classification technique. The proposed approach has shown that features extracted from torque and leakage lack a better discriminating power on its own, in classifying the running condition of seals throughout its service life. The classifier built using optimal set of features from torque and leakage collectively has resulted in a high classification accuracy when compared to random forest and logistic regression, even for the data collected at a different operating condition. Second, a data-driven approach to predict the degradation process of reciprocating seals based on friction force signal using a hybrid Particle Swarm Optimization - Support Vector Machine is presented. There is little to no knowledge on the feature that reflects the degradation of reciprocating seals and on the application of SVM in predicting the future running condition of polymeric components such as seals. Controlled run-to-failure experiments are designed and performed, and data collected from a dedicated experimental set-up is used to develop the proposed approach. A degradation feature with high monotonicity is used as an indicator of seal degradation. The pseudo nearest neighbor is used to determine the essential number of inputs for forecasting the future trend. The most challenging aspect of tuning parameters in SVM is framed in terms of an optimization problem aimed at minimizing the prediction error. The results indicate the effectiveness and better accuracy of the proposed approach when compared to GA-SVM and XGBoost. Finally, a deep neural network-based approach for estimating remaining useful life of reciprocating seals, using force and leakage signals is presented. Time domain and frequency domain statistical features are extracted from the measurements. An ideal prognostic feature should be well correlated with degradation time, monotonically increasing or decreasing and robust to outliers. The identified metrics namely: monotonicity, correlation and robustness are used to evaluate the goodness of extracted features. Each of the three metric carries a relative importance in the RUL estimation and a weighted linear combination of the metrics are used to rank and select the best set of prognostic features. The redundancy in the selected features is eliminated using Kelley-Gardner-Sutcliffe penalty function-based correlation-clustering algorithm to select a representative feature from each of the clusters. Finally, RUL estimation is modeled using a deep neural network model. Run-to-failure data collected from a reciprocating set-up was used to validate this approach and the findings show that the proposed approach can improve the accuracy of RUL prediction when compared to PSO-SVM and XGBoost regression. This research has important contribution and implications to rotary and reciprocating seal domain in utilizing sensors along with machine learning algorithms in assessing the health state and prognosis of seals without any direct measurements. This research has paved the way to move from a traditional fail-and-fix to predict-and-prevent approach in maintenance of seals. The findings of this research are foundational for developing an online degradation assessment platform which can remotely monitor the performance degradation of seals and provide action recommendations on maintenance decisions. This would be of great interest to customers and oil field operators to improve equipment utilization, control maintenance cost by enabling just-in-time maintenance and increase rate of return on equipment by predicting failures before incidents.