Electro-submersible pumps (ESPs) are now widely used to produce non-linear flow wells with high production rates and high water content to increase production. Shaft breakage is a common problem in the oil and gas industry, which can lead to production interruptions and significant economic losses. The objective of this paper is to evaluate principal component analysis (PCA) as an unsupervised machine learning technique for detecting the causes of ESP shaft breakage. The method has been successfully applied to detect ESP shaft breakage in real time in Penglai block of Bohai Oilfield, China. By plotting a two-dimensional map of the first and second principal components, different clusters in the stable, unstable and faulty regions can be identified. In this way, potential ESP shaft breaks can be detected when the clustering starts to deviate from the stable region. In addition, a PCA diagnostic model was developed to predict when an ESP shaft fracture occurs and to identify the most significant decision variables in relation to the event. This paper demonstrates that the PCA method performs well in monitoring ESP systems and accurately predicting impending ESP shaft fracture.
brief
Electrosubmersible pumps (ESPs) are now widely used in nonlinear flow wells, high production, high water content, and offshore wells, where they are simple and efficient (Ratcliff, Gomez, Cetkovic, and Madogwe, 2013). Of all the oilfield artificial lift systems, ESP is considered to be the preferred choice because of its ability to produce greater volumes at higher temperatures and deeper well depths. However, ESP systems are usually observed to be interrupted by failures. A broken pump shaft can be a serious problem for operating companies as it can result in the loss of hundreds of millions of barrels of oil. If a pump shaft breaks, the motor current will drop suddenly and production will be interrupted. Pump shaft breakage can be caused by poor pump assembly or pump deterioration.
Developments in sensors and data acquisition systems have enabled ESP systems to continuously record inlet pressure, inlet temperature, pump head, outlet pressure, outlet temperature, motor temperature, motor current, leakage current, vibration, and other data. This data is recorded at defined intervals and transmitted to the ground Remote Terminal Units (RTUs). For those wells with broken pump shafts, statistical or machine learning algorithms can be utilized for failure analysis and health monitoring.
The aim of this paper is to evaluate principal component analysis (PCA) as a monitoring tool for predicting ESP shaft breakage.
ESP Data Analysis
Over the years, the electric submersible pump (ESP) operating system has been developed and recognized as an effective means of extracting crude oil under wellbore conditions. ESP systems have attracted increasing attention due to the rapid development of electric submersible pump sensors in the petroleum industry.ESP sensors collect a large amount of data, including dynamic, static, and historical data, as shown in the figure. In the past few years, well sites using ESP technology needed to be tracked by field visits, which required a lot of human resources to make the necessary adjustments to the wellhead operating system. The oil and gas industry has begun to apply data analytics to improve production efficiency proposing ammeter charts as the primary diagnostic method for monitoring ESP performance.
In addition, the advent of Supervisory Control and Data Acquisition (SCADA) systems has facilitated the ability of field personnel to monitor and control the behavior of ESP wells.SCADA systems enable real-time continuous recording of ESP production data.
In recent years, "big data" collected by ESP sensors has become the focus of evaluating the most critical information extracted from ESP operating systems. Data-driven models incorporating machine learning have been used to determine and optimize well production.Bravo, Rodriguez, Saputelli, and Rivas Echevarria (2014) describe data analytics as an important process for collecting and analyzing big data.
Pump shaft breakage is a common failure in ESP operating systems. Data-driven modeling analysis, based on operational data acquired by ESP surface and downhole sensors, will play an important role in monitoring pump shaft breakage. There is a need to evolve from supervisory methods to data-driven model-based methods for fault diagnosis and predictive maintenance. Principal Component Analysis (PCA) is widely recognized as a preprocessing method for dimensionality reduction, eigenvalue extraction, and data visualization.PCA can be used as an unsupervised machine learning technique to analyze the causes of pump shaft breakage.
principal component analysis (PCA)
Jackson (2005) defines PCA as an unsupervised dimensionality reduction method that linearly transforms data and creates a new set of parameters called "principal components". What we know is that ESP data are often highly correlated; for example, an increase in wellhead pressure leads to an increase in inlet and outlet pressures, which ultimately leads to an increase in motor temperatures.PCA utilizes the correlation of the raw data to construct a PCA model that reduces the dimensionality of the production parameters by utilizing linear combinations and creating a new space of principal components (PCs). These PCs simplify the process by allowing the ESP system to be evaluated by a few principal components.
For those wells where pump shaft breakage exists, a PCA model can be built to analyze production data from the months prior to the pump breakage. Once a robust PCA model has been established, the cause of pump shaft breakage can be monitored and diagnosed. The basic PCA model can be represented as follows:
where X is the input matrix (n*p), denoting the original parameters; P is the loading matrix (p*k), denoting the contribution of the original parameters; T is the score matrix (n*k), denoting the relationship between the original parameters; E is the residual matrix (n*p), denoting the uncaptured variance; n is the number of time steps (Gupta, Saputelli, and Nikolaou, 2016); p is the number of number of original parameters; k is the number of principal components.
The first principal component contains the largest variance, meaning that the first principal component contains the most information. The second principal component will capture the next largest variance, having removed the information from the first principal component. In this way, the third, fourth...kth principal component can be constructed to evaluate the original system. The following figure summarizes the above.
Typically, PCA models find k principal components to construct PCs, which retains most of the information belonging to the initial system. Taking PC1 as an example, the kth principal component is represented as follows:
When dealing with two-dimensional data, Ionita and Schiopu (2010) describe this situation using the figure below.PCA is used to discover patterns in high-dimensional data and transform data in stable regions, often described as tightly aggregated or cloudy datasets. Anomalies in the ESP operating system can be detected by building a PCA model that corresponds to a normal production dataset. The first two principal components have the largest variance, and most of the information in the raw parameters can be visualized by the first two principal components alone.
PCA diagnostic model
The PCA diagnostic model is applied to identify the cause and timing of pump shaft breakage.The Hotelling T-squared statistic (T^2) and squared prediction error (SPE) are used to numerically present scalar statistics (Yue and Qin, 2001).T^2 is a univariate statistic that plays an important role in multivariate hypothesis testing, and SPE is often used for multivariate statistical process control.T^2 and SPE are applied to analyze whether decision variables meet the requirements for stable operation. In this way, the contribution of each decision variable to the fault can be determined. Fault problems will ultimately be diagnosed by ranking the contribution of each decision variable. Potential anomalies are associated with a higher or highest ranking of the relevant decision variable.
The time steps for T^2 and SPE are described below:
where x(t) is the tth time step of the input matrix; A^-1 is the inverse matrix of the covariance matrix (Westerhuis, Gurden, and Smilde, 2000); P is the eigenvectors of the covariance matrix; Pe is the residual loading matrix; and I is the unit matrix;
Case Study: ESP Broken Shaft Diagnosis
1. Selection of ESP break-axis variables
Production data from the ten ESP wells that experienced pumpshaft fractures were recorded by ESP downhole and surface sensors at 20-minute intervals. These ESP downhole and surface sensors began collecting production data from the time the ten ESP wells were placed on production until the pumpshaft fractures occurred in these wells. These ESP wells with broken pump shafts include E52ST1, C06ST1, B50ST2, E20ST2, A11ST1, B03ST1, B48ST1, E21ST1, E47ST1, and E42ST1 in Penglai block of Bohai oil field in China. two different types of data sets were collected including:
- Contains data on input variable parameters such as casing throttle, casing line pressure, casing pressure, casing gas flow, ESP inlet pressure, ESP outlet pressure, line pressure, line temperature, inlet temperature, motor current, motor leakage current, motor power, motor temperature, motor torque current, motor vibration, motor voltage, tubing throttle, and frequency of the inverter Recording.
- Contains a data record of the time of occurrence of pump shaft breakage for each well.
2. Principal component scores
A PCA model was constructed based on the captured input variables. The different principal components were ranked according to the order of decreasing variance captured. Taking well E52ST1 as an example, it was observed that eight principal components captured more than 99% of the variance of the original input parameters, as shown in Fig.
The first and second principal components had the highest variance and captured approximately 701 TP3T of variance in the raw data. Using a two-dimensional plot of the scores of principal components 1 and 2, different clusters were observed for the stable, unstable and faulty periods. During the stabilization period, the ESP operates normally and all the input variable parameters are within the normal operating range. When entering the unstable period, some input variable parameters are obviously abnormal, but the ESP still operates. In addition, pump shaft breakage and ESP failure occurred during the failure period.
The figure below shows a plot of the principal component 1 and principal component 2 scores for the historical data of pump shaft breakage in well E52ST1. The results clearly show three different clusters of stable, unstable and faulty areas as the time step increases over this time period. At the beginning, the ESP is put into production. It is observed that the input variable parameters for normal operation form a clustering of stable regions. After a long period of operation, some of the input variable parameters start to deviate from the normal operating range, but the ESP is still in operation. When this abnormal behavior occurs, an unstable region that deviates from the stable region is formed. the ESP continues to operate for an unstable period of time. Finally, the pump shaft breaks and the ESP fails. As can be observed in the figure below, the black faulty region is far from the stable region. This two-dimensional score plot of Principal Component 1 and Principal Component 2 is important for real-time monitoring of the ESP performance to compare it with the previous normal operating region and to warn the field engineer in advance of possible failures if the clustering starts to deviate from the stable region.
3. Pump shaft breakage identification
The production data obtained from the stable region is normalized and used as an input matrix (Xtraining) for the construction of the robust PCA model. In addition, historical data corresponding to the time period of the unstable or faulty region is selected as a test data set (Xtesting) and input to the PCA model. This process can be repeated for the historical pump shaft fracture events that led to the failure. In addition, a PCA diagnostic model is built to predict when pump shaft breaks occur and to identify the decision variables most responsible for pump shaft breaks.
Based on the formula, the contribution of each decision variable can be calculated by the PCA diagnostic model. Decision variables with higher or highest contribution are more relevant to pump shaft fracture. Decision variables are ranked according to their contribution to the E52ST1 well, for example, as shown in the following figure, the motor torque current has the highest contribution to the fault region. Therefore, if the clustering begins to deviate from the stable region, this contribution map can be used to diagnose the decision variables that are most responsible for pump shaft breakage on a real-time monitoring platform.
After constructing a robust PCA model, it is possible to predict the time of occurrence of pump shaft fracture. An image (IMAGE) and SPE equations are used to determine the time of potential anomalies before pump shaft fracture.Dunia and Joe Qin (1998) presented four possible detections under the PCA diagnostic model as follows:
(1) Both image metrics and SPE metrics exceeded control limits;
(2) Neither the image metrics nor the SPE metrics exceeded the control limits;
(3) The image metrics exceeded the control limits, but the SPE metrics were not exceeded;
(4) The SPE indicator exceeds the control limit, but the image indicator is not exceeded.
Typically, test results (1) and (4) are usually considered as potential cases of pump shaft fracture. This paper contains data on the time of occurrence of ESP shaft fracture for each well. The predicted ESP shaft fracture time by the PCA diagnostic model was compared in detail with the actual ESP shaft fracture time.
Image metrics and SPE metrics were calculated to predict ESP production anomalies using wells E52ST1, CO6ST1, and B50ST1 as examples, as shown below. When the ESP shaft breaks, the image and SPE were used to determine the time of breakage using the detection results (1) and (4).
The following table shows a comparison of the PCA diagnostic model's prediction of the pump shaft breakage time with the actual ESP shaft breakage time. The analysis of the following table shows that the PCA diagnostic model predicts the pump shaft breakage time slightly earlier than the actual pump shaft breakage time. Therefore, the PCA diagnostic model has excellent accuracy in predicting the fracture time of ESP shaft breakage wells and real-time learning techniques. In addition, the PCA technique can be used as a basis for developing better tools to predict ESP failures.
| Wells didn't. | PCA model predictions | Actual breakage time |
|---|---|---|
| E52ST1 | 2019-5-26 13:40 | 2019-5-26 16:00 |
| CO6ST1 | 2019-5-27 22:20 | 2019-5-28 6:40 |
| B48ST1 | 2015-10-7 21:40 | 2015-10-8 16:20 |
| E20ST2 | 2019-9-12 10:20 | 2019-9-12 12:40 |
| B50ST2 | 2019-6-15 8:00 | 2019-6-16 15:40 |
| A11ST1 | 2015-5-10 4:40 | 2015-5-10 10:00 |
| B03ST1 | 2015-8-30 15:40 | 2015-8-31 1:00 |
| E21ST1 | 2015-8-26 10:00 | 2015-8-26 23:40 |
| E47ST1 | 2018-1-14 8:40 | 2018-1-15 9:20 |
| E42ST1 | 2018-4-24 22:40 | 2018-4-25 2:00 |
reach a verdict
This thesis presents a model based on big data analytics for predicting impending pump shaft fractures in ESP operating systems. This big data model relies on real-time data collected by ESP downhole and surface sensors. It can be concluded that PCA has the potential to be used as an identification technique for predicting dynamic changes and therefore identifying impending ESP pump shaft fractures. The main conclusions of this study can be summarized as follows:
- A two-dimensional plot utilizing the scores of Principal Component 1 and Principal Component 2 can be used to identify different clusters of stable, unstable and faulty areas. With this 2D plot, if the clusters are far away from the stable region, the field engineer will be alerted to the risk of a possible ESP pump shaft fracture.
- Once a robust PCA diagnostic model has been developed, identifying the decision variables that lead to ESP pump shaft breakage is important to explain the deviation of the clusters from the stable region.
- By applying images and SPE equations, the PCA diagnostic model has excellent accuracy in predicting the time to ESP pump shaft fracture.
- PCA can be used as an important preprocessing method and unsupervised machine learning technique for predicting the development of ESP faults.
English 
Chinese 
German