Machine learning streamlines tubular connection analysis
A quick internet search will reveal a trove of definitions and detailed information on machine learning. Amidst it all, readers will find a common thread. That is, machine learning is a branch of artificial intelligence which, when employed, learns from experience and uses the knowledge gained to predict future system states. It is a broad technology that consists of many methods well suited to solving problems of regression, classification, and anomaly detection. As a result, it has found use in applications ranging from credit card fraud detection to analysis of corroding systems.
In a recent example, machine learning was applied to the make-up analysis of tubular connections. In this operation, data are collected from sensors embedded in the make-up machinery. These data often are presented to the operating technician in graphical form as a plot of torque versus revolutions, or turns, of the pin member of the connection. An example of such a graph is presented in Figure 1.
Upon completion of the make-up process, the technician is tasked with reviewing key features of the graph and comparing them to acceptance criteria to establish disposition of the connection. In the event of a rejection, the technician must diagnose the objectionable condition to determine if the connection can be salvaged, and how. To do so, he or she relies mostly upon training and experience. This is where machine learning can provide benefit.
It is generally accepted that computers are at least as capable of recalling information and associated relationships amongst data as are humans. Therefore, machine learning, when applied to the example under discussion, should be able to effectively classify and diagnose tubular connections with little or no human intervention.
Machine learning workflow
To explore this hypothesis, the workflow shown in Figure 2 was followed. As a first step, the objectives of this project were defined as:
- Develop an intelligent system for automatically classifying torque-turn graphs.
- Use the intelligent system for anomaly detection and prediction to reduce or eliminate damage to connections before it occurs.
Understanding your data
In the second step, the goal was to understand the data available for training the model. This includes answering questions such as:
- What variables or features are important to solving the problem?
- How are the data stored? In what form?
- Is the scope of the data broad enough?
The impact of a given variable may not always be known with certainty at this point in the process. If a variable does not strongly affect the result, it will be apparent when model performance is assessed in subsequent steps. It is therefore acceptable to include these data, but caution should be taken to not include too many variables for which the impact is not well understood. This avoids wasting efforts in building a model only to find that it fails in verification testing.
It is also important to ensure that the data set available is broad enough to encompass the range of input variables expected to be encountered in deployed operations. Trained algorithms perform best when exposed to data like those used to train them. If there is significant deviation from the scope of the training set, the model will fail to provide accurate responses when queried.
Data preparation is critical to any machine learning problem. During this step, data are first extracted for training. This may include the full data population or a random sampling of the data if the population is quite large. Cleansing of the data follows, to standardize data and formatting. Last, the data are verified for accuracy. As these data will be used for training the model, it is imperative that they are accurate. Failure to properly prepare the data for training can result in a model that may provide erroneous or misleading results.
Model building and training
At this point the objectives are defined, and the data selected and prepared for use. The next step is to begin building and training a machine learning model that meets the project goals. Dozens of machine learning algorithms are in common use; a detailed discussion of which would be beyond the scope of this article. However, it is worth noting that these algorithms generally can be grouped by the type of problem that they best solve, narrowing down the list of appropriate algorithms once the problem is clearly defined.
Multiple methods are used to train a machine learning model. Two methods commonly employed are those of supervised and unsupervised learning. In supervised learning, input data are provided to the model with corresponding outcomes. This method is often used for regression and classification problems.
The alternative method is unsupervised learning, which finds common use in anomaly detection problems. It differs from supervised learning in that there are no corresponding outcomes provided to the system against which the model is trained.
In both methods, the model is said to be trained when an acceptable level of maximum error has been achieved. Training is often an iterative process, the results of which are best summarized in the form of a confusion matrix. For example, the results from the first training iteration, as tested on new data (i.e., data not used for training) are shown in Table 1.
Upon completion of performance verification and final tuning, the trained model is ready to be deployed for use. Deployments may take many forms, including local hosting on a single computer, networked access, or web deployment for access from anywhere in the world. Because the bulk of the computing horsepower is required during the training operation, the deployed model often may be ported to much less powerful platforms, even mobile devices.
Even after the model is deployed, the development process does not end. Model performance should be monitored, at least periodically, to ensure that it is meeting real-world demands. It is not uncommon for model adjustments to be required after deployment. This is not necessarily a sign of a poorly performing model, but instead part of the continuous improvement process. As the model is exposed to more and more data, it will become “smarter” and “learn” from its new experiences, ultimately improving performance.
The model described as an example for connection make-up analysis was successfully deployed by Frank’s International as the intelligent connection analyzed make-up (iCAM) technology (see Figure 3). Deployment occurred after final optimization, fine tuning, and performance verification of the model to provide the requisite levels of accuracy and capability. Although not described herein, additional features have been and will continue to be added to this technology through future updates.
As the example shows, machine learning can be used effectively to analyze the connection make-up process. Doing so offers many benefits. First, the use of machine learning removes human subjectivity from the analysis, providing more consistent and accurate results. While most human operator training occurs under globally standardized programs, the experience gained by each operator will vary greatly and is limited in scope and size.
Conversely, a properly trained machine learning model incurs the benefit of learning from the collective experience of all recorded connection make-ups and recalling that information when required. Thus, the experience of a trained machine learning model equates to many lifetime’s worth of experience for a human operator.
The implementation of machine learning technology for connection evaluation also has advantages over rules-based systems in use today. Establishing rules for some conditions is straightforward (e.g., minimum torque, maximum torque, and others), but not so for others (e.g., dope squeeze, high interference, and poor sealing). To develop the necessary rules to automatically evaluate a given connection, we must create a rule or series of rules for each acceptable and rejectable condition. Those rules must be based off the prior knowledge of each condition’s signature and contributing factors.
Machine learning overcomes this challenge through its ability to identify patterns and relationships that may be imperceptible to human analysis. As a result, the trained model generally can identify and evaluate more conditions than a rules-based system and can do so more accurately and efficiently. Machine learning, as implemented here, also provides the additional benefits of predictive analysis and anomaly detection, preventing damage to threaded connections before it occurs. This translates into both time and cost savings for the tubular running operation.
From an operational perspective, implementation of automated connection analysis can reduce personnel on the rig floor. This may be achieved in two scenarios. First, using technology, such as Frank’s iCAM, allows for the computer to make a final determination of connection integrity without human intervention, removing operating technicians from the process all together.
In an alternative scenario, the human operator may be removed from the rig floor and instead monitor multiple jobs from a centralized shore location, intervening when and if necessary. This is made possible by real-time data communication technologies, such as Frank’s DISPLAY system (Figure 3), that facilitate monitoring of the connection make-up process from anywhere in the world.
As the oil & gas industry continues seeking gains in efficiency and well integrity, technologies such as machine learning are expected to play ever-increasing roles in normal operations. In a real-world application example, machine learning has been successfully applied to the evaluation of tubular connection make-ups.
The resulting benefits include accurate and consistent evaluation, real-time prediction and anomaly detection to prevent connection damage, and removing personnel from the rig floor. When combined, these translate into cost savings and improvements to both well integrity and personnel safety.
Brennan Domec, PhD. PE is director, strategic technology, Frank’s International, LLC.