Case studies:
Predicting Stuck Pipes during Drilling


The problem of stuck pipes accounts for several billions of dollars worth of non-productive time for oil & gas operators each year. Developing a method to predict this event in real-time has become high priority for the drilling industry. Automatically predicting stuck pipe events is now possible due to modern sensor techniques and advanced data analysis tools.



  • Noesis conducted a comprehensive study¬†that uses machine learning for the accurate prediction of stuck pipes.


  • First, sensors and other devices provide useful information characterizing the conditions surrounding pipes during drilling. Examples of variables involved in the monitoring of the drilling process are the pore pressure, hook load, rate of penetration, weight on bit, standpipe pressure, etc. An initial preprocessing step aimed at identifying a set of variables exhibiting high predictive power (i.e., high correlation with stuck pipe).


  • Analysis then uses multiple machine learning tools (e.g., neural networks, support vector machines, decision trees, nearest neighbor, Bayesian methods) for prediction, allowing a time-window warning before the actual failure occurs. We finally report on the use of Markov Models to establish a degree of dependency in the predictions at different time steps.



img2The experiments considered time-based data with fifteen minutes, thirty minutes, and one hour warning windows. Ensemble methods were incorporated into our analysis. We found these methods extremely accurate.

User interfaces were generated that provide information to gas & oil operators during real time drilling operations. Both interfaces are intended for scenarios where information regarding the possibility of a stuck pipe should be made in a direct and concise manner. The user interface indicates the probability of a stuck pipe divided into three regions: low, medium, and high. The right hand side of the panel display indicates the status of the current drilling operation, along with the values of variables captured by sensors in real time.


Conclusions and Benefits

The project shows a methodology to predict the occurrence of stuck pipes based on historical data from previous drilling operations. A comprehensive analysis of multiple learning algorithms on drilling data using multiple variables from existing sensors was conducted.

The methodology can be applied to other problems where the goal is to predict machine failure within a time window. It simply needs sensors continually monitoring a particular machine in operation to generate historical data. Machine learning tools can then produce accurate predictive models to prevent future failures.