Detection of Symptoms for Revealing Causes Leading to Drilling Failures

Abstract

When a diagnosis of a problem is known, the problem can usually be solved efficiently. This paper presents a method that helps reveal the most probable cause of a drilling-process failure immediately after occurrence. Normally, it takes some time to investigate and evaluate all available information before the correct cause can be determined. The method presented is targeted at reducing this time and at the same time improving the quality of the interpretation. The method relies on input parameters from the ongoing drilling process, parameters that behave irregularly and are referred to as symptoms or errors when the irregularity is severe. Recognized symptoms are used as input parameters in a knowledge-modeling method that relates symptoms to failures. The method was verified on the basis of historical drilling data, and its ability to point out the correct causes with high probability was demonstrated. Field testing the method is yet to be performed.

Drillstring Washout Diagnosis Using Friction Estimation and Statistical Change Detection

“…There a high fidelity model was fitted to data and used to detect abnormalities. Knowledge-modeling was used for classification of different incidents by [10] and a Bayesian network was shown to detect sensor and process faults in [11].…”

Incident detection and isolation in drilling using analytical redundancy relations

“…Detection of other incidents was studied in [3] using a high-fidelity model, and a knowledge-based method was used in [4]. Lost circulation, formation fluid influx, and drillstring washout have many similarities to the problem of leak diagnosis in open water channels, see, e.g., [5].…”

Fault diagnosis of downhole drilling incidents using adaptive observers and statistical change detection

“…If the multivariate Gaussian distribution is used these values are 1.6 × 10 −3 and 0.80, respectively, which is considerably higher. The t-distribution is hence providing better detection properties, where isolation in particular would be challenging using a multivariate Gaussian distribution in the GLRT decision function (11). This can also can be seen in Fig.…”

“…Simple hydraulics models and observers were used by [4,5], a high fidelity model was fitted to data in [9,10], and [11] applied a knowledge-modeling method. Due to measurement noise, a statistical cumulative sum (CUSUM) algorithm was tested on flow measurements in [2], and in [12], skewness of the statistical distribution was used to detect poor hole cleaning.…”