Page 13: of Offshore Engineer Magazine (May/Jun 2020)

12 14

DRILLING & COMPLETIONS • detector inspection dataset, containing more than 2,000 re- Numerical test results should be stored as separate felds in cords, from an offshore operator. Generally, such records the management system. This will greatly simplify subsequent contain information regarding failures or miscalibration. The machine analysis and trending • study used a linear support vector machine (SVM) algorithm Operators should consider revising how they report tests in and natural language text processing to classify inspection re- their management systems as some types of text, such as guid- cords based on the free-text feld. ance notes (boilerplate text), can detrimentally impact the pre-

Before using an ML model, the dataset was processed and dictive power of ML models • cleaned. This transformed the raw maintenance text data into Existing records should be reworked with additional felds, the limited feature set. While it is the most time-consuming to facilitate future verifcation and trending activities • part of the program, the importance of maintaining a high- Training is essential to demonstrate and facilitate the po- quality dataset cannot be overstated. tential use of maintenance logs for trending. For example, staff

The ML model was then trained to make predictions. This leaving good comments and descriptions will make text clas- step used a subset of the data as a testing basis, while the ML sifcation easier.

model used the training subset to determine its internal pa-

CONCLUSION rameters to make predictions.

To verify if initial assumptions were correct, an offshore The DNV GL study demonstrated that a trained SVM al- verifer performed a manual verifcation of the entire dataset. gorithm can rapidly identify records with potential anomalies,

Overall, this gives confdence in the predictions based on the although the predictive power is dependent on the pre-pro- unverifed dataset. cessing cleaning steps. Additional prediction improvements

Nevertheless, the company re-trained the model using the can be gained by limiting the feature set to a known vocabu- verifed classes. DNV GL intends to further refne this ap- lary of important words and phrases, informed by expert proach as it is rolled out across its offshore verifcation services knowledge. This method works well, even for highly skewed and activities. datasets with few recorded fails (Figure 3).

By gathering data from a variety of assets in the North Sea, ML ultimately gives maintenance and reliability teams a the ML algorithm will develop to be more robust and reliable more focused approach to check records by specifcally target- (Figure 2). ing the anomalous maintenance records.

Several other classifcation models were also trained to com- Crucially, by reducing the amount of time spent reviewing pare results; however, it is diffcult to draw a direct compari- non-erroneous maintenance records, this method increases son between each of these classifers as each one is subjective project effciency and allows more useful fndings and rec- to its own set of hyperparameters. ommendations to be made. Importantly, this also enables

As part of the research, the company identifed several pit- operators to reconstruct data missing from their manage- falls, particularly in the way operators currently record their ment systems, as well as identify potential systemic misre- data. For instance: porting issues.

Figure 2: Confusion matrix showing ML accuracy against manually verifed training and testing data Figure 3: More time/focus on records of concern allows engineers to target the highest-risk threats

MAY/JUNE 2020 OFFSHORE ENGINEER 13