Automatic event identification and extraction from daily drilling reports using an expert system and artificial intelligence

• Published in: Journal of petroleum science & engineering Vol. 205; p. 108939
• Main Authors: Cinelli, Lucas P., de Oliveira, José F.L., de Pinho, Vinicius M., Passos, Wesley L., Padilla, Rafael, Braz, Patrick F., Galves, Breno, Dalvi, Domenica P., Lewenfus, Gabriela, Ferreira, Jonathas O., Ji, Anthony Y.Y., de Oliveira, Felipe L., Gonçalves, Clemente J.C., Netto, Sergio L., da Silva, Eduardo A.B., de Campos, Marcello L.R.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2021
• Subjects: Automatic event extraction; Daily drilling report; Drilling data log; Natural language processing; Oil well drilling; Daily drilling report; Oil well drilling; Automatic event extraction; Natural language processing; Drilling data log
• ISSN: 0920-4105; 1873-4715
• DOI: 10.1016/j.petrol.2021.108939

This work addresses the problem of extracting events from human-written daily drilling reports (DDRs) in an automated way. Two distinct approaches based on an expert system and artificial intelligence techniques are proposed: rule-based language processing (RBLP) and deep neural networks (DNN). The RBLP employs regular expressions that are manually constructed, during the so-called building process, in order to identify the events of interest. The novelty of the present approach is to deal with multi-label classification of DDRs using RBLP and transformers, which provide a powerful DNN architecture. The events of interest are drilling failures such as ‘bump’, ‘drag’, ‘kick’, ‘loss of circulation’, and ‘stuck pipe’. Both algorithms are developed based on a training data set of 4,355 DDRs and evaluated on a test data set of 300 DDRs, all of them written in Brazilian Portuguese but can be readily adapted/replicated to any other language. Average true positive rates (TPR) of 97.30% for RBLP and 85.61% for transformers-DNN were obtained, with average false negative rates (FNR) of 2.70% and 14.39%, respectively. The corresponding false positive rates (FPR) were 4.90% and 13.52%. Transformers-DNN has superior performance if the underrepresented classes are disregarded. In this case, the average TPR was 96.79% for RBLP and 97.32% for transformers-DNN, with an average FNR of 3.21% and 2.68%, respectively. The corresponding FPR changed to 2.37% and 1.81%. The test results indicate that the two proposed approaches can lead to very significant improvements in the efficiency of the otherwise manual annotation processes, which are typically error prone and very time consuming. •Rule-based language processing and deep neural networks for information extraction.•Identification of the most frequent faults occurring during drilling process.•Results validated using a comprehensive daily drilling report (DDR) database.•Multilabel classification of daily drilling report failures.•State-of-the-art DNN techniques and a framework with few hyperparameters.