Using machine learning to detect mentions of drilling and operational problems in text. Over 5,000 public domain sentences have been labelled to train a predictive machine learning model to detect wellbore drilling and operational ‘problems’ (including reservoir and production) in documents, reports and logs.
This can support alerts & monitoring, health & safety, search & discovery as well as analogues & learning by quickly extracting where problems have or are being encountered, in volumes of unstructured text which are too vast for a person to ever realistically read through.
The model generalises, capable of surfacing types of problems that were not even present in the original training set. For example in the image shown, ‘swelling formation’ is detected in the first sentence as a potential problem. Whilst ‘swelling clays’ was labelled in the training set, the phrase ‘swelling formation’ was not. The model has inferred this based on statistical word context.
The techniques are also useful for scanned content with OCR errors, note that ‘stuck pipe’ and ‘swelling clays’ are detected even though they have spelling errors introduced by the OCR process – not uncommon!
This model adds to the existing CNN derived ML models in GeoClassifier(R) for detecting well names and subsurface topics. These models are fully integrated with the OpportunityFinder(R) algorithm for the energy transition, mining and geohealth sectors.
#digitaltransformation #machinelearning #naturallanguageprocessing #georesources