Core Drilling Accident Discrimination Model Based On Machine Learning

Because of the concealment,it's difficult to detect accidents of geological drilling engineering in the first place,which may cause economic loss,inefficiency and casualties during the drilling work.Recognition system of operating conditions plays an important role in drilling engineering,it can minimize the effect of accidents in the early state of emergencies.To achieve the effect of real-time recognition,combining with drilling monitor,intelligent classifying algorithms based on machine learning could map multiple parameters onto various operating conditions,have advantages in operating conditions which could not be expressed mathematically.In this paper,abnormal conditions-characteristic parameters relationship is established according to the analysis of parameters' changing trends in five abnormal conditions:bit-burnt accident,paste drilling,pipe sticking,drill pipe fracture and lost circulation.After the analysis and process of the obtained data,take the difference of drilling parameters as normal conditions' sample data,and the abnormal conditions' virtual sample is established by combining normal threshold with abnormal parameters' changing trends.The question that lack of samples in geological drilling engineering which is used in algorithm training is solved.Recognition of operating conditions in drilling engineering is a classification task in machine learning.This article compares multiple classification algorithms,choose k nearest-neighbor,support vector machine and logistic regression to build classification models,take Python as the programming language and use sklearn to achieve the whole procedure.After training and testing by 3 samples which have different abnormal sample size,9 obtained models are analyzed by several evaluation indexes,the accuracy rate is above 90%.According to the models' generalization abilities,optimize 2 k nearest-neighbor models,1 SVM model and 2 logistic models.As the final discriminant module,these 5 models in parallel adopt multiple decisions,the total accuracy rate is above 90% and the identifying speed is 30.5ms by testing.