DIGITAL SAMPLING: MULTIVARIATE PATTERN RECOGNITION,MACHINE LEARNING, AND EQUATION OF STATE. A REAL-TIME APPROACH TO EVALUATE CLEAN FORMATION-FLUID PROPERTIES AND MUD-FILTRATE CONTAMINATION

摘要

Acquiring physical samples from an open hole is usuallya one-opportunity event where a formation tester is sentdownhole with a limited number of sample chambers,either on a logging-while-drilling (LWD) or wirelineconveyance system. The samples are acquired, retrieved,and sent to a laboratory for analysis, which takes placeweeks to months later. By the time the laboratory hasperformed an analysis, the section has been cemented,and perhaps the rig has finished operations and movedonto the next phase. Success of the sampling operation ispredicated on the samples being acquired from the rightlocations (where to sample?), at the right time tominimize drilling fluid-filtrate contamination (when tosample?), and in a manner that preserves the integrity ofthe sample and is representative of the formation fluid(how to sample?). Digital sampling is a technique thatthat can be used to both optimize the when, where, andhow of physical samples taken and further augment theinformation collected with sensor analysis fromlocations that are not physically sampled.This work shows a new workflow that can be used toextrapolate clean fluid properties with moderately highcontaminationlevels in a rapid pumpout. Based on theextrapolated clean fluid properties, an operator canmake a decision whether to continue the pumpout toobtain physical samples or abort the pumpout if the fluidproperties extrapolated (digital sampling) at the locationare sufficient for the operation decision making. Theworkflow starts with applying principal componentanalysis (PCA) to a multichannel sensor measurement offluid pumped out of the formation during a formation testsampling operation. Because the fluid pumped outcontains only two endmembers (clean formation fluidand mud filtrate), the PCA scores of sensormeasurements form a line in the PCA space, and solutionbands of endmembers can be estimated based onphysical constraint of sensor measurements (nonnegative etc.). Then, a trend-fitting method is used topredict the asymptote of the first principal componentscore. The asymptote value can be inverted to sensorsignal using PCA inversion, and the sensor signalrepresents the clean formation-fluid measurement.Lastly, machine-learning-based composition models canbe used to predict the clean fluid compositions based onthe sensor signal. The composition data then is used topredict fluid physical properties, such as bubblepoint,viscosity, and compressibility, using an Equation ofState (EOS) model.A series of rapid pumpouts at different depths can beused to map a formation for selection of where to sample,constrain contamination models to improvecontamination estimation, determine when to sample,and optimize the pumpout parameters to obtain arepresentative sample in the shortest period of time.We have applied this workflow to a number of formationsampling jobs at multiple wells, the realtime resultsmatch with the laboratory analysis result in term ofcontamination level and clean fluid properties(compositions, GOR, bubblepoint, density, etc.)