| Abstract:Although COSL has successfully developed downhole optical spectrum analyzers, there are still shortcomings in interpretation, especially intelligent fluid identification driven by downhole spectral data in real time. In this paper, an intelligent fluid identification method driven by spectral data in real time is proposed for the first time. Firstly, based on a large number of measured spectral and fluid data, a database of various fluids is established. The spectral data and fluid types in the database are then preprocessed separately. The spectral data of 256 channels are further reduced by principal component analysis, and the first 10 principal components are chosen as input variables. Fluids are divided into 5 types (output variables): gas, oil, water, slurry (emulsified fluid, or invalid measurement), oil/gas-water mixture. Next, 23 pattern recognition models are used for modeling of fluid identification, including various trees, discriminants, support vector machines, K-nearest neighbor methods, artificial neural networks and so on. All the models are then analyzed and compared, and the best model is selected to be embedded in the formation test tool. The test accuracy of the 23 models after training is 78.1% - 99.9%. The artificial neural network is the best (accuracy of 99.9%). Finally, the spectral data from more than 40 sampling stations are employed to further examine the best model. The results show that the best model can accurately predict fluid types. The new intelligent identification method lays a solid foundation for eliminating the influence of water on the oil and gas spectra and analyzing the composition of oil and gas more accurately, thus reducing operation risks and saving operation time and costs. |