Study Of Cuttings Identification Using Laser Induced Breakdown Spectroscopy

Cutting logging is an important technique in geo-logging whether for oil exploration ordrilling engineering, and cutting identification is one of the most vital links in the processof cutting logging. Up to now lots of work was down by our laboratory, whichdeveloped a new kind of cutting identification method based on digital image processingtechnique, to displace the traditional method with the naked eye. Laser-inducedbreakdown spectroscopy (LIBS) has been shown great potential in geo-logging field withthe advantages of rapid analysis, simultaneous multi-element detection, in-situ andstand-off analysis capability. However, high-resolution, broadband LIBS spectra have abig data size which contains thousands of variables including atomic emission lines, ionicemission lines, molecular bands, and background emission. Additionally, theshot-to-shot variation is an inherent issue with all LIBS experiments due to the lasermaterial interaction, sample inhomogeneous, self-absorption and matrix effect. So theability to quickly process all of the data in a useful manner is a big challenge during LIBSused for qualitative identification. In this thesis, we used LIBS combined withchemometrics methods to realize the rapid and valid identification of different cuttingsfrom geo-logging field.The thesis began with a brief background introduction of this work, after that was thedescription of laser induced plasma’s fundamental, a detailed review of current situationof the pattern recognition based on LIBS, and a brief introduction of the experimental anddata processing methods used in this thesis. Then the bulk of author’s contribution,within the general group effort, was described just as follows.The LIBS spectra of6cutting samples from geo-logging field were collected to get theoriginal LIBS data. An experimental set up was built based on the thesis purpose andlaboratory conditions, and then after wavelength and intensity calibration we didtime-resolved measurement to get the best LIBS signal. The final laser energy wasdetermined as15mJ, and the detection delay and width were respectively800ns and8μs.At last we obtained the original100single-shot spectra from each kind of cutting sample.PLS-DA method was chosen for the LIBS data processing by an optimized design of the model analysis. Firstly we did some spectra pretreatment and intensity normalizationapproach was determined for the best. Then the analysis process of PLS-DA model forclassification was designed in detail: all samples were divided into the calibration set andtest set, combined with LOO-CV method to select the best latent variables; after that twodiscriminant approaches were used with the name of highest probability method andCMU method, and then we determined several indexes to indicate the performance of thePLS-DA model, including RMSEP, percent classified correctly, misclassified, unclassified,effectively unclassified and run time of the model.On the basis of the PLS-DA design, feature extraction method was studied to reduceinterference of the LIBS noise and background in the whole spectra. We built thefeature model with27variables which were selected from the whole24041variables, andcompared to the whole spectra model in sensitivity, robustness and efficiency. In theaspect of sensitivity, the RMSEP of feature model was lower than whole spectra model,0.2681vs.0.3250, and the feature model could give a higher percent classified correctlyas well whether in highest probability method or CMU method,88.33%vs.86.67%,81.67vs.68.33%, respectively. Also it showed a better performance in robustness andefficiency through feature extraction, with50%vs.33%for percent effectivelyunclassified and0.1s vs.7.5s for run time of the two PLS-DA models.Additionally, a preliminary research of SVM method used in cuttings’ identification hasbeen taken to solve the non-linear problem that existed in LIBS data, because essentiallythe PLS-DA model above is a kind of linear method. Compared with PLS-DA model,the percent classified correctly of whole spectra SVM model and feature SVM modelwere93.33%vs.86.67,86.67%vs.88.33%, respectively. After that a fusion method ofPLS-DA and SVM was developed which can give the best performance while modelpredicting, with the percent classified correctly of95%for whole spectra input and91.67%for feature input. The obtained results demonstrate that compared with PLS-DA,SVM model has great potential to be used in cuttings’ identification as a non-linearmethod.Finally, a summary of the work was presented, and followed by some suggests of possiblefuture developments including the extension of SVM method and the sensor fusion ofLIBS combined with digital image processing technique.