Research On Intelligent Logging Lithofacies Identification In Complex Wellbore Environment

Currently,China’s petroleum development has entered the middle stage.The main target of geological exploration has gradually shifted to unconventional reservoirs,which are extremely difficult to explore.Hence,the petroleum exploration technology is in urgent need of innovative development.Geophysical logging is a key technology for petroleum exploration.It measures various geophysical properties of formations at different depths in the wellbore with specific equipments.Based on these observational data,lithofacies identification aims to discriminate the lithologic composition of the formation,which provides the basis for real-time drilling,geological evaluation,and reservoir modeling.The traditional logging lithofacies identification is done by empirical logging-lithofacies models established by geophysicists.With the accumulation of logging and geological data and the shift of exploration target to the heterogeneous reservoir,the limitation of traditional logging lithofacies identification mothods in characterizing the complex relationship between loggings and geological parameters is emerging.Besides,the identification performance hits a bottleneck and the identification efficiency needs to be further improved.Therefore,it is necessary to develop intelligent logging lithofacies identification methods to fully mine the potential relationship between logging signals and geological information,which is of great significance for boosting the development of artificial intelligence in the petroleum industry,deepening the understanding of geologic structures,and improving the efficiency of petroleum exploration.Affected by natural and non-natura factors such as sedimentary environment,geologic structure,and the intrusion of mud and mud cake during drilling operation,the wellbore environment is very complex,resulting in the following problems:First,there are discrepancies in the distribution of logging data on different wells.Second,it is difficult to frequently perform coring operations on the target well to obtain reliable lithofacies labels.Third,the responses of logging curves are heterogeneous,making it difficult to capture the response characteristics accurately.These problems greatly affect the accuracy of logging lithofacies identification.To solve the aforementioned problems,this thesis conducts research from the following two aspects.On the one hand,this thesis concentrates on two typical distribution discrepancies of logging data,namely,marginal distribution discrepancy and joint distribution discrepancy.Specifically,this thesis studies how to reduce the distribution discrepancy of logging data from different wells,by utilizing the hybird constraints of the physical properties of loggings and the deposition laws of formations and querying the labels of the most informative samples,in the case of limited target labels.On the other hand,this thesis concentrates on the heterogeneous responses of logging curves.Specifically,this thesis studies the representation learning of the combined patterns of shapes and numerical values of logging curves,based on the principles of logging responses,to extract the logging features that are general to different wells,so as to overcome the problem of distribution discrepancy between logging data from different wells.The main work of this thesis can be summarized as follows:(1)To solve the problem of marginal distribution discrepancy of logging data from different wells,this thesis proposes an unsupervised logging transfer method based on hybrid physical constraints.First,the high-level feature representations of the source domain and target domain are jointly extracted.Then,a domain discrepancy loss is imposed on the feature representations to align the feature distributions of the two domains;and simultaneously,label smoothness constraints of feature space and geographic space are imposed on the target-domain feature representation to preserve the latent structure of target unlabeled data,therby prompting the feature to be matched with the correct class during the alignment process.Finally,the model is supervised under the source labels to learn the features that are both domain invariant and discriminative.The training process is monitored by a robust early stopping criterion,allowing the model to terminate its iterations adaptively.Experimental results on real-world logging dataset demonstrate that the proposed method outperforms existing methods on each sourcetarget combination in lithofacies identification,implying that it can effectively reduce the marginal distribution discrepancy of logging data.(2)To solve the problem of joint distribution discrepancy of logging data from different wells,this thesis proposes a logging transfer method based on active adaptation with maximum discrepancy.First,two classifiers with maximum discrepancy are trained on the source domain.Based on the prediction results and confidences of the two classifiers,the target samples are selectively screened for requesting true labels or assigning pseudo-labels.Then,the reliability of the pseudo-labels is detected based on the expected risk reduction criterion,with the aid of the queried target true labels.Afterwards,the rectified pseudo-labeled samples,together with the true-labeled target samples,are used to weight the source instances,thereby promoting the source distribution to approach the target distribution.By repeating the above learning process,the number of target labels can be gradually enriched,and the distribution discrepancy between the source domain and the target domain can be gradually reduced.Experimental results on real-world logging dataset demonstrate that,compared with existing methods,the proposed method can more effectively alleviate the performance degradation of lithofacies prediction caused by the joint distribution discrepancy of loggings,with fewer target label queries.(3)To solve the problem of heterogeneous responses of logging curves,this thesis proposes a logging feature extraction method based on a pixel-enhanced fully convolution.First,the logging-lithofacies mapping relationship is modeled as semantic segmentation.Then,a statistics-guided pixel enhancement module is designed to capture the micro-detailed features of loggings.This module generates global statistical embedding for each logging,emphasizes or suppresses specific logging values with the statistical embedding,and exchanges information between logging values.Finally,the micro-detailed features are fused with the macro-semantic features,extracted by a backbone fully convolutional segmentation network,to form aggregated feature representations,which are mapped to pixel-level lithofacies output.Experimental results on real-world logging dataset demonstrate that modeling the logging-lithofacies mapping relationship as semantic segmentation and extracting the combined patterns of logging curves can better reflect the response characteristics of logging curves than traditional modeling strategies,and thus achieves better performance for lithofacies identification.In summary,this thesis solves the problem of logging distribution discrepancy from different perspectives by explicitly reducing the distribution discrepancy under two typical distribution discrepancies and implicitly extracting general features of combined patterns of loggings.Therefore,this thesis realizes effective lithofacies identification in complex wellbore environment and enriches the intelligent logging lithofacies identification methodology,which provides feasible technologies for determining the relationship between logging data and geological information and improving the efficiency of petroleum exploration.