Research On Logging Curve Reconstruction Based On The Integration Of Multi-source Information And Machine Learning

In the field of oil exploration engineering,logging data is the first source of information about oil and gas resources in underground reservoirs.Geophysicists construct an accurate geological model from complete log curve data,then design the next strategy in the oil and gas development.However,logging curves are often missing or distorted due to some subjective or objective factors in the actual well exploration process,and several logging curves are complete loss in the worst case.These situations have caused great inconvenience for the subsequent geological modeling.Therefore,this paper uses the modern signal processing techniques combined with machine learning and deep learning methods to deeply explore the information representational relationships of inter-well logging curves,constructing a series of predictive models for logging curve recovery without additional costs.To address the problem that traditional machine learning models explore the mapping relationships insufficiently between logging curves and the lateral correlation relationships between different wells,a logging curve recovery method based on multi-dimensional feature set which composed by spatial-temporal-spectral features,and an integrated machine learning model is proposed.The method extracts the base features from three dimensions firstly,and also constructs an invariant feature system between different wells,which together construct a multi-dimensional feature set from both horizontal and vertical aspects.In addition,three machine learning models are used to build an integrated machine learning model to enhance the model’s ability to utilize information.Conducting experiments with actual engineering data from the oilfield,the method predicts the entire acoustic curves more accurately,while the multi-dimensional feature set gives a powerful outreach to the predictive model and greatly improves the robustness of the model.Although traditional neural network models can express the non-linear mapping relationships between logging curves,but they are not able to focus on the key information in a timely manner when faced with the complex and variable reservoir and geological information within a well.To address this problem,a long and short-term memory neural network(LSTM)model with attention representation is constructed by adding a layer of attention mechanism layer after the output layer of LSTM,which will help the predictive model focus on the key information by giving higher weight to important information.To reduce the redundancy of input and improve the training speed of the model,the input logging signal will be decomposed and the redundant components are removed after calculating the correlation of each modal component with the original signal.Experiments were conducted to recover several logging curves within multiple wells,and compared with other predictive models.The experimental results showed that this prediction method has higher recovery accuracy and higher stability.To further improve the accuracy of prediction and to enhance the stability and convergence of the model,an improved generative adversarial network(GAN)model for logging curve recovery is proposed,which enable GAN models not only used for generating images,but also employed in the field of reconstruction of non-linear time-series signals.The improved GAN model uses a three-layer LSTM as the generator,and a two-layer convolutional neural network(CNN)as the discriminator.Meanwhile,in order to solve the problem about non-convergence of the loss function when generating the discrete time series data,the multi-scale texture features of the logging signal are extracted from the gradient amplitude and gradient direction using the local binary model,which will further improves the prediction accuracy and stability of the model.Through several experiments,logging curves generated by this predictive method have extremely well correlated with the original curves,that really provide a new idea for the implementation of logging curve recovery.