Intelligent Evaluation Method of Cement Bond Quality Based on Convolutional Neural Network

The quality of cement bond is related to the safety of oil and gas well production and the service life of casing. At present, acoustic variable density logging (VDL) is the most widely used method for evaluating cementing quality in oil fields. The data interpretation of VDL still needs to rely on manpower, and the accuracy of interpretation results is restricted by human factors, and the workload is heavy. Oilfields have accumulated a large number of practically verified VDL interpretation results. It is of great research value and application potential to sort out these historical data and mine them with the help of deep learning technology, and establish an intelligent analysis method instead of humans to explain the cementing quality. In this study, the VDL cementing quality evaluation reports of several oil wells were collected. Through data preprocessing, the acoustic variable density images were standardized and segmented along the borehole direction. The cementation conditions of the first interface and the second interface corresponding to each segment of the acoustic variable density image were marked, and a sample set for cement bond quality evaluation was established. The cementing quality evaluation problem is transformed into an image classification problem, and the convolutional neural network method is introduced. On the basis of LeNet5, AlexNet and other classic image recognition architectures, considering the characteristics of acoustic variable density images, a personalized convolutional neural network (CBQNet) for cementing quality evaluation is designed, including 28 layers and more than 32 million learnable parameters. Using historical cementing quality evaluation samples to train and analyze the performance of convolutional neural network, the results show that: CBQNet has a training accuracy rate of 95.9% and a verification accuracy rate of 95.4% in the first interface cementing quality evaluation. In the cementing quality evaluation of the second interface, the training accuracy rate reached 90.8%, and the verification accuracy rate reached 88.1%. It shows that the convolutional neural network realizes efficient and accurate interpretation of cementing quality by mining and learning the interpretation results of historical VDL data, and provides a new method for cementing quality evaluation.