Abstract: The quality of cementing is crucial for the production efficiency and lifespan of oil and gas wells.Currently,the most widely used method is acoustic amplitude variable density logging for evaluation.However,the interpretation process is complex,and decisions related to major risks need to be made based on the results of cementing interpretation.Therefore,the evaluation of cementing quality must be undertaken by experienced experts,which is time-consuming and labor-intensive.In order to improve the efficiency of cementing interpretation,we used convolutional neural networks such as VGG and ResNet to automatically interpret cementing quality,but the accuracy was insufficient.Therefore,we proposes a method of parallel con-nection between multi-layer perceptions and convolutional neural networks(MLP-CNN),where acoustic amplitude data is input into multi-layer perceptions and variable density logging images are input into convolutional neural networks;We modifies the structure of convolutional neural networks by setting convolutional kernels of different sizes to extract information at different scales for features with varying density maps,such as the thickness,brightness,and shape of stripes.We used 9000 data from the Fuyuan block of the Tarim Oilfield for training and validation.The results showed that compared to traditional convolutional networks such as VGG and ResNet,the MLP and CNN parallel networks effectively improved the accuracy of cementing quality recognition,with an evaluation accuracy of 90%.Furthermore,compared to a single scale convolutional kernel,the convolutional neural network algorithm with multiple convolutional kernels of different sizes is more suitable for extracting features from variable density cementing images.We modified the structure of the convolutional neural network and established an MLP-CNN neural network with three convolutional kernels of different sizes,which improved the accuracy by 5%compared to the MLP-CNN model with a single convolutional kernel;meanwhile,we compared the time complexity and spatial complexity of seven networks.The findings revealed that the MLP-CNN parallel network efficiently mitigates a substantial number of ineffective convolutions,thereby reducing model computational costs and enhancing computational efficiency.Finally,in order to test the transferability of the model,we used 60000 data from the Manshen and Yueman blocks of the Tarim Oilfield for testing,and the evaluation accuracy reached 89%,indicating a satisfactory migration effect and robust performance of the model.

Key words: cementing quality evaluation, deep learning, convolutional neural network, multi-layer perceptron, image feature extraction

摘要： 固井质量的好坏关系到油气井的产量和寿命,目前最常用的方法是使用声幅—变密度测井进行评估,但是解释过程复杂,且与重大风险相关的决策需要根据固井解释结果做出.因此,固井质量评价必须由经验丰富的专家进行解释,耗时耗力.为了提高固井解释的效率,本文基于VGG、ResNet等卷积神经网络对固井质量进行自动解释,但是准确率不足.于是,本文提出一种多层感知机和卷积神经网络并联的方法(MLP-CNN),声幅数据输入到多层感知机中,变密度图输入卷积神经网络中;针对变密度图存在不同尺度信息的特征(条纹的粗细、明暗、形状),本文修改了卷积神经网络的结构,设置了大小不同的卷积核,提取不同尺度信息.本文使用了塔里木油田富源区块的 9000 个数据进行训练和验证,结果表明,相较于传统的VGG、ResNet等卷积网络,MLP和CNN并联网络有效提高了固井质量识别的准确率,评价精度为 90%,并且相较于单一尺度卷积核,多个大小不同卷积核的卷积神经网络算法更适合于固井变密度图像特征的提取,本文修改了卷积神经网络部分结构,建立的带有 3 个尺寸不同卷积核的MLP-CNN神经网络比单一卷积核的MLP-CNN模型提高了 5%的准确率;同时,本文对比了 7 种网络的时间复杂度和空间复杂度,结果表明,MLP-CNN并联网络能有效避免大量的无效卷积,节省了模型计算成本,提高模型的计算效率.最后,为了测试模型的迁移性,本文使用塔里木油田满深和跃满区块的6万条数据进行了测试,评价准确率达 89.16%,迁移效果良好,模型具有较强的鲁棒性.

关键词: 固井质量评价, 深度学习, 卷积神经网络, 多层感知机, 图像特征提取

CLC Number: 

• TE256 TP18