摘要
现有钻头磨损评估方法中,存在人工特征提取过程可能无法完全提取正确分类所需的信号动态特征,及需要对各个统计量进行大量计算等问题。为此,提出了一种新的基于改进卷积神经网络支持向量机(CNN-SVM)的钻头磨损程度评估算法。该算法将采集的近钻头原始振动数据导入CNN-Softmax模型,通过训练好的CNN模型从近钻头数据中提取主要的特征参数,将提取的稀疏特征向量输入SVM并进行故障分类,利用遗传算法实现SVM参数的优化选择,最后应用t分布随机邻域法近邻嵌入,使其故障特征学习过程可视化,以评估其特征提取能力。采用该算法对钻头磨损的现场试验数据进行了分析。分析结果表明:基于改进CNN-SVM的井下钻头磨损状态评估算法准确率高达98.33%。所得结论可为实现钻头磨损状态的进一步监测提供理论支撑。
The existing bit wear evaluation methods have the problems such as the inability of fully extracting the dynamic feature of signals needed by correct classification in the manual feature extraction process and the need for massive calculation of various statistics.Therefore,a new bit wear rate evaluation algorithm based on improved CNN-SVM was proposed in the paper.This algorithm imported the collected near-bit raw vibration data into the CNN-Softmax model,extracted the main feature parameters from the near-bit data through the trained CNN model,input the extracted sparse feature vectors into SVM for fault classification,used genetic algorithm to achieve optimization selection of SVM parameters,and finally used t-distribution stochastic neighborhood method to conduct near neighbor embedding to visualize the fault feature learning process and evaluate its feature extraction ability.In addition,this algorithm was used to analyze the field test data of bit wear.The analysis results show that the accuracy of the downhole bit wear status evaluation algorithm based on CNN-SVM is as high as 98.33%.The conclusions provide theoretical support for realizing further monitoring of bit wear status.
作者
李玉梅
邓杨林
李基伟
李乾
杨磊
于丽维
Li Yumei;Deng Yanglin;Li Jiwei;Li Qian;Yang Lei;Yu Liwei(Beijing Key Laboratory of High Dynamic Navigation Technology,Beijing Information Science&Technology University;MOE Key Laboratory of Modern Measurement&Control Technology,Beijing Information Science&Technology University;Sinopec Offshore Oilfield Services Company;Engineering Technology Research Institute of Xinjiang Oilfield Company)
出处
《石油机械》
北大核心
2024年第6期12-19,共8页
China Petroleum Machinery
基金
国家自然科学基金青年科学基金项目“干热岩储层双重介质射孔簇内复杂多裂缝起裂及扩展机理研究”(52104001)
国家自然科学基金面上项目“底部钻具高频扭转振动响应机理及识别方法研究”(52274003)。
关键词
钻头磨损状态评估
卷积神经网络
支持向量机
特征提取可视化
平均池化采样
bit wear status evaluation
convolutional neural networks(CNN)
support vector machine(SVM)
feature extraction visualization
mean pooling sampling
