作者：张俊玲1, 段礼祥2, 王志喜3, 王文权4
作者单位：1. 中国石油大学（北京）机械与储运工程学院，北京 102249;
2. 中国石油大学（北京）安全与海洋工程学院，北京 102249;
3. 川庆钻探工程有限公司井下作业公司，四川 成都 610000;
4. 川庆钻探工程有限公司安全环保质量监督检测研究院，四川 广汉 618300
关键词：压裂泵;故障诊断;KNN;TrAdaBoost;迁移学习;多分类
摘要：
压裂泵工作状态复杂，不易满足传统故障诊断中训练数据多且独立同分布的条件，故障诊断准确率不高。基于TrAdaBoost的迁移学习方法可有效解决上述问题，但模型训练时间较长，且只适用于二分类问题。为此，提出一种基于K-TrAdaBoost迁移学习的压裂泵故障诊断方法。该方法将大量带标签的辅助训练集与少量带标签的目标训练集结合构成足够多的训练集，通过选取K近邻算法（KNN）的最优K值，并引入高斯函数，优化惩罚因子C，计算辅助训练集与目标训练集的相似性，得到辅助训练数据集的初始权重，从而降低TrAdaBoost迭代次数，减少训练时间。迭代结束后，在模型内部引入多分类器，改变模型输出机制，实现多种故障类型的诊断。实验结果表明：所提方法可解决传统诊断方法训练数据集不足且无法独立同分布的问题，降低TrAdaBoost模型的训练时间，实现K-TrAdaBoost多种故障类型的诊断，提高压裂泵故障诊断的准确率。
Research on fault diagnosis method of fracturing pump based on K-TrAdaBoost transfer learning
ZHANG Junling1, DUAN Lixiang2, WANG Zhixi3, WANG Wenquan4
1. College of Mechanical and Transportation Engineering, China University of Petroleum(Beijing), Beijing 102249, China;
2. College of Safety and Ocean Engineering, China University of Petroleum(Beijing), Beijing 102249, China;
3. CCDC Down Hole Service Company, Chengdu 610000, China;
4. HSE Quality Surveillance & Inspection Research Institute, CNPC Chuanqing Drilling Engineering Company Limited, Guanghan 618300, China
Abstract: The working condition of fracturing pump is complicated. It is not easy to meet the conditions of large and independent and identically distributed training data in traditional fault diagnosis, resulting in low fault diagnosis accuracy. The transfer learning method based on TrAdaBoost solves the above problems, but the model takes a long time to train and is only suitable for binary classification problems. This paper presents a fracturing pump fault diagnosis method based on K-Tradaboost transfer learning method. A large number of auxiliary training sets with labels are combined with a small number of target training sets with labels to form enough training sets. The optimal K value of the K-Nearest Neighbor(KNN) is selected. The penalty factor C is optimized by introducing Gaussian function. The similarity between the auxiliary training set and the target training set is calculated, and the initial weight of the auxiliary training data set is obtained. The iteration times of TrAdaBoost and the training time are reduced. After the iteration, multiple classifiers are introduced into the model to change the model output mechanism and realize multiple fault type diagnosis. The experimental results show that the proposed method solves the traditional diagnostic methods problem. Reduces the training time of TrAdaBoost model. K-TrAdaBoost realizes the diagnosis of various fault types and improves the accuracy of fracturing pump fault diagnosis.
Keywords: fracturing pump;fault diagnosis;KNN;TrAdaBoost;transfer learning;multi-classification
2021, 47(10):7-11,40 收稿日期: 2021-03-09;收到修改稿日期: 2021-04-25
基金项目: 国家自然科学基金资助项目（51674277）；中国石油集团公司项目（2019-F30）
作者简介: 张俊玲（1992-），女，辽宁大连市人，博士研究生，研究方向为机械状态监测和智能诊断
参考文献
[1] 刘曰武, 高大鹏, 李奇, 等. 页岩气开采中的若干力学前沿问题[J]. 力学进展, 2019, 49(1): 1-236
[2] 闫贺. 往复式压缩机在线监测系统开发与应用[D]. 北京: 北京化工大学, 2010.
[3] 张俊玲, 陈志刚, 杜小磊, 等. 基于局部均值分解样本熵与支持向量机的2000型压裂车液力端故障诊断[J]. 计算机应用, 2018, 38(S2): 60-62, 67
[4] 杜小磊, 陈志刚, 许旭, 等. 基于小波, WAE和LSTM的压裂车故障诊断[J]. 石油机械, 2019, 47(10): 88-93, 106
[5] 但长林, 李三雁, 张彬. 基于样本熵和SVM的滚动轴承故障诊断方法研究[J]. 中国测试, 2020, 46(11): 37-42
[6] DAI W, YANG Q, XUE G R, et al. Boosting for transfer learning[C]//In Proceedings of the Twenty-Fourth International Conference on Machine Learning, 2007.
[7] 王昕, 赵飞, 蒋佐富, 等. 迁移学习和卷积神经网络电力设备图像识别方法[J]. 中国测试, 2020, 46(5): 108-113
[8] FREUND Y, SCHAPIRE R E. A decision-theoretic generalization of on-line learning and an application to boosting[J]. Journal of Computer and System Sciences, 1999, 55: 119-139
[9] CHEN Y Y, WANG Z B. Cross components calibration transfer of NIR spectroscopy model through PCA and weighted ELM-based TrAdaBoost algorithm[J]. Chemometrics and Intelligent Laboratory Systems, 2019, 192: 103824
[10] 徐桂芝, 林放, 宫铭鸿, 等. 基于TrAdaBoost的跨脑辨识P300电位研究[J]. 生物医学工程学杂志, 2019, 36(4): 531-540
[11] 刘万军, 李天慧, 曲海成. 基于Fisher准则和TrAdaboost的高光谱相似样本分类算法[J]. 国土资源遥感, 2018, 30(4): 41-48
[12] TANG D J, YANG X H, WANG X S. Improving the transferability of the crash prediction model using the TrAdaBoost. R2 algorithm[J]. Accident Analysis and Prevention, 2020: 105551
[13] 陈仁祥, 陈思杨, 杨黎霞, 等. 改进TrAdaBoost多分类算法的滚动轴承故障诊断[J]. 振动与冲击, 2019, 38(15): 36-41, 48
[14] 王楠. 基于KNN的改进算法研究及其在图像去噪的应用[D]. 桂林: 广西师范学院, 2016.
[15] 张彼德, 梅婷, 王涛. 电力变压器故障诊断的k值自适应加权KNN算法研究[J]. 湖北电力, 2020, 44(2): 6-12
[16] 黄娟娟. 基于KNN的文本分类特征选择与分类算法的研究与改进[D]. 厦门: 厦门大学, 2014.
