基于改进深度森林的滚动轴承剩余寿命预测方法  

Prediction Method of Remaining Useful Life of Rolling Bearings Based on Improved GcForest

作　　者：王玉静[1] 王诗达 康守强[1] 王庆岩[1] V.I.MIKULOVICH WANG Yujing;WANG Shida;KANG Shouqiang;WANG Qingyan;V.I.MIKULOVICH(School of Elctrical and Electronic Engineering,Harbin University of Science and Technology,Harbin 150080,Heilongjiang Province,China;Belarusian State University,Minsk 220030,Belarus)

机构地区：[1]哈尔滨理工大学电气与电子工程学院,黑龙江省哈尔滨市150080 [2]白俄罗斯国立大学,白俄罗斯明斯克220030

出　　处：《中国电机工程学报》2020年第15期5032-5042,共11页Proceedings of the CSEE

基　　金：国家自然科学基金项目(51805120);黑龙江省自然科学基金项目(LH2019E058);黑龙江省本科高校青年创新人才培养计划(UNPYSCT-2017091)。

摘　　要：针对现有人工智能方法在滚动轴承剩余寿命预测中存在精度差、运算效率低的问题,提出一种基于深层迭代特征(deep iterative features,DIF)级联CatBoost(cascade catboost,CasCatBoost)的滚动轴承剩余寿命预测新方法。该方法是一种改进的新型深度森林算法,首先对由快速傅里叶变换得到的滚动轴承频域信号进行迭代计算,得到迭代特征。为了减小内存的消耗,将深度森林中的多粒度扫描结构替换为卷积神经网络,提取迭代特征的深层特征,并构建性能退化特征集。然后对可实现GPU并行加速的单一CatBoost模型进行集成,引入决定系数R2构建CasCatBoost结构以提高模型的表征能力,选取模型最后一个级联层的平均寿命百分比p表示输出。最后运用一次函数对p进行拟合,预测出轴承的剩余寿命。利用PHM2012数据库对滚动轴承剩余寿命进行预测,所提方法的预测平均误差为10.57%、平均得分为0.426。For the problems that the existing artificial intelligence methods have poor precision and low computational efficiency in the prediction of remaining useful life(RUL)of rolling bearings,a new method of predicting RUL of rolling bearings was proposed based on deep iterative feature(DIF)cascaded CatBoost(CasCatBoost).This method is an improved new multi-grained cascade forest(gcForest)algorithm.Firstly,the frequency domain signal of rolling bearing was obtained using fast Fourier transform,and the iterative feature(IF)obtained by iterative operation.To reduce the memory consumption,the multi-grained scanning structure in the gcForest was replaced by convolutional neural networks(CNN),the deep feature DIF of IF was extracted,and the performance degradation feature set was constructed.Then,a single CatBoost model that can realize GPU parallel acceleration was integrated,and the determination coefficient R2 was introduced to construct the CasCatBoost structure for improving the representation ability of the model.The average life percentage p of the last cascade layer of the model was selected as the output.Finally,linear function was used to fit p and the RUL of rolling bearing was predicted.PHM2012 database was used for predicting the RUL of rolling bearing,and the prediction average error of the proposed method is 10.57%,the average score is 0.426.

关 键 词：滚动轴承 卷积神经网络 深层迭代特征 深度森林 剩余寿命预测 

分 类 号：TN911[电子电信—通信与信息系统] TH165[电子电信—信息与通信工程]