基于组序列多分支CNN-LSTM的风机轴承和齿轮箱故障诊断研究  

Fault diagnosis of wind turbine bearing and gearbox based on group-sequence and multi-branch CNN-LSTM model

作　　者：许启发[1,2] 程启亮 蒋翠侠[1,2] 汪湘湘 XU Qi-fa;CHENG Qi-liang;JIANG Cui-xia;WANG Xiang-xiang(School of Management,Hefei University of Technology,Hefei 230009,China;Key Laboratory of Process Optimization and Intelligent Decision-Making,Ministry of Education,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]合肥工业大学管理学院,安徽合肥230009 [2]合肥工业大学过程优化与智能决策教育部重点实验室,安徽合肥230009

出　　处：《机电工程》2022年第8期1050-1060,共11页Journal of Mechanical & Electrical Engineering

基　　金：国家自然科学基金面上项目(72171070);安徽省重点研究与开发计划面上攻关项目(202004a05020020)。

摘　　要：现实工业环境中,单点数据的采集时间通常为几秒甚至更短,现有基于单点数据的风机轴承和齿轮箱故障智能诊断算法难以取得满意结果,为此,提出了一种具有注意力机制的组序列多分支卷积神经网络长短期记忆网络(CNN-LSTM)模型,即GSMBCLAM模型。首先,将具有相同采样间隔的连续多点数据合并成组序列样本,并将组序列波形和频谱同时输入2个不同的一维卷积神经网络(1D-CNN)中,进行了自适应特征提取;其次,采用注意力机制,将提取出来的特征与人工提取的特征进行了特征加权,并将加权后的特征进行了融合,然后将其输入到LSTM中;再次,考虑到故障分类中,不同类别误分类代价不同的问题,采用焦点损失函数代替了传统的交叉熵损失函数;最后,基于SoftMax分类器输出了诊断结果,通过一个包含54000个原始波形、频谱和人工提取特征,区分5类不同的轴承和齿轮故障和1类正常的真实数据集进行了对比实验。研究结果表明:GSMBCLAM方法在准确率、精确率、召回率、F1分数上分别达到了98.40%、98.46%、98.63%、98.30%;其效果优于只基于单点数据或单分支的模型,各项指标对比于其他深度学习竞争模型具有明显优势;焦点损失函数的引入解决了故障诊断中误分类代价不同的问题。In real industrial environments,the acquisition time of single-point data was usually several seconds or even shorter.Based on single-point data,the existing intelligent diagnosis algorithms for wind turbines were difficult to obtain satisfactory results.To this end,a group-sequence multi-branch convolution neural networks-long short term memory(CNN-LSTM)model(Gsmbclam model)with attention mechanism was proposed.Firstly,continuous multi-point data with the same sampling interval were formed into a group sequence sample,and group sequence waveforms and spectrums were input into two different one-dimensional convolution neural network(1D-CNN)s for adaptive feature extraction.Secondly,the extracted features and manually features were weighted through the attention mechanism,and the weighted features were fused and input into LSTM.Then,considering the different misclassification costs of different categories in fault classification,focal loss function was used to replace the traditional cross entropy loss function.Finally,the diagnosis results were output by the SoftMax classifier.Comparison tests are performed using a real data set containing 54,000 raw waveforms,spectra,and manually features,which comprise five different types of bearing and gear faults and one type of normal data.The results show that the accuracy,precision,recall and F1 score of the proposed method reach 98.40%,98.46%,98.63%and 98.30%,respectively.Its performance is better than the model based on single point data or single branch.Each index has obvious advantages compared with other competing deep learning models.The introduction of focus loss function solves the problem of different classification costs in fault diagnosis.

关 键 词：齿轮箱故障诊断 轴承故障诊断 组序列 多分支 卷积神经网络长短期记忆网络 焦点损失函数 

分 类 号：TH132.4[机械工程—机械制造及自动化] TH133.3[电气工程—电机] TM315