【作者】 韩冰；

【导师】 佟庆彬；

【作者基本信息】 北京交通大学 ， 电气工程（专业学位）， 2021， 硕士

【摘要】 伴随着现代科学技术的超高速发展和机械设备复杂化和智能化的逐步加深,机械设备的运行工况和工作环境越来越复杂。当机械设备出现故障时,依靠专家对采集到的信号进行分析并判断出故障类型费时费力,且在错综复杂的工况下缺乏通用性,已经不能满足大型机械设备故障诊断的要求。本文针对该问题,以旋转机械中最常用的零部件轴承作为研究对象,根据深度学习搭建复数卷积神经网络,利用该网络对轴承信号进行特征提取和分类。具体内容如下:本文首先设计2层卷积层的一维卷积神经网络,采用全局均值池化层(Global Average Pooling Layer)对网络结构进行改进,通过对比不同优化算法和学习率的实验结果得到了模型最优训练参数。对振动信号进行诊断的实验结果证明,直接使用振动信号作为输入可以取得较好的预测性能,但仍存在优化的空间。针对二维卷积神经网络对图像分类问题具有较好的性能,本文设计8层二维卷积神经网络,通过滑窗取样将一维时域信号制作成图片,送入二维卷积神经网络进行实验验证。实验结果证明,模型在西储大学轴承数据集上的识别率可以达到85%左右,但由于图片数据集数据量较大,导致网络训练耗时严重。针对传统卷积神经网络仅能处理单一类型的数据,利用傅里叶变换可以将原始实数数据转换成同时携带时域和频域信息的复数数据,设计了8层一维复数卷积神经网络,该网络接收复数数据作为输入。由于同时提取轴承信号中的时域和频域的故障特征,经过实验验证,该模型可以达到100%的准确率。利用PCA对故障特征进行降维,对复数卷积神经网络分类过程进行可视化。针对实际应用中对轴承故障标签信息的不足,提出了特征迁移和卷积神经网络相结合的方法。首先通过卷积神经网络对轴承原始时域数据进行特征提取,再利用TCA算法和JDA算法对特征提取后的数据进行特征迁移,利用KNN算法根据源域数据特征预测目标域数据种类,将预测值与实际值对比得出预测准确率。针对实际应用中对轴承故障测量信息的不足,设计基于模型迁移的复数卷积神经网络。先利用源域数据对网络进行训练,然后将网络前2层卷积层进行迁移,利用目标域数据对迁移后的网络进行微调和分类测试。特征迁移和模型迁移的实验结果均证明,变负载情况时源域和目标域数据之间特征差异较小,可迁移性较好,而变故障直径和变采样位置时源域和目标域数据之间特征差异较大,可迁移性较差。更多还原

【Abstract】 With the rapid development of modern science and technology and the gradual deepening of the complexity and intelligence of mechanical equipment,the operating conditions and working environment of mechanical equipment are becoming more and more complex.When a mechanical device fails,it takes a lot of time and effort to analyze the collected signals and judge the type of failure by experts,and the lack of universality in complex working conditions can no longer meet the requirements of fault diagnosis for large mechanical devices.In order to solve this problem,a complex-value convolution neural network is built based on deep learning,which is used to extract and classify the bearing signals.The details are as follows:In this thesis,a one-dimensional convolutional neural network with two convolution layers is designed,and the global average pooling layer is used to improve the network structure.By comparing the experimental results of different optimization algorithms and learning rates,the optimal training parameters of the model are obtained.The experimental results of vibration signal diagnosis show that the direct use of vibration signal as input can achieve better prediction performance,but there is still room for optimization.In view of the good performance of two-dimensional convolutional neural network for image classification,this thesis designs eight layers of two-dimensional convolutional neural network,and makes one-dimensional time-domain signal into a picture by sliding window sampling,and then sends it to two-dimensional convolutional neural network for experimental verification.The experimental results show that the recognition rate of the model can reach 85% on the bearing data set of CWRU,but the network training time is serious due to the large amount of picture data set.For the traditional convolution network can only process a single type of data,the original real value data can be converted to complex value data with both time and frequency domain information by using Fourier transform.An 8-layer one-dimensional complex-value convolution network is designed,which receives complex value data as input.Because the fault features in both time and frequency domains of bearing signals are extracted at the same time,the model can achieve 100% accuracy.PCA is used to reduce the dimension of fault features,and the complex-value convolution neural network classification process is visualized.In view of the insufficiency of bearing fault label information in practical application,a method combining feature migration and convolution neural network is presented.First,the original bearing time domain data is extracted by convolution neural network,then the extracted data is migrated by TCA algorithm and JDA algorithm,and then the KNN algorithm is used to predict the target domain data type according to the source domain data characteristics,and the prediction accuracy is obtained by comparing the predicted value with the actual value.To overcome the shortage of bearing fault measurement information in practical application,a complex-value convolution neural network based on model migration is designed.The source domain data is used to train the network,then the first two convolution layers of the network are migrated,and the migrated network is fine-tuned and classified by the target domain data.The experimental results of both feature migration and model migration show that the difference of feature between source domain and target domain data is small and the migrability is good under variable load conditions,while the difference of feature between source domain and target domain data is large and the migration is poor under variable fault diameter and sampling location.更多还原