SIA OpenIR  > 工业控制网络与系统研究室
基于数据驱动的球磨机故障诊断方法
其他题名Data driven fault diagnosis method of ball mill
曲星宇1,2
导师曾鹏
分类号TH17
关键词数据驱动 故障诊断 离群检测 降噪自编码 递归神经网络
索取号TH17/Q86/2017
页数105页
学位专业控制理论与控制工程
学位名称博士
2017-11-30
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门工业控制网络与系统研究室
摘要本文选取流程工业典型代表矿山磨选系统为研究对象,以矿山核心机械设备——球磨机作为研究案例开展研究。首先对球磨机现场采集运行数据进行定义,针对机械设备特性和数据特点提出基于加权随机投影树的t-SNE降维方法,在降维处理后对故障原因及表征进行分类,提出此高维数据样本为严重的不均衡分类情况,为故障诊断算法的选取提供依据。在此基础上,将研究的内容分为高维小数据样本事件、高维大数据样本事件和故障诊断预测性事件。针对球磨机阶段运行小样本事件,即球磨机未在全工况运行时(调试期、试车期、检修期、半投产短时运行阶段)产生数据为小数据样本,提出基于局部权重角度离群算法的球磨机故障诊断方法,采用角度离群算法(ABOD)在高维空间中计算离群度,针对算法高时间复杂度问题,采用FastVOA算法将数据集正交投影于随机超平面上为频矩参数,降低算法的时间复杂度。提出LW-FastVOA算法增加局部权重,降低算法对多聚簇间离群点遗漏率,提高算法诊断准确率。针对球磨机连续运行大样本事件,即球磨机在全工况运行时(全负荷投产运行阶段)产生数据为大数据样本。提出基于DropOut降噪自编码算法的球磨机故障诊断。首先提出采用传统BP神经网络建立故障诊断模型,针对BP神经网络高维输入特征向量辨识能力差的问题,结合Autoencoder网络和Softmax分类器,对构建好的新网络进行训练实现故障预测。由于原始的自动编码器抗干扰性能力较弱,泛化能力不强,并且有一定过拟合的缺陷,提出基于DropOut降噪自编码方法,在特定的工况情况下,诊断的结果明显改善。针对球磨机时序相关性故障预测事件,即球磨机运行过程中,在数据的采集间隔固定的情况下,故障诊断问题就转变成了时间序列预测问题,对故障诊断的过程就带有了预测性。提出采用基于GRU可控门算法的球磨机故障诊断方法,首先根据时间序列预测的特点,提出采用RNN循环神经网络方法,解决故障诊断时间预测问题,因RNN网络反向传播存在梯度弥散问题,提出采用RNN-LSTM长短记忆循环网络方法,LSTM通过控制回传的残差解决指数下降问题,提高网络优化效率。最后提出采用GRU可控门算法优化了可控门结构解决了以上问题。实验结果表明,在评价指标进一步优化的情况下,故障诊断具有了预测性。最后,提出了球磨机故障诊断实验平台的构建与实现过程。首先构建球磨机故障诊断实验平台体系架构的硬件平台,然后对球磨机故障诊断实验平台的软件平台,给出关键开发界面和核心代码,最后给出基于搭建的硬件平台上实现基于数据驱动的球磨机故障诊断系统的方法和现场应用效果。
其他摘要In this paper, mining process is mainly used to be the research object and take the ball mill as an example to carry on the research. At first, the field data of ball mill are defined. According to the characteristics of the mechanical equipment features and data, dimension reduction method of t-SNE based on weighted random projection tree was proposed.The cause and characterization of the fault are classified after the reduction treatment, the point was put forward that the high-dimensional data samples are serious unbalanced classification, which provides the basis for the selection of fault diagnosis algorithm. On this basis, the content of the study is divided into high dimensional small data sample events, high dimensional large data sample events and fault diagnosis predictive events. Running small sample event for ball mill stage, that is, the data generated when the ball mill is not in the full operating conditions (commissioning period, maintenance period, half-production short-term operation phase), the FastVOA based on local weight is proposed in this paper. First angle-based outlier detection (ABOD) is used to measure the outliers of data points in high dimensional space. In order to solve the high time complexity of ABOD algorithm, the FastVOA algorithm is used, which project the data set orthogonally on the hyperplane and estimate the moment to reduce the time complexity of the algorithm. Finally, the LW-FastVOA algorithm is proposed to increase the local weight reduction algorithm of data points to improve the accuracy of the algorithm. Continuous operation of large sample of ball mill, that is, the ball mill is in full operating conditions (full load operation stage), a fault diagnosis model based on Dropout auto encoder is proposed. First the BP neural network is used to build the fault diagnosis model. And aiming at the problem that the BP neural network performs badly when high dimension feature vector inputs, the combination of Autoencoder network and Softmax classifier is proposed, which use the new network to train to predict the failure. Finally, due to the weak anti-jamming performance of the original automatic encoder, the generalization ability is not strong, and there is a certain over-fitting defects, the method based on the DropOut noise reduction auto encoder is proposed, verified that there is a clear improvement .on the evaluation of fault diagnosis. The event of timing correlation failure of ball mill is predicted, that is, when the data collection interval is fixed during the operation of the ball mill, the fault diagnosis problem is transformed into a time series prediction problem. In the case of longitudinal high dimensional data, the horizontal time series prediction is added, the fault diagnosis process is predictive. The method based on GRU controlled gate is proposed in this paper. Firstly, according to the characteristics of time series prediction, RNN cyclic neural network method is proposed to solve the problem of fault diagnosis time prediction, and there is a gradient dispersion problem in RNN network back propagation. Therefore, RNN LSTM is used to improve the efficiency of network optimization by controlling the residuals of feedback. Finally, GRU controllable gate algorithm is proposed to solve the problems above. The experimental results show that the fault diagnosis is predictive in the case of further optimization of the evaluation index. Finally, the construction and implementation process of ball mill fault diagnosis experiment platform is introduced. Firstly, the hardware platform of the f the ball mill fault diagnosis experiment is given. Then, the software platform of the ball mill fault diagnosis experiment is introduced, and the key development interface and core program are given. Finally, the results of field application and method based on the data driven for ball mill fault diagnosis system are shown.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/21280
专题工业控制网络与系统研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
曲星宇. 基于数据驱动的球磨机故障诊断方法[D]. 沈阳. 中国科学院沈阳自动化研究所,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于数据驱动的球磨机故障诊断方法.pdf(3962KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[曲星宇]的文章
百度学术
百度学术中相似的文章
[曲星宇]的文章
必应学术
必应学术中相似的文章
[曲星宇]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。