基于改进NPE算法的间歇过程故障诊断研究

发布时间：2018-01-08 11:05

本文关键词：基于改进NPE算法的间歇过程故障诊断研究　出处：《兰州理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着工业过程朝着智能化、大规模和集成化方向发展,生产过程变的越来越复杂。复杂系统容易受到外部环境的干扰和自身设备的老化而发生各种故障,要实现工业过程自动化来提高生产效率,就要求控制系统处于稳定状态,需要对生产过程进行实时监控。间歇生产方式由于自身的优点在实际应用中越来越广,间歇过程的产品是批次输出,一旦生产过程的某个时刻或某个变量发生故障,而得不到及时解决,都会使整个批次产品达不到要求,从而带来巨大的经济损失,因此间歇过程的过程监控和故障诊断就更为重要。对于复杂的间歇过程,我们很难获得准确的过程模型,基于数据驱动的多元统计方法,不需要得到过程的精确模型,只需过程的历史和在线数据就能实现对过程的监控和故障诊断,随着传感器技术的发展和广泛应用于工业过程,过程的历史和时时数据的获取变得相对容易,为基于数据驱动的故障诊断的广泛应用提供可能。如何在海量的过程数据中充分提取对过程监控有用的特征信息是过程监控与故障诊断的关键,相比与主元分析(PCA)算法只关注过程的全局特征信息,邻域保持嵌入(NPE)算法从数据的局部特征信息出发来挖过出更多的间歇过程的细节特征信息。本文主要针对间歇生产过程在监控和故障诊断过程中存在的问题,对邻域保持嵌入算法(NPE)进行了改进,其主要的研究内容如下:1.针对间歇过程的强非线性,传统方法只是对数据的协方差矩阵进行分解,忽略数据高阶统计量信息,导致无法充分提取非线性过程的有效信息造成诊断效果不佳的问题,结合核算法和高维统计量信息提出了一种基于统计量的多向核邻域保持嵌入(SPA-MKNPE)算法。该算法首先引入统计量模式分析(SPA)方法将样本数据投影到统计量样本空间中,可以更充分地提取非线性数据的高阶统计量信息;然后在统计量空间中通过核函数将统计量样本映射到高维核空间,用以解决数据的非线性;最后在高维核空间中应用邻域保持嵌入算法充分提取数据的局部结构来对间歇过程进行监控,检测到过程故障后用变量贡献图法来诊断出故障变量。通过青霉素发酵过程来验证SPA-MKNPE算法对强非线性的间歇过程故障诊断的有效性。2.针对间歇生产过程数据很难满足单一的高斯分布,通常既包含高斯成分又包含非高斯成分的混合多分布,传统方法在特征提取时不能兼顾数据的全局和局部特征的问题,提出了MGNPE-LICA的故障诊断算法。首先用D检验法将原始数据分成高斯空间和非高斯空间,对于高斯空间用MGNPE算法在充分提取数据的局部结构特征时兼顾数据的全局特征;对于非高斯空间用MLICA算法在解决非高斯问题的同时更好地保持数据的全局和局部信息。再将两个空间的监控指标合成一个联合监控指标对过程进行监控。通过青霉素发酵过程验证了MGNPE-LICA算法的在解决间歇过程数据多分布的有效性。3.针对间歇过程的三维数据在展开成二维的过程中必然会导致数据内在结构破坏和全局与局部特征信息在提取过程中无法兼顾的问题,提出了张量全局-局部邻域保持嵌入算法(TGNPE),首先用张量分解的方法直接对三维数据进行建模,而不对数据进行展开,有效地保存了数据的内部结构,再用邻域保持嵌入算法充分提取数据局部结构信息的同时兼顾数据的全局信息,实现对数据特征信息更加充分地提取,TGNPE算法检测到故障后用贡献图法诊断出故障变量。通过青霉素发酵过程验证了TGNPE算法更利于间歇过程数据信息的提取,诊断精度更高。4.针对间歇过程在时序和空间上的动态特性,提出了基于张量分解的动态邻域保持嵌入算法(TDNPE)。首先用张量分解的方法把间歇过程数据看成二阶张量,在张量空间对数据建模避免数据展开的向量化过程,导致数据内部结构被破坏;然后用动态邻域保持嵌入算法在张量空间提取过程特征信息,同时考虑空间和时序上的局部特征,有效解决了过程的动态特性。通过青霉素发酵过程的仿真实验验证了TDNPE算法的有效性。
[Abstract]:With the development of industrial process toward intelligent, large-scale and integrated development, the production process becomes more and more complex. The complex system is vulnerable to the interference of the external environment and their own aging equipment and all kinds of faults, to realize the automation of industrial process to improve production efficiency, requires the control system is stable, the need for real-time monitoring of the the production process of batch process. Because of its advantages of more and more widely in practical applications, batch process products are batch output, once a time of the production process or a variable fault, but it is not solved in time, will make the whole batch of products meet the requirements, so as to bring huge economic losses therefore, process monitoring and fault diagnosis of batch process is more important. For the batch process is complicated, it is difficult to obtain accurate process model, based on data driven Multivariate statistical methods, does not need the precise model of the process, only the history and online data can be achieved on the process monitoring and fault diagnosis, along with the development of sensor technology and widely used in the industrial process, the process of acquiring data and history always become easy, may provide for the wide application of data driven fault diagnosis based on the data in the process. How to fully extract the feature information useful for monitoring process is the key process monitoring and fault diagnosis, compared with principal component analysis (PCA) algorithms only focus on the global features of the process of information, neighborhood preserving embedding (NPE) algorithm from the local feature information of data dig out minutiae information for batch process more. This article mainly aims at the existing in the batch process monitoring and fault diagnosis problems in the process of neighborhood preserving embedding algorithm (NPE) Has been improved, and its main research contents are as follows: 1. for strong nonlinear batch process, the traditional approach is to decompose the covariance matrix of the data, ignoring the data information of high order statistics, to fully extract effective information of nonlinear process caused by the problem of poor diagnostic effect, combined with the accounting method and the high dimensional statistics information is proposed a multi kernel neighborhood preserving embedding (SPA-MKNPE) algorithm based on statistics. Firstly, statistic pattern analysis (SPA) method of sample data will be projected to the statistics in the sample space, you can more fully extract the information of high order statistics and nonlinear data; statistics in space by kernel function will be mapped into high statistics dimensional kernel space is used to solve the nonlinear data; finally the application of neighborhood in the high dimensional kernel space is fully extracted according to local node number preserving embedding algorithm Structure to monitor batch process, detect process after fault by using variable contribution plot method to diagnose the fault variables. Through the penicillin fermentation process to validate the SPA-MKNPE algorithm for nonlinear batch process fault diagnosis of the effectiveness of.2. for batch process data is very difficult with single foot Gauss distribution, mixing usually includes Gauss component it contains non Gauss component distribution, traditional methods can not take into account the data of global and local features in feature extraction, we propose a fault diagnosis algorithm of MGNPE-LICA. Firstly, D test was used for the original data into Gauss space and Gauss space, the Gauss space MGNPE algorithm is used to extract the local structure in full feature data when considering the global features of the data; for non Gauss space with MLICA algorithm in solving the problem of non Gauss and better maintain the global and local data The Ministry of information. Then the two space monitoring index of the synthesis of a joint monitoring indicators to monitor the process of penicillin fermentation process. Through the validation of the MGNPE-LICA algorithm in solving the distributed batch process data the validity of.3. for 3D data of batch process in the process of two-dimensional in internal structure will inevitably lead to destruction and global data with the local feature information cannot be taken into account in the process of extraction problem, we propose a tensor global local neighborhood preserving embedding algorithm (TGNPE). The method first uses tensor decomposition model directly on 3D data, data without expansion, effectively preserve the internal structure of the data, and then the neighborhood preserving embedding algorithm fully extracted data locally the structure information of both global information data and realize more fully to extract the data feature information, TGNPE algorithm to detect faults after use The contribution of graph fault diagnosis variable. Through the penicillin fermentation process to verify the TGNPE extraction method is more conducive to the data information of batch process, higher diagnostic accuracy of.4. for dynamic characteristics of the batch process in time and space, and proposes a dynamic neighborhood tensor decomposition keep embedding (TDNPE) algorithm based on tensor decomposition method first uses. The batch process data as two order tensor in the tensor space data to the quantification process of data modeling to avoid, the internal structure of the data is destroyed; and then use dynamic neighborhood preserving embedding algorithm in tensor Space Extraction of feature information, and the local feature space and time, effectively solves the dynamic characteristics of the process. Through the simulation experiment of penicillin fermentation process to verify the effectiveness of the TDNPE algorithm.

【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP277

【相似文献】