CN111598220B - Gas turbine fault prediction method based on correlation analysis - Google Patents

Gas turbine fault prediction method based on correlation analysis Download PDF

Info

Publication number
CN111598220B
CN111598220B CN202010404291.6A CN202010404291A CN111598220B CN 111598220 B CN111598220 B CN 111598220B CN 202010404291 A CN202010404291 A CN 202010404291A CN 111598220 B CN111598220 B CN 111598220B
Authority
CN
China
Prior art keywords
monitoring node
node vector
monitoring
vector
value
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010404291.6A
Other languages
Chinese (zh)
Other versions
CN111598220A (en
Inventor
杨静
方宝富
沈安波
樊高金
江刘峰
朱尤杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei University of Technology
Original Assignee
Hefei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hefei University of Technology filed Critical Hefei University of Technology
Priority to CN202010404291.6A priority Critical patent/CN111598220B/en
Publication of CN111598220A publication Critical patent/CN111598220A/en
Application granted granted Critical
Publication of CN111598220B publication Critical patent/CN111598220B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M15/00—Testing of engines
    • G01M15/14—Testing gas-turbine engines or jet-propulsion engines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Chemical & Material Sciences (AREA)
  • Combustion & Propulsion (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

本发明公开了一种基于相关性分析的燃气轮机故障预测方法，其步骤包括：1、以流的方式逐个读入待处理的监测节点向量；2、对每个当前读入的监测节点向量，与系统已经进行读入的监测节点向量，进行相关性分析；3、对已经选择的相关监测节点进行冗余分析；4、对于新加入的监测节点进行定向，确定与其他监测节点间的因果关系，重复步骤1‑4，直至读入的监测节点向量的数量超过极限值，最终获得相应的监测系统因果结构图，并用于训练故障预测模型；从而得到故障预测模型，以实现对故障进行更加准确的预测。本发明能获得更加精准的故障预测模型，从而能对故障进行更加准确的预测。The invention discloses a gas turbine fault prediction method based on correlation analysis, the steps of which include: 1. Read in the monitoring node vectors to be processed one by one in a flow mode; 2. For each currently read monitoring node vector, and The system has already read in the monitoring node vector and performs correlation analysis; 3. Perform redundancy analysis on the selected related monitoring nodes; 4. Orientate the newly added monitoring nodes to determine the causal relationship with other monitoring nodes. Repeat steps 1-4 until the number of monitoring node vectors read in exceeds the limit value, and finally obtain the corresponding causal structure diagram of the monitoring system, and use it to train the fault prediction model; thus obtain the fault prediction model to achieve more accurate fault prediction predict. The present invention can obtain a more accurate fault prediction model, so that faults can be predicted more accurately.

Description

一种基于相关性分析的燃气轮机故障预测方法A Gas Turbine Fault Prediction Method Based on Correlation Analysis

技术领域technical field

本发明属于数据挖掘领域，具体地说是一种基于相关性分析的燃气轮机故障预测方法。The invention belongs to the field of data mining, in particular to a gas turbine fault prediction method based on correlation analysis.

背景技术Background technique

目前国内燃气轮机状态监测和故障诊断研究现状最近有很大进步,但是技术还相对比较落后,应用成果较少。随着大数据技术的兴起，如何将大数据相关技术应用于燃气轮机状态监测和故障诊断是一个值得研究的课题。燃气轮机机组在运行时不断地产生大量的监测数据，基于这些海量的运行监测数据，开展燃气轮机机组状态分析、性能监测和故障智能诊断预测研究，具有非常重要的现实意义。通过数据建模，可以对燃气轮机机组状态进行实时的健康评估，预测状态趋势，在没有发生重大故障前提前预警，可以早期发现燃气轮机故障，从而避免经济损失、提供维修建议、有助于燃机的安全可靠地运行。然而，这些数据的分布往往是任意的，彼此之间的关系往往具有非线性的特点，对于这种非线性数据的研究是具有一定的挑战。这些运行数据构成一个复杂的网络系统，辨识该复杂系统的网络节点间的联系，有助于燃气轮机的状态监测和故障预测。At present, the research status of domestic gas turbine condition monitoring and fault diagnosis has made great progress recently, but the technology is still relatively backward, and the application results are few. With the rise of big data technology, how to apply big data related technology to gas turbine condition monitoring and fault diagnosis is a topic worth studying. Gas turbine units continuously generate a large amount of monitoring data during operation. Based on these massive operating monitoring data, it is of great practical significance to carry out research on status analysis, performance monitoring, and fault intelligent diagnosis and prediction of gas turbine units. Through data modeling, real-time health assessment of gas turbine unit status can be performed, status trends can be predicted, early warnings can be made before major failures occur, and gas turbine failures can be detected early, thereby avoiding economic losses, providing maintenance suggestions, and contributing to the development of gas turbines. Operate safely and reliably. However, the distribution of these data is often arbitrary, and the relationship between each other is often nonlinear, so it is a certain challenge to study this nonlinear data. These operating data constitute a complex network system, and identifying the connections between the network nodes of this complex system is helpful for the state monitoring and failure prediction of gas turbines.

而描述复杂网络间关系的杰出模型就是由美国加州大学的Judea Pearl提出的基于概率论和图论的贝叶斯网络模型，并凭杰出的贡献获得2011年度图灵奖。Hoyer等对于贝叶斯网络因果模型进行更进一步扩展，提出了加性噪声模型，该模型可以建模非高斯非线性的数据。而燃气轮机机组的运行数据恰恰也是非高斯非线性的。所以，基于加性噪声模型对燃气轮机机组的运行数据进行分析是非常有意义的研究方向。至于加性噪声模型的结构学习，Hoyer等基于非线性回归和基于HSIC标准提出了识别因果结构的方法，Mooij等提出了基于HSIC回归的算法，Zhang等提出了两阶段的算法，Tillman等人提出了kPC算法，Yamada等提出了最小二乘独立性回归的方法，Mooij等提出了基于最大后验的方法，Zhang等提出一个基于核的条件独立测试，Peters等提出了基于后续独立测试的回归方法，Zhang等提出一个基于回归的条件独立型测试的方法等，Nowzohour等基于惩罚性的似然的方法等等。The outstanding model describing the relationship between complex networks is the Bayesian network model based on probability theory and graph theory proposed by Judea Pearl of the University of California, USA, and won the 2011 Turing Award for his outstanding contributions. Hoyer et al. further extended the Bayesian network causal model and proposed an additive noise model, which can model non-Gaussian nonlinear data. However, the operating data of the gas turbine unit is also non-Gaussian nonlinear. Therefore, it is a very meaningful research direction to analyze the operating data of the gas turbine unit based on the additive noise model. As for the structure learning of additive noise models, Hoyer et al. proposed a method for identifying causal structures based on nonlinear regression and HSIC criteria, Mooij et al. proposed an algorithm based on HSIC regression, Zhang et al. proposed a two-stage algorithm, Tillman et al. proposed the kPC algorithm, Yamada et al. proposed the method of least squares independence regression, Mooij et al. proposed a method based on maximum a posteriori, Zhang et al. proposed a kernel-based conditional independent test, Peters et al. proposed a regression method based on subsequent independent tests , Zhang et al. proposed a regression-based conditional independence test method, etc., Nowzohour et al. based on the penalty-likelihood method and so on.

燃气轮机机组数据还具有高维特性，处理高维数据的一种方法就是降维，通常采用主成分分析法、独立成分分析等方法，而这些方法都要事先知道所有数据维的信息并一次载入内存，但有时燃气轮机机组数据维数巨大无法一次载入内存。并且可能不断出现新的测点数据，造成数据的特征空间是动态的、未知的。处理动态高维的数据近年来兴起一种基于流特征的数据分析方法，目前是数据挖掘领域一个新兴的研究方向，可以有效的处理高维大数据。Gas turbine unit data also has high-dimensional characteristics. One method to deal with high-dimensional data is dimensionality reduction, usually using methods such as principal component analysis and independent component analysis, and these methods must know all data dimensions in advance and load them at once. memory, but sometimes the data dimension of the gas turbine unit is too large to be loaded into the memory at one time. And new measuring point data may appear continuously, resulting in a dynamic and unknown feature space of the data. Handling dynamic high-dimensional data In recent years, a data analysis method based on flow features has emerged. It is currently an emerging research direction in the field of data mining, which can effectively process high-dimensional big data.

目前这些方法的主要局限包括：The main limitations of these current approaches include:

(1)上述多数算法的计算复杂度比较大，满足不了燃气轮机机组运行数据的在线实时学习；(1) The computational complexity of most of the above algorithms is relatively large, which cannot meet the online real-time learning of the operating data of the gas turbine unit;

(2)燃气轮机机组数据维数巨大无法一次载入内存，并且可能不断出现新的测点数据，现存的方法不能处理这种情形。(2) The data dimension of the gas turbine unit is too large to be loaded into the memory at one time, and new measurement point data may appear continuously, and the existing methods cannot handle this situation.

发明内容Contents of the invention

本发明为克服现有技术存在的不足之处，提出了一种基于相关性分析的燃气轮机故障预测方法，以期能获得更加精准的故障预测模型，从而能对故障进行更加准确的预测。In order to overcome the deficiencies in the prior art, the present invention proposes a gas turbine fault prediction method based on correlation analysis, in order to obtain a more accurate fault prediction model, thereby enabling more accurate fault prediction.

本发明为解决技术问题采用如下技术方案：The present invention adopts following technical scheme for solving technical problems:

本发明一种基于相关性分析的燃气轮机故障预测方法，是应用于燃气轮机系统中，并每隔一段时间对所述燃气轮机系统中Z个监测节点的运行状态进行监测，从而得到m次监测值向量所组成的燃气轮机运行数据集D，记为D＝{sam1,sam2,...,samv,...,samm}，其中，samv表示第v次监测值向量，且

Figure BDA0002490695650000021
表示第v次监测值向量中第i个监测节点的监测值；1≤v≤m，1≤i≤Z；将m次监测下的第i个监测节点的监测值所组成的向量定义为Xi，其特点是，所述燃气轮机故障预测是按如下步骤进行：A gas turbine failure prediction method based on correlation analysis of the present invention is applied to a gas turbine system, and monitors the operating states of Z monitoring nodes in the gas turbine system at regular intervals, thereby obtaining m times of monitoring value vectors Composed gas turbine operation data set D, recorded as D={sam 1 , sam 2 ,...,sam v ,...,sam m }, where sam v represents the vth monitoring value vector, and
Figure BDA0002490695650000021
Indicates the monitoring value of the i-th monitoring node in the v-th monitoring value vector; 1≤v≤m, 1≤i≤Z; the vector formed by the monitoring value of the i-th monitoring node under m monitoring is defined as X i , which is characterized in that the gas turbine failure prediction is performed according to the following steps:

步骤1、定义时刻t，并初始化t＝0；Step 1, define time t, and initialize t=0;

定义监测节点个数Z的极限值为max，即Z≤max；Define the limit value of the number of monitoring nodes Z to be max, that is, Z≤max;

步骤2、定义所选择的监测节点向量集为EF，并初始化第t时刻所选择的监测节点向量集为

Figure BDA0002490695650000022
Step 2. Define the selected monitoring node vector set as EF, and initialize the monitoring node vector set selected at the tth moment as
Figure BDA0002490695650000022

步骤3、定义变量j，并初始化j＝1；Step 3, define variable j, and initialize j=1;

步骤4、判断j≤Z是否成立，若成立，从燃气轮机运行数据集D中读取具有m个取值的第j个监测节点向量Xj；并初始化第j个监测节点向量Xj的相关监测节点向量集MB(Xj)为空、初始化第j个监测节点向量Xj的新增监测节点向量集FA(Xj)为空、初始化第j个监测节点向量Xj的冗余监测节点向量集FD(Xj)为空；再执行步骤5；否则，表示获得最终的由监测节点构成的因果结构图，其中，每个监测节点的父节点和子节点都是与相应监测节点相关的监测节点，并执行步骤16；Step 4. Determine whether j≤Z is true, if true, read the j-th monitoring node vector X j with m values from the gas turbine operation data set D; and initialize the related monitoring of the j-th monitoring node vector X j The node vector set MB(X j ) is empty, the newly added monitoring node vector set FA(X j ) of the j-th monitoring node vector X j is initialized to be empty, and the redundant monitoring node vector of the j-th monitoring node vector X j is initialized The set FD(X j ) is empty; go to step 5 again; otherwise, it means to obtain the final causal structure graph composed of monitoring nodes, where the parent node and child nodes of each monitoring node are monitoring nodes related to the corresponding monitoring node , and execute step 16;

步骤5、判断j＝1是否成立，若成立，则将所述第j个监测节点向量Xj加入所述第t时刻所选择的监测节点向量集EFt中，从而获得第t+1时刻所述选择的监测节点向量集EFt+1；再将t+1赋值给t、将j+1赋值给j后，返回步骤4；否则执行步骤6；Step 5, judging whether j=1 is established, if established, then adding the j monitoring node vector X j to the monitoring node vector set EF t selected at the t time, so as to obtain the t+1 time The monitoring node vector set EF t+1 selected above; then assign t+1 to t and j+1 to j, then return to step 4; otherwise, execute step 6;

步骤6、对所述第j个监测节点向量Xj进行相关性分析；Step 6, performing a correlation analysis on the jth monitoring node vector X j ;

步骤7、判断所述第j个监测节点向量Xj的相关监测节点向量集MB(Xj)是否为空集，若为空集，则返回步骤4；否则，将第j个监测节点向量Xj加入第t时刻所选择的监测节点向量集EFt中，从而获得第t+1时刻所选择的监测节点向量集EFt+1＝EFt∪Xj；再将t+1赋值给t后，执行步骤8；Step 7, judging whether the related monitoring node vector set MB(X j ) of the jth monitoring node vector X j is an empty set, if it is an empty set, then return to step 4; otherwise, set the jth monitoring node vector X j is added to the monitoring node vector set EF t selected at the tth moment, so as to obtain the monitoring node vector set EF t+1 =EF t ∪X j at the t+1th moment; and then assign t+1 to t , go to step 8;

步骤8、定义变量k，并初始化k＝1；Step 8, define variable k, and initialize k=1;

步骤9、对所述第t时刻所选择的监测节点向量集EFt的第k个监测节点向量Xk进行冗余校验分析；Step 9, performing redundancy check analysis on the kth monitoring node vector X k of the monitoring node vector set EF t selected at the tth moment;

步骤10、将k+1赋值给k，并判断k＞j是否成立，若成立，则执行步骤11；否则返回步骤9执行；Step 10, assign k+1 to k, and judge whether k>j is true, if true, execute step 11; otherwise, return to step 9 for execution;

步骤11、定义变量count；并初始化count＝0；初始化k＝1；Step 11, define variable count; and initialize count=0; initialize k=1;

步骤12、判断所述第k个监测节点向量Xk的相关监测节点向量集MB(Xk)是否为空集，若为空集，则从所述第t时刻所选择的监测节点向量集EFt中删除所述第k个监测节点向量Xk后，再将count+1赋值给count后，执行步骤13；否则直接执行步骤13；Step 12. Judging whether the related monitoring node vector set MB(X k ) of the kth monitoring node vector X k is an empty set, if it is an empty set, then from the monitoring node vector set EF selected at the tth moment After deleting the k-th monitoring node vector X k in t , assigning count+1 to count, execute step 13; otherwise, directly execute step 13;

步骤13、将k+1赋值给k后；判断k＞j是否成立，若成立，则将j-count赋值给j后，获得更新的第t时刻所选择的监测节点向量集EFt′，记为EFt′＝{X′1,X′2,...,X′i,...X′j-count}；X′i表示更新的第t时刻的监测节点向量集EFt′中第i个监测节点向量，并有，

Figure BDA0002490695650000031
表示第i个监测节点向量X′i中第v个监测值；1≤i≤j-count；再将j-count赋值给j后，执行步骤14；否则返回步骤12执行；Step 13. After assigning k+1 to k; judge whether k>j is true, if it is true, assign j-count to j, and obtain the updated monitoring node vector set EF t ' selected at the tth moment, record EF t ′={X′ 1 ,X′ 2 ,...,X′ i ,...X′ j-count }; X′ i represents the updated monitoring node vector set EF t ′ at the tth moment The i-th monitoring node vector, and has,
Figure BDA0002490695650000031
Indicates the v-th monitoring value in the i-th monitoring node vector X′ i ; 1≤i≤j-count; after assigning j-count to j, execute step 14; otherwise, return to step 12 for execution;

步骤14、对所述第t时刻所选择的监测节点向量集EFt′的第j个监测节点向量X′j进行在线的局部定向，从而得到第t时刻因果结构图；Step 14, performing online local orientation on the jth monitoring node vector X'j of the monitoring node vector set EF t ' selected at the tth moment, so as to obtain the causal structure diagram at the tth moment;

步骤15、将j+1赋值给j，返回步骤4；Step 15, assign j+1 to j, and return to step 4;

步骤16、任意选择一个监测节点的监测节点向量，并作为LSTM神经网络模型的输出，再将与所选择的监测节点相关的监测节点向量作为LSTM神经网络模型的输入，从而训练LSTM神经网络模型，从而得到故障预测模型；Step 16, arbitrarily select the monitoring node vector of a monitoring node, and use it as the output of the LSTM neural network model, and then use the monitoring node vector related to the selected monitoring node as the input of the LSTM neural network model, thereby training the LSTM neural network model, So as to get the fault prediction model;

步骤17、实时监测任意一个监测节点的运行状态并获得相应的燃气轮机运行数据集，并利用所述故障预测模型得到实时监测的监测节点的预测值，再将所述预测值与实时监测的监测节点的真实值进行比较，当超过所设定的阈值，则表示相应监测节点可能发生故障，并给出预警提示。Step 17. Real-time monitor the operating status of any monitoring node and obtain the corresponding gas turbine operating data set, and use the fault prediction model to obtain the predicted value of the real-time monitored monitoring node, and then compare the predicted value with the real-time monitored monitoring node Compared with the real value of , when it exceeds the set threshold, it means that the corresponding monitoring node may fail, and an early warning prompt is given.

本发明所述的基于相关性分析的燃气轮机故障预测方法的特点也在于，所述步骤6是按如下步骤进行：The gas turbine failure prediction method based on correlation analysis of the present invention is also characterized in that said step 6 is carried out as follows:

步骤6.1、设置相关性阈值为α；Step 6.1, setting the correlation threshold to α;

步骤6.2、定义变量w；并初始化w＝1；定义变量θ；Step 6.2, define variable w; and initialize w=1; define variable θ;

步骤6.3、用希尔伯特-施密特独立性准则计算第j个监测节点向量Xj和第w个监测节点向量Xw的相关程度HSICjw；Step 6.3, using the Hilbert-Schmidt independence criterion to calculate the correlation degree HSIC jw between the jth monitoring node vector X j and the wth monitoring node vector Xw ;

步骤6.4、将相关程度HSICjw赋值给θ，并判断θ≥α是否成立，若成立，则表示第j个监测节点向量Xj与第w个监测节点向量Xw相关，并执行步骤6.5；否则，表示第j个监测节点向量Xj与第w个监测节点向量Xw独立，并执行步骤6.6；Step 6.4. Assign the degree of correlation HSIC jw to θ, and judge whether θ≥α is established. If it is established, it means that the j-th monitoring node vector X j is related to the w-th monitoring node vector X w , and perform step 6.5; otherwise , indicating that the j-th monitoring node vector X j is independent from the w-th monitoring node vector X w , and perform step 6.6;

步骤6.5、将第j个监测节点向量Xj加到第w个监测节点向量Xw的相关监测节点向量集MB(Xw)中，即MB(Xw)＝MB(Xw)∪Xj，将第j个监测节点向量Xj加入新增监测节点向量集FA(Xw)中，即FA(Xw)＝FA(Xw)∪{Xj}，从而更新第w个监测节点向量Xw的相关监测节点向量集MB(Xw)和新增监测节点向量集FA(Xw)；同时将第w个监测节点向量Xw加到第j个监测节点向量Xj的相关监测节点向量集MB(Xj)，即MB(Xj)＝MB(Xj)∪Xw，将第w个监测节点向量Xw加入新增监测节点向量集FA(Xj)中，即FA(Xj)＝FA(Xj)∪{Xw}，从而更新第j个监测节点向量Xj的相关监测节点向量集MB(Xj)和新增监测节点向量集FA(Xj)；再执行步骤6.6；Step 6.5. Add the j-th monitoring node vector X j to the related monitoring node vector set MB(X w ) of the w-th monitoring node vector X w , that is, MB(X w )=MB(X w )∪X j , add the j-th monitoring node vector X j to the new monitoring node vector set FA(X w ), that is, FA(X w )=FA(X w )∪{X j }, thereby updating the w-th monitoring node vector The related monitoring node vector set MB(X w ) of X w and the new monitoring node vector set FA(X w ); at the same time, add the wth monitoring node vector X w to the related monitoring node of the jth monitoring node vector X j Vector set MB(X j ), that is, MB(X j )=MB(X j )∪X w , add the wth monitoring node vector X w to the new monitoring node vector set FA(X j ), that is, FA( X j )=FA(X j )∪{X w }, thus updating the related monitoring node vector set MB(X j ) of the jth monitoring node vector X j and the new monitoring node vector set FA(X j ); Execute step 6.6;

步骤6.6、将k+1赋值给k，并判断k＞j-1是否成立，若成立，则执行步骤7；否则返回步骤6.3执行。Step 6.6. Assign k+1 to k, and judge whether k>j-1 is true, and if so, execute step 7; otherwise, return to step 6.3 for execution.

步骤9在线的冗余校验分析是按如下步骤进行：The online redundancy check analysis of step 9 is carried out according to the following steps:

步骤9.1、设置冗余度阈值β；计算第k个监测节点向量Xk的相关监测节点向量集MB(Xk)中的监测节点向量个数，记为Sk；Step 9.1, setting the redundancy threshold β; calculating the number of monitoring node vectors in the related monitoring node vector set MB(X k ) of the kth monitoring node vector X k , denoted as S k ;

步骤9.2、定义变量s；并初始化s＝1；定义变量σ；Step 9.2, define variable s; and initialize s=1; define variable σ;

步骤9.3、获取所述相关监测节点向量集MB(Xk)中第s个监测节点向量的下标记为τs；Step 9.3, obtaining the subscript of the sth monitoring node vector in the related monitoring node vector set MB(X k ) as τ s ;

步骤9.4、用希尔伯特-施密特独立性准则计算第τs个监测节点向量

Figure BDA0002490695650000051
和第k个监测节点向量Xk的相关程度
Figure BDA0002490695650000052
Step 9.4, use the Hilbert-Schmidt independence criterion to calculate the τ sth monitoring node vector
Figure BDA0002490695650000051
The degree of correlation with the kth monitoring node vector X k
Figure BDA0002490695650000052

步骤9.5、将所述相关程度

Figure BDA0002490695650000053
赋值给σ，判断σ≤β是否成立，若成立，则表示第τs个监测节点向量
Figure BDA0002490695650000054
和第k个监测节点向量Xk不相关，即为冗余监测节点，并执行步骤9.6；否则，表明第τs个监测节点向量
Figure BDA0002490695650000055
和第k个监测节点向量Xk相关，并执行步骤9.7；Step 9.5, the correlation degree
Figure BDA0002490695650000053
Assign a value to σ to judge whether σ≤β is established, and if it is established, it means that the τ sth monitoring node vector
Figure BDA0002490695650000054
It is irrelevant to the k-th monitoring node vector X k , that is, it is a redundant monitoring node, and execute step 9.6; otherwise, it indicates that the τ s -th monitoring node vector
Figure BDA0002490695650000055
Correlate with the kth monitoring node vector X k , and execute step 9.7;

步骤9.6、从所述第k个监测节点向量Xk的相关监测节点向量集MB(Xk)中删除所述第τs个监测节点向量

Figure BDA0002490695650000056
即
Figure BDA0002490695650000057
并将第τs个监测节点向量
Figure BDA0002490695650000058
加入所述第k个监测节点向量Xk的冗余监测节点向量集FD(Xk)中，即
Figure BDA0002490695650000059
从所述第τs个监测节点向量
Figure BDA00024906956500000510
的相关监测节点向量集
Figure BDA00024906956500000511
中删除所述第k个监测节点向量Xk，即MB(Xτs)＝MB(Xτs)-Xk，并将第k个监测节点向量Xk加入所述第τs个监测节点向量
Figure BDA00024906956500000512
的冗余监测节点向量集
Figure BDA00024906956500000513
即
Figure BDA00024906956500000514
Step 9.6. Delete the τ s monitoring node vector from the related monitoring node vector set MB(X k ) of the k monitoring node vector X k
Figure BDA0002490695650000056
Right now
Figure BDA0002490695650000057
and the τ s monitoring node vector
Figure BDA0002490695650000058
Add to the redundant monitoring node vector set FD(X k ) of the kth monitoring node vector X k , namely
Figure BDA0002490695650000059
From the τ s monitoring node vector
Figure BDA00024906956500000510
The related monitoring node vector set of
Figure BDA00024906956500000511
Delete the k-th monitoring node vector X k in , that is, MB(X τs )=MB(X τs )-X k , and add the k-th monitoring node vector X k to the τ s -th monitoring node vector
Figure BDA00024906956500000512
Redundancy monitoring node vector set of
Figure BDA00024906956500000513
Right now
Figure BDA00024906956500000514

步骤9.7、将s+1赋值给s；并判断s＞Sk是否成立，若成立，则执行步骤10；否则返回步骤9.3执行。Step 9.7. Assign s+1 to s; and judge whether s>S k is true, and if so, execute step 10; otherwise, return to step 9.3 for execution.

所述步骤14是按如下步骤进行：Described step 14 is to carry out as follows:

步骤14.1、设置方向支持度阈值为γ；Step 14.1, set the direction support threshold to γ;

步骤14.2、从第j个监测节点向量Xj的相关监测节点向量集MB(Xj)任选一个监测节点向量Xg，并将第g个监测节点向量Xg从相关监测节点向量集MB(Xi)中删除；Step 14.2. Select a monitoring node vector X g from the related monitoring node vector set MB(X j ) of the j-th monitoring node vector X j , and select the g-th monitoring node vector X g from the related monitoring node vector set MB( X i );

步骤14.3、当第g个监测节点向量Xg作为第j个监测节点向量Xj的父监测节点向量，即Xg→Xj时，利用最小二乘互信息方法的p-value值计算Xg→Xj方向的支持度，并记为p-value(Xg,Xj)；Step 14.3. When the g-th monitoring node vector X g is used as the parent monitoring node vector of the j-th monitoring node vector X j , that is, X g → X j , use the p-value value of the least squares mutual information method to calculate X g →The support degree in the X j direction, and recorded as p-value(X g ,X j );

步骤14.4、当第j个监测节点向量Xj作为第g个监测节点向量Xg的父监测节点向量，即Xj→Xg时，利用最小二乘互信息方法的的p-value值计算Xj→Xg方向的支持度，并记为p-value(Xj,Xg)；Step 14.4. When the j-th monitoring node vector X j is used as the parent monitoring node vector of the g-th monitoring node vector X g , that is, X j → X g , use the p-value value of the least squares mutual information method to calculate X The support degree of j →X g direction, and recorded as p-value(X j ,X g );

步骤14.5、如果p-value(Xg,Xj)＞γ或p-value(Xj,Xg)≤γ，则表示Xg→Xj方向的支持度较大，并定向为Xg→Xj；Step 14.5. If p-value(X g ,X j )>γ or p-value(X j ,X g )≤γ, it means that the support degree of X g →X j direction is large, and it is oriented as X g → Xj ;

如果p-value(Xj,Xg)＞γ或p-value(Xg,Xj)≤γ，则表示Xj→Xg方向的支持度较大，并定向为Xj→Xg；If p-value(X j ,X g )＞γ or p-value(X g ,X j )≤γ, it means that the support degree in the direction of X j →X g is large, and it is oriented as X j →X g ;

如果p-value(Xj,Xg)≤γ或p-value(Xg,Xj)≤γ或p-value(Xg,Xj)＞γ或p-value(Xj,Xg)＞γ，则表示两监测节点向量间没有因果关系，不用定向；If p-value(X j ,X g )≤γ or p-value(X g ,X j )≤γ or p-value(X g ,X j )＞γ or p-value(X j ,X g ) >γ, it means that there is no causal relationship between the two monitoring node vectors, and no orientation is required;

步骤14.6、如果第j个监测节点向量Xj的相关监测节点向量集MB(Xj)为空，则执行步骤15，否则，则返回步骤14.2执行。Step 14.6. If the related monitoring node vector set MB(X j ) of the jth monitoring node vector X j is empty, execute step 15; otherwise, return to step 14.2 for execution.

与已有技术相比，本发明的有益效果体现在：Compared with the prior art, the beneficial effects of the present invention are reflected in:

1、针对燃气轮机机组运行监测数据分布往往是任意的，彼此之间的关系往往具有非线性的特点，本发明基于希尔伯特-施密特独立性准则对监测节点的相关性进行研究，是一个新研究，结合局部学习策略实现监测节点相关监测节点的学习，显著的降低了学习的复杂度，从而满足了燃气轮机状态实时监测的需要。1. The distribution of monitoring data for the operation of gas turbine units is often arbitrary, and the relationship between them often has nonlinear characteristics. The present invention studies the correlation of monitoring nodes based on the Hilbert-Schmidt independence criterion. A new study, combined with local learning strategies to realize the learning of monitoring nodes related to monitoring nodes, significantly reduces the complexity of learning, thus meeting the needs of real-time monitoring of gas turbine status.

2、针对燃气轮机机组运行监测数据，研究监测节点彼此之间的因果联系，通常的贪婪搜索的方法复杂度较大，满足不了在线学习的时效性，本发明采用最小二乘互信息的方式进行定向，计算复杂度显著的降低，满足了在线的燃气轮机状态实时监测。2. For the monitoring data of the gas turbine unit operation, the causal relationship between the monitoring nodes is studied. The usual greedy search method is complex and cannot meet the timeliness of online learning. The present invention adopts the least square mutual information method for orientation , the computational complexity is significantly reduced, and the online real-time monitoring of gas turbine status is satisfied.

3、本发明针对燃气轮机机组运行数据的动态、高维性，以流的方式进行处理，可以处理高维、动态的燃气轮机机组运行监测数据。基于相关性分析和冗余校验从而实现了监测节点的相关相关监测节点集的在线更新，融入实时定向，实现了局部因果结构的快速在线调整，流的处理方式可以降低学习的时间复杂度，从而满足了在线学习的时效性要求，适用于高维动态的燃气轮机机组运行数据。3. The present invention aims at the dynamic and high-dimensional nature of the gas turbine unit operation data, and processes it in a flow manner, and can process high-dimensional and dynamic gas turbine unit operation monitoring data. Based on correlation analysis and redundancy check, the online update of related monitoring node sets of monitoring nodes is realized, and real-time orientation is integrated to realize fast online adjustment of local causal structure. The stream processing method can reduce the time complexity of learning, Therefore, the timeliness requirement of online learning is met, and it is suitable for high-dimensional dynamic gas turbine unit operation data.

具体实施方式Detailed ways

本实施例中，一种基于相关性分析的燃气轮机故障预测方法，是应用于燃气轮机系统中，并每隔一段时间对燃气轮机系统中Z个监测节点的运行状态进行监测，持续记录一段时间，假设共监测m次，从而得到m次监测值向量所组成的燃气轮机运行数据集D，记为D＝{sam1,sam2,...,samv,...,samm}，其中，samv表示第v次监测值向量，且

Figure BDA0002490695650000071
表示第v次监测值向量中第i个监测节点的监测值；1≤v≤m，1≤i≤Z；将m次监测下的第i个监测节点的监测值所组成的向量定义为Xi，
Figure BDA0002490695650000072
表示第i个监测节点向量Xi具有m个取值，因为对第i个监测节点进行了m次监测，就有m个监测值。该燃气轮机故障预测方法目的是为了找出监测节点间的关系，找到与任意监测节点相关性较强的监测节点，并在该方法的基础上，使用神经网络的方法对于监测节点的未来趋势进行预测，从而对为燃气轮机的运行状态进行监测和故障预警。具体的说，该燃气轮机故障预测是按如下步骤进行：In this embodiment, a gas turbine failure prediction method based on correlation analysis is applied to a gas turbine system, and monitors the operating status of Z monitoring nodes in the gas turbine system at regular intervals, and keeps recording for a period of time, assuming a total of Monitor m times, so as to obtain the gas turbine operation data set D composed of m monitoring value vectors, recorded as D={sam 1 ,sam 2 ,...,sam v ,...,sam m }, where, sam v Indicates the vth monitoring value vector, and
Figure BDA0002490695650000071
Indicates the monitoring value of the i-th monitoring node in the v-th monitoring value vector; 1≤v≤m, 1≤i≤Z; the vector formed by the monitoring value of the i-th monitoring node under m monitoring is defined as X i ,
Figure BDA0002490695650000072
Indicates that the i-th monitoring node vector X i has m values, because the i-th monitoring node has been monitored for m times, and there are m monitoring values. The purpose of this gas turbine fault prediction method is to find out the relationship between the monitoring nodes, find the monitoring node with strong correlation with any monitoring node, and on the basis of this method, use the neural network method to predict the future trend of the monitoring node , so as to monitor the operating status of the gas turbine and provide early warning of failures. Specifically, the fault prediction of the gas turbine is carried out as follows:

步骤1、定义时刻t，并初始化t＝0；Step 1, define time t, and initialize t=0;

定义监测节点个数Z的极限值为max，即Z≤max；Define the limit value of the number of monitoring nodes Z to be max, that is, Z≤max;

步骤2、定义所选择的监测节点向量集为EF，并初始化第t时刻所选择的监测节点向量集为

Figure BDA0002490695650000073
Step 2. Define the selected monitoring node vector set as EF, and initialize the monitoring node vector set selected at the tth moment as
Figure BDA0002490695650000073

步骤3、定义变量j，并初始化j＝1；Step 3, define variable j, and initialize j=1;

步骤4、判断j≤Z是否成立，若成立，从燃气轮机运行数据集D中读取具有m个取值的第j个监测节点向量Xj；并初始化第j个监测节点向量Xj的相关监测节点向量集MB(Xj)为空、初始化第j个监测节点向量Xj的新增监测节点向量集FA(Xj)为空、初始化第j个监测节点向量Xj的冗余监测节点向量集FD(Xj)为空；再执行步骤5；否则，表示获得最终的由监测节点构成的因果结构图，且每个监测节点的父节点和子节点都是与相应监测节点相关的监测节点，并执行步骤16；Step 4. Determine whether j≤Z is true, if true, read the j-th monitoring node vector X j with m values from the gas turbine operation data set D; and initialize the related monitoring of the j-th monitoring node vector X j The node vector set MB(X j ) is empty, the newly added monitoring node vector set FA(X j ) of the j-th monitoring node vector X j is initialized to be empty, and the redundant monitoring node vector of the j-th monitoring node vector X j is initialized The set FD(X j ) is empty; go to step 5 again; otherwise, it means that the final causal structure diagram composed of monitoring nodes is obtained, and the parent node and child node of each monitoring node are monitoring nodes related to the corresponding monitoring node, And execute step 16;

步骤5、判断j＝1是否成立，若成立，则将第j个监测节点向量Xj加入第t时刻所选择的监测节点向量集EFt中，从而获得第t+1时刻选择的监测节点向量集EFt+1；再将t+1赋值给t、将j+1赋值给j后，返回步骤4；否则执行步骤6；Step 5. Judging whether j=1 is established, if it is established, add the j-th monitoring node vector X j to the monitoring node vector set EF t selected at the t-th moment, so as to obtain the monitoring node vector selected at the t+1-th moment Set EF t+1 ; then assign t+1 to t and j+1 to j, then return to step 4; otherwise, go to step 6;

步骤6、对第j个监测节点向量Xj进行相关性分析；Step 6, performing a correlation analysis on the jth monitoring node vector X j ;

步骤6.1、设置相关性阈值为α；Step 6.1, setting the correlation threshold to α;

步骤6.2、定义变量w；并初始化w＝1；定义变量θ；Step 6.2, define variable w; and initialize w=1; define variable θ;

步骤6.3、用希尔伯特-施密特独立性准则计算第j个监测节点向量Xj和第w个监测节点向量Xw的相关程度HSICjw；Step 6.3, using the Hilbert-Schmidt independence criterion to calculate the correlation degree HSIC jw between the jth monitoring node vector X j and the wth monitoring node vector Xw ;

依据公式(1)计算独立性准则HSICjw的值：Calculate the value of the independence criterion HSIC jw according to the formula (1):

Figure BDA0002490695650000081
Figure BDA0002490695650000081

n是向量的Xj和Xw维度数，H，K，L都是n行n列的矩阵，Kij＝k(xi,xj)，Lij＝l(xi,xj)，Kij和Lij是映射的核函数，H＝I-n-111T，1是n×1的全1向量，trace是矩阵的迹运算。n is the number of X j and X w dimensions of the vector, H, K, and L are all matrixes of n rows and n columns, K ij =k( xi ,x j ), L ij =l( xi ,x j ), K ij and L ij are the kernel functions of the mapping, H=In -1 11 T , 1 is an n×1 full 1 vector, and trace is the trace operation of the matrix.

步骤6.4、将相关程度HSICjw赋值给θ，并判断θ≥α是否成立，若成立，则表示第j个监测节点向量Xj与第w个监测节点向量Xw相关，并执行步骤6.5；否则，表示第j个监测节点向量Xj与第w个监测节点向量Xw独立，并执行步骤6.6；Step 6.4. Assign the degree of correlation HSIC jw to θ, and judge whether θ≥α is established. If it is established, it means that the j-th monitoring node vector X j is related to the w-th monitoring node vector X w , and perform step 6.5; otherwise , indicating that the j-th monitoring node vector X j is independent from the w-th monitoring node vector X w , and perform step 6.6;

步骤6.5、将第j个监测节点向量Xj加到第w个监测节点向量Xw的相关监测节点向量集MB(Xw)中，即MB(Xw)＝MB(Xw)∪Xj，将第j个监测节点向量Xj加入新增监测节点向量集FA(Xw)中，即FA(Xw)＝FA(Xw)∪{Xj}，从而更新第w个监测节点向量Xw的相关监测节点向量集MB(Xw)和新增监测节点向量集FA(Xw)；同时将第w个监测节点向量Xw加到第j个监测节点向量Xj的相关监测节点向量集MB(Xj)，即MB(Xj)＝MB(Xj)∪Xw，将第w个监测节点向量Xw加入新增监测节点向量集FA(Xj)中，即FA(Xj)＝FA(Xj)∪{Xw}，从而更新第j个监测节点向量Xj的相关监测节点向量集MB(Xj)和新增监测节点向量集FA(Xj)；再执行步骤6.6；Step 6.5. Add the j-th monitoring node vector X j to the related monitoring node vector set MB(X w ) of the w-th monitoring node vector X w , that is, MB(X w )=MB(X w )∪X j , add the j-th monitoring node vector X j to the new monitoring node vector set FA(X w ), that is, FA(X w )=FA(X w )∪{X j }, thereby updating the w-th monitoring node vector The related monitoring node vector set MB(X w ) of X w and the new monitoring node vector set FA(X w ); at the same time, add the wth monitoring node vector X w to the related monitoring node of the jth monitoring node vector X j Vector set MB(X j ), that is, MB(X j )=MB(X j )∪X w , add the wth monitoring node vector X w to the new monitoring node vector set FA(X j ), that is, FA( X j )=FA(X j )∪{X w }, thus updating the related monitoring node vector set MB(X j ) of the jth monitoring node vector X j and the new monitoring node vector set FA(X j ); Execute step 6.6;

步骤6.6、将k+1赋值给k，并判断k＞j-1是否成立，若成立，则执行步骤7；否则返回步骤6.3执行。Step 6.6. Assign k+1 to k, and judge whether k>j-1 is true, and if so, execute step 7; otherwise, return to step 6.3 for execution.

步骤7、判断第j个监测节点向量Xj的相关监测节点向量集MB(Xj)是否为空集，若为空集，则返回步骤4；否则，将第j个监测节点向量Xj加入第t时刻所选择的监测节点向量集EFt中，从而获得第t+1时刻所选择的监测节点向量集EFt+1＝EFt∪Xj；再将t+1赋值给t后，执行步骤8；Step 7. Determine whether the related monitoring node vector set MB(X j ) of the j-th monitoring node vector X j is an empty set, and if it is an empty set, return to step 4; otherwise, add the j-th monitoring node vector X j to In the monitoring node vector set EF t selected at the tth moment, the monitoring node vector set EF t+1 selected at the t+1st moment is obtained EF t+1 = EF t ∪X j ; after assigning t+1 to t, execute Step 8;

步骤8、定义变量k，并初始化k＝1；Step 8, define variable k, and initialize k=1;

步骤9、对第t时刻所选择的监测节点向量集EFt的第k个监测节点向量Xk进行冗余校验分析；Step 9, performing redundancy check analysis on the kth monitoring node vector X k of the monitoring node vector set EF t selected at the tth moment;

步骤9.1、设置冗余度阈值β；计算第k个监测节点向量Xk的相关监测节点向量集MB(Xk)中的监测节点向量个数，记为Sk；Step 9.1, setting the redundancy threshold β; calculating the number of monitoring node vectors in the related monitoring node vector set MB(X k ) of the kth monitoring node vector X k , denoted as S k ;

步骤9.2、定义变量s；并初始化s＝1；定义变量σ；Step 9.2, define variable s; and initialize s=1; define variable σ;

步骤9.3、获取相关监测节点向量集MB(Xk)中第s个监测节点向量的下标记为τs；Step 9.3, obtaining the subscript of the sth monitoring node vector in the relevant monitoring node vector set MB(X k ) is τ s ;

步骤9.4、用希尔伯特-施密特独立性准则计算第τs个监测节点向量

Figure BDA0002490695650000091
和第k个监测节点向量Xk的相关程度
Figure BDA0002490695650000092
Step 9.4, use the Hilbert-Schmidt independence criterion to calculate the τ sth monitoring node vector
Figure BDA0002490695650000091
The degree of correlation with the kth monitoring node vector X k
Figure BDA0002490695650000092

步骤9.5、将相关程度

Figure BDA0002490695650000093
赋值给σ，判断σ≤β是否成立，若成立，则表示第τs个监测节点向量
Figure BDA0002490695650000094
和第k个监测节点向量Xk不相关，即为冗余监测节点，并执行步骤9.6；否则，表明第τs个监测节点向量
Figure BDA0002490695650000095
和第k个监测节点向量Xk相关，并执行步骤9.7；Step 9.5, the degree of correlation
Figure BDA0002490695650000093
Assign a value to σ to judge whether σ≤β is established, and if it is established, it means that the τ sth monitoring node vector
Figure BDA0002490695650000094
It is irrelevant to the k-th monitoring node vector X k , that is, it is a redundant monitoring node, and execute step 9.6; otherwise, it indicates that the τ s -th monitoring node vector
Figure BDA0002490695650000095
Correlate with the kth monitoring node vector X k , and execute step 9.7;

步骤9.6、从第k个监测节点向量Xk的相关监测节点向量集MB(Xk)中删除第τs个监测节点向量

Figure BDA0002490695650000096
即
Figure BDA0002490695650000097
并将第τs个监测节点向量
Figure BDA0002490695650000098
加入第k个监测节点向量Xk的冗余监测节点向量集FD(Xk)中，即
Figure BDA0002490695650000099
从第τs个监测节点向量
Figure BDA00024906956500000910
的相关监测节点向量集
Figure BDA00024906956500000911
中删除第k个监测节点向量Xk，即
Figure BDA00024906956500000912
并将第k个监测节点向量Xk加入第τs个监测节点向量
Figure BDA00024906956500000913
的冗余监测节点向量集
Figure BDA00024906956500000914
即
Figure BDA00024906956500000915
Step 9.6. Delete the τ s monitoring node vector from the related monitoring node vector set MB(X k ) of the k monitoring node vector X k
Figure BDA0002490695650000096
Right now
Figure BDA0002490695650000097
and the τ s monitoring node vector
Figure BDA0002490695650000098
Add to the redundant monitoring node vector set FD(X k ) of the kth monitoring node vector X k , namely
Figure BDA0002490695650000099
From the τ sth monitoring node vector
Figure BDA00024906956500000910
The related monitoring node vector set of
Figure BDA00024906956500000911
Delete the kth monitoring node vector X k in , namely
Figure BDA00024906956500000912
And add the k-th monitoring node vector X k to the τ s -th monitoring node vector
Figure BDA00024906956500000913
Redundancy monitoring node vector set of
Figure BDA00024906956500000914
Right now
Figure BDA00024906956500000915

步骤9.7、将s+1赋值给s；并判断s＞Sk是否成立，若成立，则执行步骤10；否则返回步骤9.3执行。Step 9.7. Assign s+1 to s; and judge whether s>S k is true, and if so, execute step 10; otherwise, return to step 9.3 for execution.

步骤10、将k+1赋值给k，并判断k＞j是否成立，若成立，则执行步骤11；否则返回步骤9执行；Step 10, assign k+1 to k, and judge whether k>j is true, if true, execute step 11; otherwise, return to step 9 for execution;

步骤11、定义变量count；并初始化count＝0；初始化k＝1；Step 11, define variable count; and initialize count=0; initialize k=1;

步骤12、判断第k个监测节点向量Xk的相关监测节点向量集MB(Xk)是否为空集，若为空集，则从第t时刻所选择的监测节点向量集EFt中删除第k个监测节点向量Xk后，再将count+1赋值给count后，执行步骤13；否则直接执行步骤13；Step 12. Determine whether the related monitoring node vector set MB(X k ) of the kth monitoring node vector X k is an empty set, and if it is an empty set, delete the monitoring node vector set EF t selected at the tth moment After k monitoring node vectors X k , assign count+1 to count, then execute step 13; otherwise, directly execute step 13;

步骤13、将k+1赋值给k后；判断k＞j是否成立，若成立，则将j-count赋值给j后，获得更新的第t时刻所选择的监测节点向量集EFt′，记为EFt′＝{X′1,X′2,...,X′i,...X′j-count}；X′i表示更新的第t时刻的监测节点向量集EFt′中第i个监测节点向量，并有，

Figure BDA0002490695650000101
表示第i个监测节点向量X′i中第v个监测值；1≤i≤j-count；再将j-count赋值给j后，执行步骤14；否则返回步骤12执行；Step 13. After assigning k+1 to k; judge whether k>j is true, if it is true, assign j-count to j, and obtain the updated monitoring node vector set EF t ' selected at the tth moment, record EF t ′={X′ 1 ,X′ 2 ,...,X′ i ,...X′ j-count }; X′ i represents the updated monitoring node vector set EF t ′ at the tth moment The i-th monitoring node vector, and has,
Figure BDA0002490695650000101
Indicates the v-th monitoring value in the i-th monitoring node vector X′ i ; 1≤i≤j-count; after assigning j-count to j, execute step 14; otherwise, return to step 12 for execution;

步骤14、对第t时刻所选择的监测节点向量集EFt′的第j个监测节点向量X′j进行在线的局部定向，从而得到第t时刻因果结构图；Step 14. Carry out online local orientation to the jth monitoring node vector X'j of the monitoring node vector set EF t ' selected at the tth moment, so as to obtain the causal structure diagram at the tth moment;

步骤14.1、设置方向支持度阈值为γ；Step 14.1, set the direction support threshold to γ;

步骤14.2、从第j个监测节点向量Xj的相关监测节点向量集MB(Xj)任选一个监测节点向量Xg，并将第g个监测节点向量Xg从相关监测节点向量集MB(Xi)中删除；Step 14.2. Select a monitoring node vector X g from the related monitoring node vector set MB(X j ) of the j-th monitoring node vector X j , and select the g-th monitoring node vector X g from the related monitoring node vector set MB( X i );

步骤14.3、当第g个监测节点向量Xg作为第j个监测节点向量Xj的父监测节点向量，即Xg→Xj时，利用最小二乘互信息方法的p-value值计算Xg→Xj方向的支持度，并记为p-value(Xg,Xj)；采用最小二乘互信息方法的p-value表示方向的支持度原因在于，如MakotoYamada和Masashi Sugiyama所著的文献《Dependence Minimizing Regression withModel Selection forNon-Linear Causal Inference underNon-GaussianNoise》，对于非线性非高斯的数据，最小二乘互信息方法的p-value可以衡量方向的支持度，计算方法见上述论文。Step 14.3. When the g-th monitoring node vector X g is used as the parent monitoring node vector of the j-th monitoring node vector X j , that is, X g → X j , use the p-value value of the least squares mutual information method to calculate X g → The support degree in the X j direction, and recorded as p-value(X g , X j ); the p-value using the least squares mutual information method indicates the support degree of the direction. The reason is that, such as the literature written by MakotoYamada and Masashi Sugiyama "Dependence Minimizing Regression with Model Selection forNon-Linear Causal Inference underNon-GaussianNoise", for nonlinear and non-Gaussian data, the p-value of the least squares mutual information method can measure the support of the direction, and the calculation method is shown in the above paper.

步骤14.4、当第j个监测节点向量Xj作为第g个监测节点向量Xg的父监测节点向量，即Xj→Xg时，利用最小二乘互信息方法的的p-value值计算Xj→Xg方向的支持度，并记为p-value(Xj,Xg)；Step 14.4. When the j-th monitoring node vector X j is used as the parent monitoring node vector of the g-th monitoring node vector X g , that is, X j → X g , use the p-value value of the least squares mutual information method to calculate X The support degree of j →X g direction, and recorded as p-value(X j ,X g );

步骤14.5、如果p-value(Xg,Xj)＞γ或p-value(Xj,Xg)≤γ，则表示Xg→Xj方向的支持度较大，并定向为Xg→Xj；Step 14.5. If p-value(X g ,X j )>γ or p-value(X j ,X g )≤γ, it means that the support degree of X g →X j direction is large, and it is oriented as X g → Xj ;

如果p-value(Xj,Xg)＞γ或p-value(Xg,Xj)≤γ，则表示Xj→Xg方向的支持度较大，并定向为Xj→Xg；If p-value(X j ,X g )＞γ or p-value(X g ,X j )≤γ, it means that the support degree in the direction of X j →X g is large, and it is oriented as X j →X g ;

如果p-value(Xj,Xg)≤γ或p-value(Xg,Xj)≤γ或p-value(Xg,Xj)＞γ或p-value(Xj,Xg)＞γ，则表示两监测节点向量间没有因果关系，不用定向；If p-value(X j ,X g )≤γ or p-value(X g ,X j )≤γ or p-value(X g ,X j )＞γ or p-value(X j ,X g ) >γ, it means that there is no causal relationship between the two monitoring node vectors, and no orientation is required;

步骤14.6、如果第j个监测节点向量Xj的相关监测节点向量集MB(Xj)为空，则执行步骤15，否则，则返回步骤14.2执行。Step 14.6. If the related monitoring node vector set MB(X j ) of the jth monitoring node vector X j is empty, execute step 15; otherwise, return to step 14.2 for execution.

步骤15、将j+1赋值给j，返回步骤4；Step 15, assign j+1 to j, and return to step 4;

步骤16、任意选择一个监测节点的监测节点向量，并作为LSTM神经网络模型的输出，再将与所选择的监测节点相关的监测节点向量作为LSTM神经网络模型的输入，从而训练LSTM神经网络模型，从而得到故障预测模型；Step 16, arbitrarily select the monitoring node vector of a monitoring node, and use it as the output of the LSTM neural network model, and then use the monitoring node vector related to the selected monitoring node as the input of the LSTM neural network model, thereby training the LSTM neural network model, So as to get the fault prediction model;

步骤17、实时监测任意一个监测节点的运行状态并获得相应的燃气轮机运行数据集，并利用故障预测模型得到实时监测的监测节点的预测值，再将预测值与实时监测的监测节点的真实值进行比较，当超过所设定的阈值，则表示相应监测节点可能发生故障，并给出预警提示。Step 17. Monitor the operation status of any monitoring node in real time and obtain the corresponding gas turbine operation data set, and use the fault prediction model to obtain the predicted value of the real-time monitored monitoring node, and then compare the predicted value with the real value of the real-time monitored monitoring node In comparison, when the set threshold is exceeded, it indicates that the corresponding monitoring node may fail, and an early warning prompt is given.

Claims (4)

1. A gas turbine fault prediction method based on correlation analysis is applied to a gas turbine system, and the operation states of Z monitoring nodes in the gas turbine system are monitored at intervals, so that a gas turbine operation data set D consisting of m monitoring value vectors is obtained and is recorded as D = { sam 1 ,sam 2 ,...,sam v ,...,sam m Wherein, sam v Represents a vector of the v-th monitored value, and
Figure FDA0002490695640000011
Figure FDA0002490695640000012
representing the monitoring value of the ith monitoring node in the v-th monitoring value vector; v is more than or equal to 1 and less than or equal to m, i is more than or equal to 1 and less than or equal to Z; defining a vector formed by monitoring values of the ith monitoring node under m-time monitoring as X i The method is characterized in that the fault prediction of the gas turbine is carried out according to the following steps:
step 1, defining a time t, and initializing t =0;
defining the limit value of the number Z of the monitoring nodes as max, namely Z is less than or equal to max;
step 2, defining the selected monitoring node vector set as EF, and initializing the selected monitoring node vector set at the t-th moment as
Figure FDA0002490695640000013
Step 3, defining a variable j, and initializing j =1;
step 4, judging whether the j is less than or equal to Z, if so, reading a jth monitoring node vector X with m values from a gas turbine operation data set D j (ii) a And initializing the jth monitor node vector X j Relevant monitoring node vector set MB (X) j ) Initializing jth monitoring node vector X for null j Newly added monitor node vector set FA (X) j ) Initializing jth monitoring node vector X for null j Redundant monitor node vector set FD (X) j ) Is empty; then step 5 is executed; otherwise, the final causal structure diagram formed by the monitoring nodes is obtained, wherein the father node and the child node of each monitoring node are the monitoring nodes related to the corresponding monitoring node, and step 16 is executed;
step 5, judging whether j =1 is true, if so, determining the jth monitoring node vector X j Adding the selected monitoring node vector set EF at the t moment t Thereby obtaining the selected monitoring node vector set EF at the t +1 th moment t+1 (ii) a Assigning t +1 to t and j +1 to j, and returning to the step 4; otherwise, executing step 6;
step 6, for the jth monitoring node vector X j Performing correlation analysis;
step 7, judging the jth monitoring node vector X j Of the relevant monitoring node vector set MB (X) j ) If the set is an empty set, returning to the step 4; otherwise, the jth monitoring node vector X is used j Adding the monitoring node vector set EF selected at the t moment t FromAnd obtaining a monitoring node vector set EF selected at the t +1 th moment t+1 ＝EF t ∪X j (ii) a Assigning t +1 to t, and executing the step 8;
step 8, defining a variable k, and initializing k =1;
step 9, for the monitoring node vector set EF selected at the t-th moment t Of the kth monitoring node vector X k Performing redundancy check analysis;
step 10, assigning k +1 to k, judging whether k is greater than j, and executing step 11 if k is greater than j; otherwise, returning to the step 9 for execution;
step 11, defining a variable count; and initializing count =0; initializing k =1;
step 12, judging the k-th monitoring node vector X k Relevant monitoring node vector set MB (X) k ) Whether the vector set is an empty set or not, if the vector set is the empty set, selecting a monitoring node vector set EF from the t-th moment t Deleting the k-th monitoring node vector X k Then, after assigning the count +1 to the count, executing step 13; otherwise, directly executing step 13;
step 13, assigning k +1 to k; judging whether k is more than j, if so, assigning j-count to j, and then obtaining an updated monitoring node vector set EF selected at the t-th moment t ', is marked as EF t ′＝{X′ 1 ,X′ 2 ,...,X′ i ,...X′ j-count }；X′ i Set of monitoring node vectors EF representing the updated t-th moment t The ith monitor node vector in' and, in combination,
Figure FDA0002490695640000021
Figure FDA0002490695640000022
representing the ith monitor node vector X' i The middle-th monitoring value; i is more than or equal to 1 and less than or equal to j-count; assigning j-count to j, and executing step 14; otherwise, returning to the step 12 for execution;
step 14, for the monitoring node vector set EF selected at the t-th time t 'j-th monitor node vector X' j Carrying out on-line local orientation to obtain a t-time cause and effect structure diagram;
step 15, assigning j +1 to j, and returning to the step 4;
step 16, selecting a monitoring node vector of a monitoring node at will, and using the monitoring node vector as the output of the LSTM neural network model, and using the monitoring node vector related to the selected monitoring node as the input of the LSTM neural network model, so as to train the LSTM neural network model, thereby obtaining a fault prediction model;
and step 17, monitoring the operation state of any monitoring node in real time, obtaining a corresponding gas turbine operation data set, obtaining a predicted value of the monitoring node monitored in real time by using the fault prediction model, comparing the predicted value with a real value of the monitoring node monitored in real time, indicating that the corresponding monitoring node is likely to have a fault when the predicted value exceeds a set threshold value, and giving an early warning prompt.
2. The correlation analysis-based gas turbine engine failure prediction method of claim 1, wherein the step 6 is performed as follows:
step 6.1, setting a correlation threshold value as alpha;
step 6.2, defining a variable w; and initializing w =1; defining a variable theta;
step 6.3, calculating the jth monitoring node vector X by using a Hilbert-Schmidt independence criterion j And w-th monitor node vector X w Degree of correlation HSIC jw ；
Step 6.4, correlating degree HSIC jw Assigning to theta, judging whether theta is more than or equal to alpha, and if so, representing the jth monitoring node vector X j And w-th monitor node vector X w Correlating and executing step 6.5; otherwise, the j monitoring node vector X is represented j And the w-th monitor node vector X w Independently, and perform step 6.6;
step 6.5, monitoring the jth node vector X j Vector X added to the w-th monitor node w In a related monitorVector set MB (X) of measured nodes w ) In (1), i.e. MB (X) w )＝MB(X w )∪X j The jth monitor node vector X j Adding newly added monitoring node vector set FA (X) w ) In (1), namely FA (X) w )＝FA(X w )∪{X j H, thereby updating the w-th monitor node vector X w Of the relevant monitoring node vector set MB (X) w ) And a newly added monitoring node vector set FA (X) w ) (ii) a At the same time, the w-th monitoring node vector X w Vector X added to jth monitor node j Of the relevant monitoring node vector set MB (X) j ) I.e. MB (X) j )＝MB(X j )∪X w The w-th monitor node vector X w Adding newly added monitoring node vector set FA (X) j ) In (1), namely FA (X) j )＝FA(X j )∪{X w H, thereby updating the jth monitor node vector X j Relevant monitoring node vector set MB (X) j ) And a newly added monitoring node vector set FA (X) j ) (ii) a Then step 6.6 is executed;
6.6, assigning k +1 to k, judging whether k is greater than j-1, and if so, executing step 7; otherwise, the step 6.3 is returned to for execution.
3. The correlation analysis-based gas turbine engine failure prediction method of claim 1, wherein the step 9 of on-line redundancy check analysis is performed by the steps of:
step 9.1, setting a redundancy threshold value beta; calculating the kth monitoring node vector X k Of the relevant monitoring node vector set MB (X) k ) The number of the monitoring node vectors in (1) is recorded as S k ；
Step 9.2, defining a variable s; and initializing s =1; defining a variable sigma;
step 9.3, obtaining the relevant monitoring node vector set MB (X) k ) The vector of the middle s-th monitoring node is denoted by tau s ；
Step 9.4, calculating the t < th > by using Hilbert-Schmidt independence criterion s Vector of monitoring nodes
Figure FDA0002490695640000031
And the kth monitoring node vector X k Degree of correlation of
Figure FDA0002490695640000041
Step 9.5, correlating the degree of correlation
Figure FDA0002490695640000042
Assigning to sigma, judging whether sigma is less than or equal to beta, if yes, representing the tau s Vector of monitoring nodes
Figure FDA0002490695640000043
And the kth monitoring node vector X k Irrelevant, namely, the node is a redundant monitoring node, and step 9.6 is executed; otherwise, it indicates the τ th s Vector of monitoring nodes
Figure FDA0002490695640000044
And the kth monitoring node vector X k Correlation and step 9.7 is performed;
step 9.6, from the k-th monitoring node vector X k Of the relevant monitoring node vector set MB (X) k ) In which the τ th is deleted s A vector of monitoring nodes
Figure FDA0002490695640000045
Namely that
Figure FDA0002490695640000046
And will be tau s Vector of monitoring nodes
Figure FDA0002490695640000047
Adding the kth monitoring node vector X k Redundant monitor node vector set FD (X) k ) In, i.e.
Figure FDA0002490695640000048
From the th τ s Vector of monitoring nodes
Figure FDA0002490695640000049
Relevant monitoring node vector set
Figure FDA00024906956400000410
Deleting the k-th monitoring node vector X k I.e. by
Figure FDA00024906956400000411
And the kth monitoring node vector X k Adding the τ th s A vector of monitoring nodes
Figure FDA00024906956400000412
Redundant monitoring node vector set
Figure FDA00024906956400000413
Namely, it is
Figure FDA00024906956400000414
Step 9.7, assigning s +1 to s; and judging S > S k If yes, executing step 10; otherwise, the step 9.3 is returned to for execution.
4. The correlation analysis-based gas turbine engine fault prediction method of claim 1, wherein said step 14 is performed by:
step 14.1, setting a direction support degree threshold value as gamma;
step 14.2 from the jth monitor node vector X j Of the relevant monitoring node vector set MB (X) j ) Optionally one monitoring node vector X g And the g-th monitoring node vector X g Vector set MB (X) from related monitoring nodes i ) Deleting;
step 14.3, when the g-th monitoring node vector X g As the jth monitor node vector X j Of parent monitoring node vectors, i.e. X g →X j When usingP-value calculation X of least square mutual information method g →X j Support of orientation and is denoted as p-value (X) g ,X j )；
Step 14.4, when the jth monitoring node vector X j As the g-th monitor node vector X g Of parent monitoring node vectors, i.e. X j →X g Then, X is calculated by using the p-value of the least square mutual information method j →X g Support of orientation and is denoted as p-value (X) j ,X g )；
Step 14.5, if p-value (X) g ,X j ) Gamma or p-value (X) j ,X g ) Gamma is less than or equal to gamma, then X is represented g →X j The direction has larger support degree and is oriented as X g →X j ；
If p-value (X) j ,X g ) Gamma or p-value (X) g ,X j ) Gamma is less than or equal to gamma, then X is represented j →X g The direction has larger support degree and is oriented as X j →X g ；
If p-value (X) j ,X g ) Gamma or p-value (X) g ,X j ) Gamma or p-value (X) g ,X j ) Gamma or p-value (X) j ,X g ) If the value is more than gamma, no causal relationship exists between the two monitoring node vectors, and no orientation is needed;
step 14.6, if the jth monitoring node vector X j Of the relevant monitoring node vector set MB (X) j ) If it is empty, step 15 is executed, otherwise, step 14.2 is returned to.
CN202010404291.6A 2020-05-13 2020-05-13 Gas turbine fault prediction method based on correlation analysis Active CN111598220B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010404291.6A CN111598220B (en) 2020-05-13 2020-05-13 Gas turbine fault prediction method based on correlation analysis

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010404291.6A CN111598220B (en) 2020-05-13 2020-05-13 Gas turbine fault prediction method based on correlation analysis

Publications (2)

Publication Number Publication Date
CN111598220A CN111598220A (en) 2020-08-28
CN111598220B true CN111598220B (en) 2023-03-14

Family

ID=72185425

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010404291.6A Active CN111598220B (en) 2020-05-13 2020-05-13 Gas turbine fault prediction method based on correlation analysis

Country Status (1)

Country Link
CN (1) CN111598220B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119472562A (en) * 2025-01-13 2025-02-18 山东历控能源有限公司 A smart energy consumption analysis method and system

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011043108A1 (en) * 2009-10-09 2011-04-14 株式会社日立製作所 Equipment status monitoring method, monitoring system, and monitoring program
CN109061341A (en) * 2018-07-10 2018-12-21 杭州安脉盛智能技术有限公司 Kalman filtering transformer fault prediction technique and system neural network based

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011043108A1 (en) * 2009-10-09 2011-04-14 株式会社日立製作所 Equipment status monitoring method, monitoring system, and monitoring program
CN109061341A (en) * 2018-07-10 2018-12-21 杭州安脉盛智能技术有限公司 Kalman filtering transformer fault prediction technique and system neural network based

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于SVM-BN的天然气长输管道燃气轮机故障预警方法研究;王明达等;《天然气工业》;20171125(第11期);全文 *

Also Published As

Publication number Publication date
CN111598220A (en) 2020-08-28

Similar Documents

Publication Publication Date Title
CN110321603B (en) Depth calculation model for gas path fault diagnosis of aircraft engine
Caceres et al. A probabilistic Bayesian recurrent neural network for remaining useful life prognostics considering epistemic and aleatory uncertainties
Shang et al. A remaining life prediction of rolling element bearings based on a bidirectional gate recurrent unit and convolution neural network
Gao et al. A neural network-based joint prognostic model for data fusion and remaining useful life prediction
Ye et al. Health condition monitoring of machines based on long short-term memory convolutional autoencoder
Ayodeji et al. Causal augmented ConvNet: A temporal memory dilated convolution model for long-sequence time series prediction
Xu et al. Global attention mechanism based deep learning for remaining useful life prediction of aero-engine
CN110807257A (en) Method for predicting residual life of aircraft engine
Rengasamy et al. Deep learning approaches to aircraft maintenance, repair and overhaul: A review
Chang et al. Remaining useful life prediction for rolling bearings using multi-layer grid search and LSTM
CN112001110B (en) Structural damage identification monitoring method based on vibration signal space real-time recurrent graph convolutional neural network
CN112488235A (en) Elevator time sequence data abnormity diagnosis method based on deep learning
Ma et al. Deep recurrent convolutional neural network for remaining useful life prediction
CN111122811A (en) A fault monitoring method for sewage treatment process based on OICA and RNN fusion model
CN113295399B (en) Converter valve element state evaluation method, device, electronic equipment and storage medium
Huang et al. Bayesian neural network based method of remaining useful life prediction and uncertainty quantification for aircraft engine
CN118228086A (en) Bridge deformation monitoring method, device and medium
CN114676887A (en) A river water quality prediction method based on graph convolution STG-LSTM
CN115081484A (en) Fault diagnosis method of aero-engine sensor based on CRJ-OSELM algorithm
Hong et al. Remaining useful life prediction using time-frequency feature and multiple recurrent neural networks
CN113962318A (en) Transformer state anomaly detection method based on internal and external data fusion and machine learning
CN111598220B (en) Gas turbine fault prediction method based on correlation analysis
Zhao et al. Remaining useful life prediction method based on convolutional neural network and long short-term memory neural network
CN111859799A (en) Method and device for evaluating data accuracy based on complex electromechanical system coupling relationship model
Chen et al. Reliability analysis using deep learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant