Nothing Special   »   [go: up one dir, main page]

CN112101480B - Multivariate clustering and fused time sequence combined prediction method - Google Patents

Multivariate clustering and fused time sequence combined prediction method Download PDF

Info

Publication number
CN112101480B
CN112101480B CN202011034152.5A CN202011034152A CN112101480B CN 112101480 B CN112101480 B CN 112101480B CN 202011034152 A CN202011034152 A CN 202011034152A CN 112101480 B CN112101480 B CN 112101480B
Authority
CN
China
Prior art keywords
monitoring
causal
variables
variable
network
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
CN202011034152.5A
Other languages
Chinese (zh)
Other versions
CN112101480A (en
Inventor
谢军太
黄婧
高智勇
高建民
姜洪权
席越
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202011034152.5A priority Critical patent/CN112101480B/en
Publication of CN112101480A publication Critical patent/CN112101480A/en
Application granted granted Critical
Publication of CN112101480B publication Critical patent/CN112101480B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Probability & Statistics with Applications (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a multivariate clustering and fused time sequence combined prediction method, which aims at solving the problems that the existing neural network model has no specific learning mechanism and the data structure characteristic information is not sufficiently mined, and provides a multivariate clustering and fused time sequence combined prediction method from the multivariate directional coupling angle by combining the advantages of a graph convolution neural network and a long-term and short-term memory network. Firstly, researching the causal transfer relationship among variables based on coupled Glanduger causal measurement analysis; secondly, establishing a directed weighting network according to the causal analysis result of the variables, extracting the node and edge weight characteristics of the directed weighting network, embedding the weight of the target variable into a graph convolution neural network for training, and realizing the accurate classification of the monitored variables; and finally, taking a non-target monitoring variable time sequence contained in the community where the target monitoring variable is located as input, and predicting the target monitoring variable based on the long-term and short-term memory neural network. The method is verified by applying a monitoring sequence of a compressor unit in a chemical production system, and the result shows that: the method is superior to the traditional node classification method in the aspects of prediction accuracy and calculation complexity, and the method can keep stronger prediction capability in the abnormal state of the system.

Description

Multivariate clustering and fused time sequence combined prediction method
Technical Field
The invention relates to the technical field of time sequence prediction, in particular to a multivariate clustering and fused time sequence combined prediction method.
Background
The complex electromechanical system of the process industry represented by the energy chemical industry relates to the exchange of various media such as energy, information, substances and the like, and has the characteristics of harsh process conditions, large-scale device, continuous processing process, fine production control requirement and the like. The mutual coupling of all monitored variables in the system substantially forms a network diagram representing the dynamic change of the complex electromechanical system, and the basis for constructing the complex network is to determine the representation relationship of all nodes, connecting edges and even weights of all nodes, so that various researchers perform various researches on the correlation relationship among the variables. For example, when constructing a complex comprehensive network of marine traffic forms, Sui et al determines the convergence and divergence relationship between vessels by calculating the approach rate, thereby determining whether two vessels are connected; zhou et al calculate correlations between settlement data of different monitoring points using Pearson correlation coefficients, and construct a complex network representing changes in the surrounding environment of the underground building; zhou et al, combined with Chichi information criteria, Granger causal relationship tests and Pearson correlation, etc. methods to generate network rules; g.f.zebende et al quantify air temperature and humidity based on DCCA cross-correlation coefficients; paulo Ferreira et al use DCCA correlation coefficients to find out the relationship between oil prices and the stock market and calculate the difference of DCCA correlation coefficients that vary with time; yunke Zhang et al use the Pearson correlation coefficient for detecting the correlation characteristics of the current derivative, and based on these characteristics, construct a novel long-distance HVDC transmission line protection scheme to detect internal and external faults; in fact, coupling interaction between complex electromechanical system variables is obvious, nonlinear effects are prominent, causal relationships have obvious directivity, Pearson correlation and other methods cannot well process nonlinear non-stationary sequences, correlation coefficients solved by DCCA and other methods cannot represent directivity, and therefore the method has shortcomings in measuring nonlinear relationships between variables and representing directions of continuous edges. Glange causal tests were first proposed by glange in 1969 and have been widely used in the fields of meteorological science, economics, and brain science because they can quantitatively describe the linear relationship between variables. But since the Glankey causal test can only determine the linear causal relationship between variables, the application of the method in the multivariate causal relationship analysis of a nonlinear system is limited. In order to solve the problem, Hu and the like can analyze the nonlinear relation between variables by utilizing a Copula function, and provide a Copula-Granger causal analysis method capable of evaluating the nonlinear relation between the variables. The method is based on a Copula theory, the multivariate joint distribution of random variables is expressed by the edge distribution of each variable and a Copula function describing the correlation of the variables, the nonlinear relation between the two variables can be described, Hu et al verify that the Granger method based on Copula is superior to other available grangeje methods; based on the method adopted by Shabbir Dastgir et al, a two-way causal relationship existing between bitcoin attention and bitcoin revenue was observed.
The graph neural network is a recently popular deep learning network, has obvious advantages in the aspect of mining the internal relation of graph structure data, directly acts on the neural network on a graph structure, embeds nodes based on the neighbor information of the nodes, and aggregates the information of each node and the surrounding nodes through the neural network. The study of neural networks in the present figure is largely divided into two important directions. One is graph data mining based on relationship significance: the method mainly aims to research data analysis and prediction of scenes which can easily form network data, and mining is carried out on the basis of the existing network data, wherein the mining comprises node and text classification, node link prediction and the like, such as social networks, natural language processing and the like. Secondly, modeling and reasoning of graph data based on relationship ambiguity have the difficulty that no obvious network data exists in a scene, a network relationship model needs to be established according to the scene, and then relevant relationship reasoning and state prediction are carried out, such as computer vision, molecular property research and the like. Meanwhile, the scholars begin to use the graph convolution neural network for the research of the complex network, for example, Yu et al convert the key node identification problem in the complex network into a regression problem, and provide an RCNN method for ranking the nodes; zhao et al propose an InfGCN deep learning model that identifies the most influential nodes in a complex network based on a graph convolution network. The learners are dedicated to the field of key node discovery and aim to discover nodes playing key roles in the structure and function of a network, but research is still carried out in the fields of complex network community discovery and link prediction, so that the graph convolutional neural network is utilized to train a complex network established based on coupling Glangel causal measure analysis, and the community structure where a target monitoring variable is located in the complex network is discovered, so that the composition of the network nodes is reasonably divided, and the accurate prediction of the target monitoring variable time sequence is realized, which is particularly important.
Disclosure of Invention
Aiming at the problems that an existing neural network model has no specific learning mechanism and the data structure characteristic information is not sufficiently mined, the multivariate directional coupling angle is utilized, the advantages of a graph convolution neural network and a long-term and short-term memory network are combined, the graph convolution neural network is utilized to train a complex network established based on coupling Glan Jack causal measure analysis, and the purpose is to find a community structure where a target monitoring variable is located in the complex network so as to reasonably divide the composition of network nodes, thereby realizing the accurate prediction of the target monitoring variable time sequence.
In order to achieve the purpose, the invention adopts the technical scheme that the multivariate clustering and fusion time sequence combination prediction method comprises the following steps:
step 1), according to data collected by a DCS, defining monitoring variables corresponding to monitoring point positions, determining a variable set of a monitoring target of a complex electromechanical system to be analyzed, selecting a monitoring variable value of each monitoring variable at a certain moment every day, and obtaining monitoring time sequences of the monitoring variables at the same moment corresponding to different continuous days;
step 2), performing CGC coupling Grangel causal measurement analysis on every two monitoring time sequences obtained in the step 1) to obtain the direction and the coupling coefficient of the coupling causal relationship;
step 3), establishing a directed weighting network model representing the bottom layer interaction dynamics of the system by taking the monitoring variables corresponding to the monitoring time sequence obtained in the step 2) as nodes, the coupling causal relationship as edges and the coupling coefficient as edges;
step 4), extracting the connection relation of the nodes in the directed weighting network in the step 3), and determining a target monitoring variable;
step 5), taking the connection relation of the nodes extracted in the step 4) as input, training based on a GCN graph convolution neural network to obtain a community division result, and obtaining a monitoring variable set contained in a community where a target monitoring variable is located;
and 6) using the daily monitoring time sequence corresponding to the monitoring variable set obtained in the step 5) as input and the daily monitoring time sequence corresponding to the target monitoring variable as output, training the LSTM long-short term memory neural network, and realizing accurate prediction of the time sequence corresponding to the target monitoring variable based on the trained LSTM long-short term memory neural network.
Further, the sampling frequency of the monitoring time sequence obtained in the step 1) is set according to the sampling cost and the monitoring precision, the length of the sample is set, and a monitoring data set is obtained from historical data of the system operation process.
Further, the analysis of the coupled granger causal measure in the step 2) comprises the following steps:
step 21), obtaining monitoring data of time series X and Y
Figure BDA0002704692070000041
And
Figure BDA0002704692070000042
m and n represent the delay dimension;
step 22), the monitoring data obtained according to step 21)
Figure BDA0002704692070000043
And
Figure BDA0002704692070000044
calculating monitoring data based on kernel estimation method
Figure BDA0002704692070000045
And
Figure BDA0002704692070000046
edge distribution of
Figure BDA0002704692070000047
And
Figure BDA0002704692070000048
step 23) estimating empirical Copula conditional density based on the edge distribution calculated in step 22)
Figure BDA0002704692070000049
Then the kernel estimation method is used for carrying out on the conditional density of the empirical Copula
Figure BDA00027046920700000410
And
Figure BDA00027046920700000411
carrying out estimation;
step 24) optimizing the Copula estimation in step 23);
step 25) taking logarithm of the Copula function estimated in the step 24), calculating expectation of a sample, and obtaining a causal relationship between every two monitoring variables;
further, in the step 24), Bernstein is applied to approximate estimation of Copula conditional density for Copula estimation to optimize.
Further, in the CGC-coupled granger causal measure analysis method, X ═ X is set as X t And Y ═ Y t Respectively, the causal relationship of X → Y can be expressed as:
Figure BDA00027046920700000412
wherein f represents a conditional probability density,
Figure BDA00027046920700000413
and
Figure BDA00027046920700000414
history information of X and Y, m and n are hysteresis orders, and t represents a monitoring time. According to the original concept of GC, the left side of the equal sign indicates the influence of the history information of the time series y and x on the current value of the prediction y, and the right side of the equal sign indicates the influence of the past value of the time series y on the current value of the prediction y.
Further, in the CGC-coupled granger causal measure analysis method, a log likelihood ratio for measuring GC values is defined according to literature:
Figure BDA0002704692070000051
where E represents the expectation of the sample space.
Further, in the CGC-coupled Granger causal measure analysis method, a high-dimensional Copula is converted into a series of low-dimensional Copula by using a recursive representation, so that an efficient and easily-implemented Granger causal estimate can be provided, and causal relationships can be converted into:
Figure BDA0002704692070000052
where h represents the conditional joint density of the random variable (X, Y), and f and g represent the conditional edge densities of the random variables X and Y, respectively, according to Sklar's theory, and according to the density function c of the condition Copula, h can be further expanded to:
Figure BDA0002704692070000053
wherein,
Figure BDA0002704692070000054
and G represents the conditional edge distribution of Y and X, respectively, and further the simplified form of the X → Y causal relationship is:
Figure BDA0002704692070000055
further, the step 2) further comprises screening the coupling causal coefficient, and the specific method comprises introducing a section of Gaussian noise with moderate intensity, calculating the coupling correlation coefficient between the section of noise and each monitoring sequence to serve as the lower threshold of the coupling correlation coefficient among the monitoring variables, and eliminating the coupling correlation coefficient lower than the lower threshold.
Further, in step 3), it is required to perform feature transformation on the causal measurement relationship between the calculated monitoring variables, use a causal coefficient obtained by CGC coupling glargine causal measurement analysis as a weight of an edge connecting between the nodes, further characterize the causal coefficient as an adjacency matrix, and define the number of non-zero elements in the ith row and the ith column and the degree of the ith vertex of the adjacency matrix, where the network G ═ V, (E) is a network composed of | V | ═ N nodes and | E | ═ M edges, the degree is the number of neighboring nodes of a node, and represents the capability of establishing a direct connection between the node and surrounding nodes, and is represented as:
Figure BDA0002704692070000061
further, in the step 5), in the training process of the GCN graph convolution neural network, an edge weight connection relation of a target monitoring variable needs to be embedded;
further, in the step 6), the input of the training of the LSTM long-short term memory neural network is a non-target monitoring variable in the monitoring variable set.
Compared with the prior art, the invention has at least the following beneficial technical effects:
aiming at the problems that the existing neural network model has no specific learning mechanism and is insufficient in data structure characteristic information mining, the invention provides a time sequence combination prediction method for multivariate clustering and fusion from the angle of multivariate directional coupling and by combining the advantages of a graph convolution neural network and a long-term and short-term memory network. Firstly, researching the causal transfer relationship among variables based on coupled Glanduger causal measurement analysis; secondly, establishing a directed weighting network according to the causal analysis result of the variables, extracting the node and edge weight characteristics of the directed weighting network, embedding the weight of the target variable into a graph convolution neural network for training, and realizing the accurate classification of the monitored variables; and finally, taking a non-target monitoring variable time sequence contained in the community where the target monitoring variable is located as input, and predicting the target monitoring variable based on the long-term and short-term memory neural network. The method is verified by applying a monitoring sequence of a compressor unit in a chemical production system, and the result shows that: compared with the traditional classification method, the method has better performance in the aspects of prediction accuracy and computational complexity.
Furthermore, the community obtained by dividing the complex network reflects and reproduces the relationship among the components of the actual system, and has stronger robustness, so that the community still shows good effect even if abnormal state prediction is carried out.
Furthermore, the method can bring certain guiding significance to the service safety control of the actual complex electromechanical system.
Drawings
FIG. 1 is a flow chart of a multivariate clustering and fusion time series combined prediction method.
Fig. 2 is a schematic diagram of a directional weighting network.
Fig. 3 is a connection relationship diagram corresponding to the node e.
Fig. 4 is a schematic diagram of the connection of the main equipment of the compressor and the monitoring point location.
Fig. 5 is a compressor block directed weighting network graph.
Fig. 6 is a network characteristic diagram of the node 34.
FIG. 7 is a diagram of GCN community division results.
FIG. 8 is a comparison graph of the predicted value and the true value of the time series corresponding to the GCN method.
FIG. 9 is a diagram showing abnormal fluctuation of the first stage pressure ratio of the supercharger.
Fig. 10 is a graph showing abnormal fluctuation in displacement of the turbocharger shaft.
Fig. 11 is a diagram showing the system parameter to V2 actual output abnormal fluctuation.
Detailed Description
The invention is described in further detail below with reference to the accompanying drawings:
the invention provides a multivariate clustering and fused time sequence combined prediction method aiming at the problems that an existing neural network model has no specific learning mechanism and is insufficient in data structure characteristic information mining, from a multivariate directional coupling angle and combining the advantages of a graph convolution neural network and a long-term and short-term memory network. Firstly, researching the causal transfer relationship among variables based on coupled Glanzey causal measurement analysis; secondly, establishing a directed weighting network according to the causal analysis result of the variables, extracting the node and edge weight characteristics of the directed weighting network, embedding the weight of the target variable into a graph convolution neural network for training, and realizing the accurate classification of the monitored variables; and finally, taking a non-target monitoring variable time sequence contained in the community where the target monitoring variable is located as input, and predicting the target monitoring variable based on the long-term and short-term memory neural network. The method is verified by applying a monitoring sequence of a compressor unit in a chemical production system, and the result shows that: compared with the traditional classification method, the method has better performance in the aspects of prediction accuracy and computational complexity.
As shown in fig. 1, a multivariate clustering and fused time series combined prediction method includes the following steps:
step 1), according to data collected by a DCS, defining monitoring variables corresponding to monitoring point positions, determining a variable set of a monitoring target of a complex electromechanical system to be analyzed, selecting a monitoring variable value of each monitoring variable at a certain moment every day, and obtaining monitoring time sequences of the monitoring variables at the same moment corresponding to different continuous days;
step 2), performing CGC coupling Grangel causal measurement analysis on every two monitoring time sequences obtained in the step 1) to obtain the direction and the coupling coefficient of the coupling causal relationship;
step 3), establishing a directed weighting network model representing the bottom layer interaction dynamics of the system by taking the monitoring variables corresponding to the monitoring time sequence obtained in the step 2) as nodes, the coupling causal relationship as edges and the coupling coefficient as edges;
step 4), extracting the connection relation of the nodes in the directed weighting network in the step 3), and determining a target monitoring variable;
step 5), taking the connection relation of the nodes extracted in the step 4) as input, training based on a GCN graph convolution neural network to obtain a community division result, and obtaining a monitoring variable set contained in a community where a target monitoring variable is located;
and 6) using the daily monitoring time sequence corresponding to the monitoring variable set obtained in the step 5) as input and the daily monitoring time sequence corresponding to the target monitoring variable as output, training the LSTM long-short term memory neural network, and realizing accurate prediction of the time sequence corresponding to the target monitoring variable based on the trained LSTM long-short term memory neural network.
And setting the sampling frequency of a monitoring sequence at the same time corresponding to the monitoring variable on different continuous days according to the sampling cost and the monitoring precision, setting the length of a sample, and acquiring a monitoring data set from historical data in the operation process of the system, wherein the monitoring sequence consists of 08:00 monitoring data of each day of twelve continuous days.
Performing CGC coupling Glanduger causal measurement analysis on the obtained monitoring sequences pairwise, comprising the following steps:
(1) obtaining historical monitoring data of time series X and Y
Figure BDA0002704692070000091
And
Figure BDA0002704692070000092
m and n represent the delay dimension;
(2) calculating the edge distribution of the variable based on a kernel estimation method according to the variable monitoring data obtained in the step (1)
Figure BDA0002704692070000093
And
Figure BDA0002704692070000094
(3) estimating empirical Copula conditional density based on the edge distribution calculated in step (2)
Figure BDA0002704692070000095
The kernel method is used again for empirical Copula estimation;
(4) optimizing the Copula estimation in the step (3), and applying Bernstein to approximate estimation of the Copula density;
(5) taking logarithm of the Copula function estimated in the step (4), calculating expectation of a sample, and obtaining a causal relationship between every two monitoring variables;
further, in the CGC-coupled granger causal measure analysis method, a log likelihood ratio for measuring GC values is defined according to literature as:
Figure BDA0002704692070000096
where E represents the expectation of the sample space.
Further, in the CGC-coupled Granger causal measure analysis method, a high-dimensional Copula is converted into a series of low-dimensional Copula by using a recursive representation, thereby providing an efficient and easy-to-implement Granger causal estimate, and converting the causal relationship into:
Figure BDA0002704692070000097
wherein h represents the condition joint density of the random variable (X, Y), f and g represent the condition edge density of the random variable X and Y, respectively, and according to the Sklar theory, according to the density function c of the condition Copula, h is further expanded to:
Figure BDA0002704692070000101
wherein,
Figure BDA0002704692070000102
and G represents the conditional edge distribution of Y and X, respectively, and further the simplified form of the X → Y causal relationship is:
Figure BDA0002704692070000103
furthermore, a coupling causal coefficient obtained through CGC coupling Glange causal measurement analysis needs to be screened, and the specific method is to introduce a section of Gaussian noise with moderate intensity, calculate a coupling correlation coefficient between the section of Gaussian noise and each monitoring sequence, serve as a lower threshold limit of the coupling correlation coefficient among all monitoring variables, and reject the coupling correlation coefficient lower than the lower threshold limit.
Further, feature conversion needs to be performed on the causal measurement relationship between the calculated monitoring variables, as shown in table 1, a causal coefficient obtained by CGC causal measurement analysis is used as a weight of a connecting edge between nodes, and is further characterized as an adjacency matrix:
TABLE 1 edge-by-edge weight representation
Figure BDA0002704692070000104
Figure BDA0002704692070000105
Defining the number of nonzero elements in the ith row and the ith column of the adjacency matrix and the degree of the ith vertex, wherein the network G (V, E) is a network formed by | V | ═ N nodes and | E | ═ M edges, and the degree refers to the number of neighbor nodes of a node, embodies the capability of establishing direct contact between the node and the surrounding nodes and is expressed as follows:
Figure BDA0002704692070000111
a directed weighting network graph built from the adjacency matrix is shown in fig. 2.
Further, in the GCN graph convolution neural network training process, an edge weight connection relation of a target monitoring variable needs to be embedded, and if a node e is selected as the target monitoring variable, the connection relation between the node and the corresponding node is shown in fig. 3.
Further, the input of the graph convolution neural network during the test process also includes real-time monitoring data of the target monitoring variable.
The invention provides a multivariate clustering and fused time sequence combined prediction method aiming at the problems that an existing neural network model has no specific learning mechanism and is insufficient in data structure characteristic information mining, from a multivariate directional coupling angle and combining the advantages of a graph convolution neural network and a long-term and short-term memory network. Firstly, researching the causal transfer relationship among variables based on coupled Glanduger causal measurement analysis; secondly, establishing a directed weighting network according to the causal analysis result of the variables, extracting the node and edge weight characteristics of the directed weighting network, embedding the weight of the target variable into a graph convolution neural network for training, and realizing the accurate classification of the monitored variables; and finally, taking a non-target monitoring variable time sequence contained in the community where the target monitoring variable is located as input, and accurately predicting the target monitoring variable based on the long-term and short-term memory neural network.
Furthermore, the method is verified by applying actual service monitoring data of the compressor unit shown in fig. 4 in a certain chemical production system, and the result shows that compared with the traditional classification method, the method has better performance in the aspects of prediction accuracy and calculation complexity, and meanwhile, communities obtained by dividing a complex network reflect and reproduce the relationship among all components of the actual system, so that the method has stronger robustness, and therefore, even if abnormal state prediction is carried out, the method still has good effect.
Example (b):
example 37 variables from table 2 associated with the service of the compressor block primary equipment (where the 38 th variable is introduced gaussian noise) were selected, including equipment and process variables. And selecting the monitoring data of 12 days when the unit normally operates and the monitoring data of 1 day when the unit is in a fault state for subsequent complex network modeling and target monitoring variable time sequence prediction, wherein the sampling interval is 1 min.
Table 2 main monitoring points and description
Figure BDA0002704692070000121
Figure BDA0002704692070000131
The method comprises the following steps: coupled granger causal measure analysis of monitoring time series
Calculating a coupling cause-and-effect relationship between monitoring time sequences based on CGC coupling Glange cause-and-effect measurement analysis, introducing a Gaussian noise sequence as comparison of coupling cause-and-effect coefficients as a lower threshold of the coupling cause-and-effect coefficients, setting the coupling cause-and-effect coefficients larger than or equal to the lower threshold to be 0, obtaining an updated coupling cause-and-effect coefficient table, and further obtaining a coupling cause-and-effect relationship matrix.
Step two: directed weighting network model establishment and target monitoring variable determination
And establishing a directed weighting network model capable of representing the interaction dynamics of the bottom layer of the system by taking the monitoring variables in the multidimensional monitoring sequence as nodes, the coupling relationship as edges and the magnitude of the coupling coefficient as edges. And performing characteristic conversion on the causal measurement relation between the monitoring variables obtained by calculation, namely taking a causal coefficient obtained by CGC coupling Glange causal measurement analysis as the weight of a connecting edge between nodes, further characterizing the causal coefficient as an adjacency matrix, establishing a directed weighting network according to the adjacency matrix, determining the target monitoring variables, and extracting the edge-weight connection relation of the target monitoring variables.
Step three: complex network community division based on GCN graph convolution neural network
And (4) extracting nodes and connection relations of the complex network model established in the step two as the input of the GCN graph convolution neural network, embedding the edge-right connection relations of the target monitoring variables extracted in the step two in the training process of the GCN to obtain a community division result of the whole complex network, and determining a monitoring variable set contained in the community where the target monitoring variables are located.
Step four: time series prediction based on LSTM long-short term memory neural network
And taking the monitoring time sequence of the nodes in the community where the target monitoring variable is located obtained in the third step as an input, and predicting the community based on the LSTM long-short term memory neural network (decap is 0.1, batch _ size is 144, epochs is 500, and estimation _ split is 0.1, and Relu is taken as an implicit layer activation function).
1. Selection of system characteristic variables and description thereof
The variables used in this example and their descriptions are shown in Table 2. From the selected variables and the description thereof, the variables used in the example comprise the process variables and the equipment monitoring quantities, because the equipment monitoring quantities and the process variables are always associated to a certain degree, the process variables can reflect the service state of the equipment to a certain degree, the equipment monitoring variables can reflect the adjustment and fluctuation conditions of the process to a certain degree, and the association analysis of the whole system is formed by the variables.
2. Directed weighting network model establishment and target monitoring variable determination
The invention considers that the service state of each monitoring point of the compressor unit in one day presents non-stable characteristic, and the monitoring points correspond to stable service state at the same time every day in a period of time.
TABLE 3 monitoring variable causal relationship values
Figure BDA0002704692070000141
Therefore, data of each monitoring point at 12:00 hours per day under the continuous 12-day normal operation state is selected, after normalization processing, measurement relations between every two of 37 monitoring variables are analyzed based on a CGC coupling Langey causal method, meanwhile, a section of Gaussian noise sequence with moderate intensity is introduced as the 38 th variable, the causal relation between each monitoring variable and the CGC method is calculated and used as a lower threshold, and when the coupling causal coefficient is lower than the lower threshold, the correlation between the two variables is not strong and is not reserved. The causal relationship values obtained by CGC calculation and lower threshold screening are shown in table 3.
The correlation coefficients are converted into an adjacency matrix, monitoring point positions are taken as network nodes, correlation coefficients among monitoring variables are taken as edge connecting weights, and a complex network with the compressor set directed weighting is constructed based on Gephi and is shown in FIG. 5. The actual outputs of the monitoring 34, i.e., the system parameters to V2, are selected as target monitoring variables, and the network characteristics implied by the node are shown in fig. 6.
3. Complex network community division based on GCN graph convolution neural network
Training is performed based on an API-oriented unsupervised graph learning open source Python framework, in the training process, edge-right connection relations corresponding to the nodes 34 are embedded, and finally, three divided communities are obtained as shown in FIG. 7, and nodes and descriptions contained in the community where the nodes 34 are located are shown in Table 4.
TABLE 4 the Community in which the node 34 is located includes the node and its description (GCN)
Figure BDA0002704692070000151
4. Time series prediction based on LSTM long-short term memory neural network
Taking a monitoring time sequence corresponding to a target node 34 in a normal state of service for one day as output, taking monitoring time sequences corresponding to community containing nodes obtained by dividing the three methods as input, and predicting based on an LSTM network (decap ═ 0.1, batch _ size ═ 144, epochs ═ 500, and estimation _ split ═ 0.1, and all using Relu as an implicit layer activation function), so as to obtain a result as shown in fig. 8, wherein the root mean square error is only 0.0250, and the mean square error is 0.0006, and an accurate prediction target is achieved.
And then, 11-hour monitoring data in an abnormal state of the compressor unit are selected to predict a monitoring time sequence corresponding to the target node, wherein the root mean square error is 0.3520, and the mean square error is 0.1240, so that a more accurate prediction effect is achieved. The reason is that when the abnormity occurs, the heat exchanger at the outlet of the steam turbine of the compressor unit is scaled, so that the exhaust pressure is increased, the vacuum degree is reduced, the liquid level is controlled frequently, and the phenomena of heating of a supporting bearing component of a steam turbine rotor and the like are caused. As shown in fig. 9 to 11, abnormal fluctuations occur in some of the monitored variables, such as the intake air temperature at the stage of the turbocharger 1 (node 18), the turbocharger shaft displacement (node 20), and the like, and significant abnormal fluctuations also occur in the actual output to V2, which is the system parameter corresponding to the node 34. When the node 34 for representing the system operation state is predicted, the coupling relation between the complex network nodes and the connecting edges of the compressor unit is considered in the community structure obtained by the GCN method. From the community division result of the complex network, the community obtained by the GCN method includes two nodes with the maximum degree value in the network, namely the node 18 and the node 20, and it is known from the meaning of the degree value that the larger the node degree value is, the stronger the connection relationship with the surrounding nodes is, the stronger the interaction between the nodes in the community is, and the more representative the representation of the system operation state, and the smaller the effect of the variables represented by other nodes on the system parameters to the actual output of V2 is.
The effectiveness of the GCN community division method can be illustrated by comparing the physical structure and meaning of each variable in an actual system. This is mainly reflected in: the variable 3 can directly reflect the heat exchange performance of the steam turbine condenser, the variables 17-20 are main monitoring variables of the booster, the variables 22-24 correspond to all monitoring points of the steam temperature behind a main steam valve of the steam turbine, and the variables are main bases of feedback regulation and control performance in normal operation of a steam turbine compressor unit.
In conclusion, the invention provides a time sequence combination prediction method for multivariate clustering and fusion aiming at the problems that the existing neural network model has no specific learning mechanism and the data structure characteristic information is not sufficiently mined, and combining the advantages of a graph convolution neural network and a long-short term memory network from a multivariate directional coupling angle. Firstly, researching the causal transfer relationship among variables based on coupled Glanzey causal measurement analysis; secondly, establishing a directed weighting network according to the causal analysis result of the variables, extracting the node and edge weight characteristics of the directed weighting network, embedding the weight of the target variable into a graph convolution neural network for training, and realizing the accurate classification of the monitored variables; and finally, taking a non-target monitoring variable time sequence contained in the community where the target monitoring variable is located as input, and predicting the target monitoring variable based on the long-term and short-term memory neural network. The method is verified by applying a monitoring sequence of a compressor unit in a chemical production system, and the result shows that: the method of the invention has good performance in the aspects of prediction accuracy and calculation complexity, and the method can keep strong prediction capability in the abnormal state of the system.
The complex network is divided into communities, the relations among all components of the actual system are reflected and reproduced, and the method has stronger robustness, so that the method still shows good effect even if abnormal state prediction is carried out, and meanwhile, the method has certain guiding significance for the service safety control of the complex electromechanical system.

Claims (9)

1. A multivariate clustering and fused time series combined prediction method is characterized by comprising the following steps:
step 1), according to data collected by a DCS, defining monitoring variables corresponding to monitoring point positions, determining a variable set of a monitoring target of a complex electromechanical system to be analyzed, selecting a monitoring variable value of each monitoring variable at a certain moment every day, and obtaining monitoring time sequences of the monitoring variables at the same moment corresponding to different continuous days;
step 2), performing CGC coupling Grangel causal measurement analysis on every two monitoring time sequences obtained in the step 1) to obtain the direction and the coupling coefficient of the coupling causal relationship;
the coupled granger causal measure analysis in the step 2) comprises the following steps:
step 21), obtaining monitoring data of time series X and Y
Figure FDA0003645874970000011
And
Figure FDA0003645874970000012
m and n represent the delay dimension;
step 22), the monitoring data obtained according to step 21)
Figure FDA0003645874970000013
And
Figure FDA0003645874970000014
calculating monitoring data based on kernel estimation method
Figure FDA0003645874970000015
And
Figure FDA0003645874970000016
edge distribution of
Figure FDA0003645874970000017
And
Figure FDA0003645874970000018
step 23) estimating empirical Copula conditional density based on the edge distribution calculated in step 22)
Figure FDA0003645874970000019
Then the kernel estimation method is used for carrying out on the conditional density of the empirical Copula
Figure FDA00036458749700000110
And
Figure FDA00036458749700000111
carrying out estimation;
step 24) optimizing the Copula estimation in step 23);
step 25) taking logarithm of the Copula function estimated in the step 24), calculating expectation of a sample, and obtaining a causal relationship between every two monitoring variables;
step 3), establishing a directed weighting network model representing the bottom layer interaction dynamics of the system by taking the monitoring variables corresponding to the monitoring time sequence obtained in the step 2) as nodes, the coupling causal relationship as edges and the coupling coefficient as edges;
step 4), extracting the connection relation of the nodes in the directed weighting network in the step 3), and determining a target monitoring variable;
step 5), taking the connection relation of the nodes extracted in the step 4) as input, training based on a GCN graph convolution neural network to obtain a community division result, and obtaining a monitoring variable set contained in a community where a target monitoring variable is located;
and 6) using the daily monitoring time sequence corresponding to the monitoring variable set obtained in the step 5) as input and the daily monitoring time sequence corresponding to the target monitoring variable as output, training the LSTM long-short term memory neural network, and realizing accurate prediction of the time sequence corresponding to the target monitoring variable based on the trained LSTM long-short term memory neural network.
2. The multivariate clustering and fusion time series combination prediction method as claimed in claim 1, wherein the sampling frequency of the monitoring time series obtained in step 1) is set according to the sampling cost and the monitoring precision, the length of the sample is set, and the monitoring data set is obtained from the historical data of the system operation process.
3. The multivariate clustering and fusion time series combination prediction method according to claim 2, wherein in the CGC coupled Glangel causal measure analysis method, X-X t And Y ═ Y t Respectively, the causal relationship of X → Y can be expressed as:
Figure FDA0003645874970000021
wherein f represents a conditional probability density,
Figure FDA0003645874970000022
and
Figure FDA0003645874970000023
history information of X and Y, respectively, m and n are hysteresis order, tIndicating a monitoring moment; according to the original concept of GC, the left side of the equal sign indicates the influence of the history information of the time series y and x on the current value of the prediction y, and the right side of the equal sign indicates the influence of the past value of the time series y on the current value of the prediction y.
4. The multivariate clustering and fusion time series combination prediction method as claimed in claim 2, wherein in the CGC coupled Glange causal measure analysis method, a log likelihood ratio for measuring GC values is defined according to literature:
Figure FDA0003645874970000024
where E represents the expectation of the sample space.
5. The method as claimed in claim 2, wherein the CGC-coupled Granger causal measure analysis method uses recursive representation to transform a high-dimensional Copula into a series of low-dimensional Copula, which provides an efficient and easy-to-implement Granger causal estimation and transforms the causal relationship into:
Figure FDA0003645874970000031
where h represents the conditional joint density of the random variable (X, Y), and f and g represent the conditional edge densities of the random variables X and Y, respectively, according to Sklar's theory, and according to the density function c of the condition Copula, h can be further expanded to:
Figure FDA0003645874970000032
wherein,
Figure FDA0003645874970000033
f and G represent the conditional edge distributions of Y and X, respectively, and the simplified form of the X → Y causal relationship is obtained as:
Figure FDA0003645874970000034
6. the multivariate clustering and fused time series combined prediction method as claimed in claim 1, wherein the step 2) further comprises screening the coupling causal coefficient, specifically, introducing a section of Gaussian noise with moderate intensity, calculating the coupling correlation coefficient between the section of noise and each monitoring sequence as the lower threshold of the coupling correlation coefficient between each monitoring variable, and rejecting the coupling correlation coefficient lower than the lower threshold.
7. The multivariate clustering and fused time series combination prediction method according to claim 1, wherein in the step 3), the causal metric relationship between the calculated monitoring variables needs to be subjected to feature transformation, the causal coefficients obtained by CGC coupling glanged causal metric analysis are used as weights of the connecting edges between the nodes, and are further characterized as an adjacency matrix, and the number of non-zero elements in the ith row and ith column and the degree of the ith vertex of the adjacency matrix are defined, wherein the network G ═ V, E) is a network composed of | V | ═ N nodes and | E | ═ M edges, and the degree is the number of neighboring nodes of a node, which embodies the ability of establishing direct connection between the node and the surrounding nodes, and is expressed as:
K i ＝∑ j∈G δ ij ，
Figure FDA0003645874970000041
8. the multivariate clustering and fusion time series combination prediction method as claimed in claim 1, wherein in the step 5), an edge-weight connection relation of target monitoring variables needs to be embedded in the GCN graph convolution neural network training process.
9. The multivariate clustering and fused time series combined prediction method as claimed in claim 1, wherein in the step 6), the input of the training of the LSTM long-short term memory neural network is a non-target monitoring variable in the monitoring variable set.
CN202011034152.5A 2020-09-27 2020-09-27 Multivariate clustering and fused time sequence combined prediction method Active CN112101480B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011034152.5A CN112101480B (en) 2020-09-27 2020-09-27 Multivariate clustering and fused time sequence combined prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011034152.5A CN112101480B (en) 2020-09-27 2020-09-27 Multivariate clustering and fused time sequence combined prediction method

Publications (2)

Publication Number Publication Date
CN112101480A CN112101480A (en) 2020-12-18
CN112101480B true CN112101480B (en) 2022-08-05

Family

ID=73782338

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011034152.5A Active CN112101480B (en) 2020-09-27 2020-09-27 Multivariate clustering and fused time sequence combined prediction method

Country Status (1)

Country Link
CN (1) CN112101480B (en)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784918B (en) * 2021-02-01 2023-06-30 中国科学院自动化研究所 Node identification method, system and device based on unsupervised graph representation learning
CN113077094B (en) * 2021-04-13 2022-08-23 南京邮电大学 Ozone prediction method based on LSTM-GCN
CN113188968A (en) * 2021-05-06 2021-07-30 长沙理工大学 Air pollutant concentration prediction method and system based on combined deep learning model
CN113539382B (en) * 2021-05-28 2024-04-05 浙江中控技术股份有限公司 Early warning positioning method and system for key technological parameters of dimethyl phosphite
CN113422763B (en) * 2021-06-04 2022-10-25 桂林电子科技大学 Alarm correlation analysis method constructed based on attack scene
CN113327220B (en) * 2021-06-24 2023-06-02 浙江成功软件开发有限公司 Ocean multi-time sequence relevance discovery method based on complex network
CN113837358B (en) * 2021-08-25 2024-06-07 华润数字科技有限公司 System strategy prediction method based on Grangel causal relationship and related equipment
CN113907771A (en) * 2021-10-07 2022-01-11 北京工业大学 Method for identifying uterine contraction conduction direction and pacing area based on uterine electromyographic signals
CN114036736B (en) * 2021-11-05 2024-05-07 大连理工大学 Causal network learning method based on local Granges causal analysis
CN114372504B (en) * 2021-12-06 2024-09-10 燕山大学 Wind turbine generator system fault early warning method based on graph neural network
CN114693106A (en) * 2022-03-28 2022-07-01 国网山东省电力公司信息通信公司 Cloud edge mode-based data model construction method for information thing networking unification
CN115062686A (en) * 2022-05-07 2022-09-16 北京交通大学 Multi-KPI (Key performance indicator) time sequence abnormity detection method and system based on multi-angle features
CN114925811A (en) * 2022-05-16 2022-08-19 北京火山引擎科技有限公司 Multivariable time sequence processing method, device, equipment and medium
CN114997488A (en) * 2022-05-27 2022-09-02 中国石油大学(北京) Comprehensive energy system load prediction method and device
CN117454299B (en) * 2023-12-21 2024-03-26 深圳市研盛芯控电子技术有限公司 Abnormal node monitoring method and system

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993281A (en) * 2019-03-28 2019-07-09 北京工业大学 A kind of causality method for digging based on deep learning
CN110826686A (en) * 2018-08-07 2020-02-21 艾玛迪斯简易股份公司 Machine learning system and method with attribute sequence

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10037367B2 (en) * 2014-12-15 2018-07-31 Microsoft Technology Licensing, Llc Modeling actions, consequences and goal achievement from social media and other digital traces
CN108171142B (en) * 2017-12-26 2019-02-12 中南大学 A kind of causal method of key variables in determining complex industrial process
AU2019366977A1 (en) * 2018-10-24 2021-05-27 Marketspringpad Ip Usa Llc. Methods and systems for interactive data management
CN109407654B (en) * 2018-12-20 2020-08-04 浙江大学 Industrial data nonlinear causal analysis method based on sparse deep neural network
CN111008363B (en) * 2019-11-21 2021-11-19 西安交通大学 Multivariable causal-driven complex electromechanical system service safety situation evaluation method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110826686A (en) * 2018-08-07 2020-02-21 艾玛迪斯简易股份公司 Machine learning system and method with attribute sequence
CN109993281A (en) * 2019-03-28 2019-07-09 北京工业大学 A kind of causality method for digging based on deep learning

Also Published As

Publication number Publication date
CN112101480A (en) 2020-12-18

Similar Documents

Publication Publication Date Title
CN112101480B (en) Multivariate clustering and fused time sequence combined prediction method
CN116757534B (en) Intelligent refrigerator reliability analysis method based on neural training network
CN106644162B (en) Ring main unit wire core temperature soft measurement method based on neighborhood preserving embedding regression algorithm
CN114363195B (en) Network flow prediction and early warning method for time and frequency spectrum residual convolution network
CN114218872B (en) DBN-LSTM semi-supervised joint model-based residual service life prediction method
CN104865827B (en) Oil pumping unit oil extraction optimization method based on multi-working-condition model
CN112784920A (en) Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN114169091B (en) Method for establishing prediction model of residual life of engineering machinery part and prediction method
CN115169724A (en) Runoff prediction method based on space-time graph convolutional neural network
CN114777192B (en) Secondary network heat supply autonomous optimization regulation and control method based on data association and deep learning
CN111222992A (en) Stock price prediction method of long-short term memory neural network based on attention mechanism
CN116579505B (en) Electromechanical equipment cross-domain residual life prediction method and system without full life cycle sample
CN116956189A (en) Current abnormality detection system, method, electronic equipment and medium
CN112580693A (en) Petrochemical process fault diagnosis method based on self-help resampling neighborhood preserving embedding
CN115310536A (en) Reservoir water level prediction early warning method based on neural network and GCN deep learning model
CN116975645A (en) Industrial process soft measurement modeling method based on VAE-MRCNN
Yang et al. Teacher–Student Uncertainty Autoencoder for the Process-Relevant and Quality-Relevant Fault Detection in the Industrial Process
CN117667495B (en) Association rule and deep learning integrated application system fault prediction method
CN118426414A (en) Industrial data enhancement method based on self-attention variation self-encoder
CN113151842B (en) Method and device for determining conversion efficiency of wind-solar complementary water electrolysis hydrogen production
CN113609766A (en) Soft measurement method based on depth probability latent model
CN116894172A (en) Wind power short-term power prediction method based on hybrid prediction model
CN116842358A (en) Soft measurement modeling method based on multi-scale convolution and self-adaptive feature fusion
CN116129212A (en) Power load prediction method and system based on graph neural network
CN106444706A (en) Industrial process fault detection method based on data neighborhood feature preservation

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