CN114519405B - Process industry multi-sensor data collaborative analysis method and system - Google Patents
Process industry multi-sensor data collaborative analysis method and system Download PDFInfo
- Publication number
- CN114519405B CN114519405B CN202210417451.XA CN202210417451A CN114519405B CN 114519405 B CN114519405 B CN 114519405B CN 202210417451 A CN202210417451 A CN 202210417451A CN 114519405 B CN114519405 B CN 114519405B
- Authority
- CN
- China
- Prior art keywords
- graph
- model
- node
- sensor
- hidden variable
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Abstract
The invention relates to a process industry multi-sensor data collaborative analysis method and system, and belongs to the technical field of data analysis. The flow industry multi-sensor data collaborative analysis method provided by the invention creatively introduces a directed graph structure with hidden variables aiming at three typical relations (causal relation, homologous relation and homogeneous relation) existing among monitoring point positions, and realizes the posterior estimation from sensor observation variables to hidden variables based on a hidden variable inference network model. And finally, learning a directed edge structure and a node value in the graph by using the graph neural network model, wherein the output characteristic vector can be used as characteristic representation of the current production condition and the equipment running state, and further, the performance of downstream applications such as prediction, control, abnormal monitoring and the like is effectively improved.
Description
Technical Field
The invention relates to the technical field of data analysis, in particular to a method and a system for collaborative analysis of multi-sensor data in process industry.
Background
In a process industrial scene, sensors for monitoring indexes such as flow, concentration, pressure, liquid level and the like are widely applied, and the data acquisition requirement of a process control system is supported. However, the existing collaborative analysis technology for multipoint timing sequence monitoring data is very deficient, and the intelligent system only realizes the analysis of a single monitoring point for sensor data. Meanwhile, the existing data analysis technology lacks a data analysis and characterization method combined with production flow and process prior knowledge, so that production state information cannot be fully mined. The incompleteness of data analysis in the production process can cause difficulty in accurately sensing the complete appearance of the production condition by an intelligent system, and the precision and the effect of downstream intelligent application are limited.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a flow industry multi-sensor data collaborative analysis method and system.
In order to achieve the purpose, the invention provides the following scheme:
a process industry multi-sensor data collaborative analysis method comprises the following steps:
introducing hidden variables according to the correlation among the multi-sensor point positions in the production system to construct a directed graph model; the correlation between the multi-sensor point positions comprises: causal, homologous, and homogeneous relationships;
assigning a structure vector to the nodes of the directed graph model;
training a hidden variable reasoning network by adopting a training data set to obtain a trained hidden variable reasoning network model; the input of the hidden variable reasoning network model is monitoring data of a sensor, and the output is a node observation value;
on the basis of the directed graph model, extending directed edges among the nodes to obtain a new directed graph model;
inputting the observed values of all nodes in the new directed graph model into a graph neural network so as to train the structure vector and the graph neural network parameters to obtain a graph neural network model; the input of the graph neural network model is a node observed value, and the output is a characteristic vector of a node;
acquiring monitoring data of multiple sensors at the current moment in a production system to be analyzed;
inputting the monitoring data into the hidden variable reasoning network model to obtain a node observation value at the current moment;
inputting the node observation value at the current moment into the graph neural network model to obtain a characteristic vector; the feature vector is used for representing observation point information corresponding to the current-time operation sensor.
Preferably, the introducing hidden variables according to the correlation among the multi-sensor point positions in the production system to construct a directed graph model specifically includes:
when the correlation among the multi-sensor point positions is a homogeneous relationship, respectively introducing a first hidden variable graph node and a second hidden variable graph node, and constructing a directed edge from the first hidden variable graph node to the second hidden variable graph node;
when the correlation among the multi-sensor point positions is a causal relationship, introducing a third hidden variable graph node, constructing a directed edge from a first sensor to the third hidden variable graph node, and constructing a directed edge from the third hidden variable graph node to a second sensor;
when the correlation among the multi-sensor point positions is the homologous relation, introducing a fourth hidden variable graph node, constructing a directed edge from the fourth hidden variable graph node to the first sensor, and constructing a directed edge from the fourth hidden variable graph node to the second sensor.
Preferably, the training of the hidden variable inference network by using the training data set to obtain the trained hidden variable inference network model specifically includes:
respectively constructing a generation model and a reasoning model based on a variational cyclic neural network model structure;
and training the parameters of the generated model and the parameters of the inference model by using a training data set by adopting a variational Bayesian method until preset training conditions are reached.
Preferably, on the basis of the directed graph model, extending the directed edges between the nodes to obtain a new directed graph model specifically includes:
determining the connection weight between any two graph nodes in the directed graph model;
determining a connectivity indicator based on the connection weight;
and when the connectivity index meets a preset condition, constructing a new directed edge in the two graph nodes to form a new directed graph model.
Preferably, the inputting the observed value of each node in the new directed graph model to the graph neural network to train the structure vector and the graph neural network parameters to obtain the graph neural network model specifically includes:
forming data characteristics of the nodes based on the observed values of the nodes;
extracting feature vectors of the nodes by using the data features of the nodes and the structural vectors according to the adjacency relation between the nodes in the new directed graph model;
predicting monitoring data of the sensor based on the extracted feature vectors;
calculating a training loss value of the graph neural network by using the mean square error based on the prediction data;
and determining the training loss value to the gradient of the model parameters and the structure vectors of the graph neural network until the training condition is met to obtain the graph neural network model.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the flow industry multi-sensor data collaborative analysis method provided by the invention creatively introduces a directed graph structure with hidden variables aiming at three typical relations (causal relation, homologous relation and homogeneous relation) existing among monitoring point positions, and realizes the posterior estimation from sensor observation variables to hidden variables based on a hidden variable inference network model. And finally, learning a directed edge structure and a node value in the graph by using the graph neural network model, wherein the output characteristic vector can be used as characteristic representation of the current production condition and the equipment running state, and further, the performance of downstream applications such as prediction, control, abnormal monitoring and the like is effectively improved.
Corresponding to the process industry multi-sensor data collaborative analysis method, the invention also provides a process industry multi-sensor data collaborative analysis system, which comprises:
the directed graph model building module is used for introducing hidden variables according to the correlation among the multi-sensor point positions in the production system to build a directed graph model; the correlation between the multi-sensor point positions comprises: causal, homologous, and homogeneous relationships;
a structure vector assigning module, configured to assign a structure vector to a node of the directed graph model;
the first model training module is used for training the hidden variable reasoning network by adopting a training data set to obtain a trained hidden variable reasoning network model; the input of the hidden variable reasoning network model is monitoring data of a sensor, and the output of the hidden variable reasoning network model is a node observed value;
the directed graph model updating module is used for expanding directed edges among the nodes to obtain a new directed graph model on the basis of the directed graph model;
the second model training module is used for inputting the observed values of all nodes in the new directed graph model into the graph neural network so as to train the structure vector and the graph neural network parameters to obtain the graph neural network model; the input of the graph neural network model is a node observed value, and the output is a characteristic vector of a node;
the monitoring data acquisition module is used for acquiring the monitoring data of the multiple sensors at the current moment in the production system to be analyzed;
the node observation value determining module is used for inputting the monitoring data into the hidden variable reasoning network model to obtain a node observation value at the current moment;
the characteristic vector determining module is used for inputting the node observed value at the current moment into the graph neural network model to obtain a characteristic vector; the feature vector is used for representing observation point information corresponding to the current-time operation sensor.
Preferably, the directed graph model building module includes:
the first directed edge construction unit is used for respectively introducing a first hidden variable graph node and a second hidden variable graph node when the correlation among the multi-sensor point positions is a homogeneous relationship, and constructing a directed edge from the first hidden variable graph node to the second hidden variable graph node;
the second directed edge construction unit is used for introducing a third hidden variable graph node when the correlation among the multi-sensor point positions is a causal relationship, constructing a directed edge from the first sensor to the third hidden variable graph node, and constructing a directed edge from the third hidden variable graph node to the second sensor;
and the third directed edge construction unit is used for introducing a fourth hidden variable graph node when the correlation among the multi-sensor point positions is the homologous relationship, constructing a directed edge from the fourth hidden variable graph node to the first sensor, and constructing a directed edge from the fourth hidden variable graph node to the second sensor.
Preferably, the first model training module comprises:
the model construction unit is used for respectively constructing a generation model and a reasoning model based on the variational cyclic neural network model structure;
and the first model training unit is used for training the parameters of the generated model and the parameters of the inference model by using a training data set by adopting a variational Bayesian method until preset training conditions are reached.
Preferably, the directed graph model updating module includes:
the connection weight determining unit is used for determining the connection weight between any two graph nodes in the directed graph model;
a connectivity index determination unit configured to determine a connectivity index based on the connection weight;
and the directed graph model updating unit is used for constructing a new directed edge in the two graph nodes to form a new directed graph model when the connectivity index meets a preset condition.
Preferably, the second model training module comprises:
a data feature generation unit configured to form a data feature of a node based on the observation value of the node;
a feature vector extraction unit, configured to extract feature vectors of nodes according to an adjacency relation between nodes in the new directed graph model by using data features of the nodes and the structure vectors;
the monitoring data prediction unit is used for predicting the monitoring data of the sensor based on the extracted feature vector;
a loss value training unit for calculating a training loss value of the graph neural network using the mean square error based on the prediction data;
and the second model training unit is used for determining the gradient of the training loss value to the model parameters and the structure vectors of the graph neural network until the training condition is met to obtain the graph neural network model.
The technical effect achieved by the process industry multi-sensor data collaborative analysis system provided by the invention is the same as that achieved by the process industry multi-sensor data collaborative analysis method provided by the invention, and therefore, the detailed description is omitted here.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
FIG. 1 is a flow chart of a process industry multi-sensor data collaborative analysis method provided by the present invention;
fig. 2 is a directed graph of association relationships between different sensor point locations according to an embodiment of the present invention; wherein, part (a) of fig. 2 is a directed graph in a causal relationship; FIG. 2 (b) is a directed graph in which the homologies are shown; FIG. 2 (c) is a directed graph in which the homologies are shown;
FIG. 3 is an overall framework diagram of a multi-sensor data collaborative analysis method according to an embodiment of the present invention;
fig. 4 is a schematic structural diagram of the process industry multi-sensor data collaborative analysis system provided by the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a flow industry multi-sensor data collaborative analysis method and system, which can effectively improve the performance of downstream applications such as prediction, control, anomaly detection and the like.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
As shown in fig. 1, the method for collaborative analysis of process industry multi-sensor data provided by the present invention includes:
and S1, introducing hidden variables according to the correlation among the multi-sensor point positions in the production system to construct a directed graph model. The correlation between the multi-sensor point positions comprises: causal relationships, homologous relationships, and homogeneous relationships. Wherein, the meaning of the homogeneous relation defined by the invention is as follows: when a material passes through a certain processing procedure, certain correlation exists between index monitoring quantities before and after processing, and the material often has a certain time delay characteristic. The causal relationship is as follows: the data of one point location directly influences and determines the change of the data of another point location, for example, the influence of the addition amount of the catalyst on the quality index of the finished product can be expressed as a causal relationship. The homologous relationship is as follows: for the same material at the same time, two monitoring technologies are adopted to measure different physical quantities or chemical quantities of the same material, for example, the flow rate and the concentration of the material are monitored simultaneously, and a certain implicit correlation exists between the two.
Based on the above-defined correlation between the multi-sensor point locations, the specific implementation process of step S1 may be:
s1-1, when the correlation among the multi-sensor point positions is a homogeneous relation, respectively introducing a first hidden variable graph node and a second hidden variable graph node, and constructing a directed edge between the first hidden variable graph node and the second hidden variable graph node. For example, for sensors in homogenous relationa 1、a 2Respectively introducing hidden variable graph nodesz 1、z 2.. Wherein, in the first placetTime of day, tojMonitoring value of individual sensora j (t) The corresponding hidden variable vector is expressed asz j (t)，z j (t)∈R k. From the point of view of the time domain,z j (t) Will transmit information toz j (t+1), and is therefore constructed in the figureBuild fromz j (t) Toz j (t+1) directed edges. Furthermore, in the order of processing and monitoring of the substances, the first assumption isjThe material monitoring information under each process step isa j (t) The process product flows intoj+1 steps, monitoring data area j+1(t) At this time, an implicit variable is assumedz j (t) Will also pass information toz j (t+1), and is therefore also constructed in the figurez j (t) Toz j (t+1) as shown in part (c) of fig. 2.
S1-2, when the correlation among the multi-sensor point positions is causal, introducing a third hidden variable graph node, constructing a directed edge from the first sensor to the third hidden variable graph node, and constructing a directed edge from the third hidden variable graph node to the second sensor. For example, suppose thattAt the moment, the monitoring values of the two sensors are respectivelya(t) Andb(t) The two have a causal relationship, i.e.a(t) Determining or influencingb(t) A change in (c). Introducing hidden variable at the momentz(t)，z(t)∈R kConstructed froma(t) Toz(t) And construct a secondaryz(t) Tob(t) As shown in part (a) of fig. 2. Representsa(t) Information decision ofz(t) In turn, influenceb(t) Distribution of (2).
S1-3, when the correlation among the multi-sensor point positions is the homologous relation, introducing a fourth hidden variable graph node, constructing a directed edge from the fourth hidden variable graph node to the first sensor, and constructing a directed edge from the fourth hidden variable graph node to the second sensor. For example, suppose thattAt the moment, the monitoring values of the two sensors are respectivelya(t) Andb(t) The two have a homologous relationship, i.e.a(t) Andb(t) The method is a monitoring numerical value obtained by adopting different monitoring means for the same material. Introducing hidden variable at the momentz(t)，z(t)∈R kWhile constructing fromz(t) Toa(t) Directed edges andz(t) Tob(t) Has a directed edge. Representa(t) Information decision ofz(t) In turn, influenceb(t) Distribution of (2).
And S2, giving a structure vector to the nodes of the directed graph model. The structure vector is represented as: v. ofi，
And initialized to a random vector. Wherein the content of the first and second substances,Mis the sum of the hidden variable node and the observation variable node in the graph. Each structure vector is optimized simultaneously during the subsequent network training process.
And S3, training the hidden variable reasoning network by adopting the training data set to obtain a trained hidden variable reasoning network model. The input of the hidden variable reasoning network model is the monitoring data of the sensor, and the output is the node observed value.
The specific implementation process of step S3 is as follows:
s3-1, respectively constructing a generation model and an inference model based on the variational recurrent neural network model structure.
And S3-2, training the parameters of the generated model and the parameters of the inference model by using a training data set by adopting a variational Bayes method until preset training conditions are reached.
For example, using training set StrainAnd training the hidden variable reasoning network. Oriented to three kinds of incidence relations, from StrainRandomly extracting the sensor data sequence s belonging to the association relation w1:=[s1，...，s w ]Wherein, in the step (A),wrepresenting the window size of the data sequence, each term s thereintRepresents the firsttAnd associating the monitoring data of the sensor at the moment. Constructing a reasoning model for given observation data s w1:And (5) estimating the value of the hidden variable node in the lower estimation graph, and training a reasoning model. According to the three types of association relations, model construction and training processes in each relation are as follows:
first, for anyA sensor having a homogeneous relationship, assuming a number of sensors in homogeneous relationshipKThus, it is firsttTime monitoring data set st，st∈R K Will also express it
. Based on the variational cyclic neural network model structure, the time range of 1:wis as followsKAn implicit variable
AndKan observed variable
Generating a model of
And reasoning model
：
For generative models, it is of the form:
wherein the content of the first and second substances,
in order to be a model of a recurrent neural network,θas a network parameter, d w1:Is a deterministic hidden variable in a recurrent neural network, the firsttThe hidden step variable is dt。d0A hidden variable is initialized for the network.
And
the respective tables carry a multivariate gaussian distribution of the diagonal covariance matrix.
For inference models, can be givenMonitoring data sequence s w1:In the case of (2), a hidden variable is estimated
The posterior distribution of (a), in the form of:
wherein
Representing a multivariate gaussian distribution with a diagonal covariance matrix,
the recurrent neural network model is described above. For each probability distribution, a fully-connected neural network is used to estimate the mean and covariance matrix diagonal vectors of the distribution.
And then a variational Bayes method is adopted, and a training set S is utilizedtrainTo latent variable reasoning model
And generating the model
Parameter (2) ofθ、φTraining is carried out, and the training target is the lower limit of the likelihood of the maximum sensor observation data:
wherein, the first and the second end of the pipe are connected with each other,
representative pair distribution
In the hope of expectation,
represents the posterior distribution of the hidden variable,
represents the a prior distribution of the hidden variables,
representing the reconstructed distribution of the observed variable.
For any one of the homology relations, the sensor monitoring data is
And
hidden variables are
Which generates a model of
. The reasoning model is
Representing given monitoring data
And
next, the estimated hidden variables
Posterior distribution of (2). Both are modeled by adopting a neural network and trained by adopting a variational Bayes method
And
training objective to maximize sensor observations
And
lower limit of evidence of (c):
wherein the content of the first and second substances,
and the prior distribution representing the hidden variable is defined as standard Gaussian distribution and does not contain the parameter to be learned.
Represents the posterior distribution of the hidden variables,
representing the reconstructed distribution of the observed variable.
For any cause and effect relationship, the sensor monitoring data are respectively
And
hidden variable is ztAnd adopting a conditional variation encoder model to model the correlation between the hidden variable and the observed variable. The generation process comprises the following steps:
wherein the content of the first and second substances,
is representative of a given
Underlying variable ztThe conditional prior distribution is in the form of a multivariate gaussian distribution with a diagonal covariance matrix, and a neural network is used for modeling.
To be at a known hidden variable ztDependent variable of
The prediction distribution of (1) is in the form of a multivariate gaussian distribution with a diagonal covariance matrix, and is modeled by a neural network.
For hidden variable ztA posteriori reasoning model
For multivariate gaussian distribution with diagonal covariance matrix, neural networks are used for modeling.
Using training sets
Training by using variational Bayes method
、
Parameter (2)
Training objective to maximize sensor observations
Lower limit of likelihood evidence of (1):
And S4, on the basis of the directed graph model, expanding directed edges among the nodes to obtain a new directed graph model. The specific implementation process of the step comprises the following steps:
and S4-1, determining the connection weight between any two graph nodes in the directed graph model. For example, for any two graph nodes, the connection weight between the nodes is defined according to the structure vector as follows:
and S4-2, determining the connectivity index based on the connection weight.
And S4-3, when the connectivity index meets the preset condition, constructing a new directed edge in the two graph nodes to form a new directed graph model. Here, the preset conditions are:
wherein, define
TopK represents the maximum K elements selected from the set, and K can be freely adjusted in practical application according to the upper limit of the calculation load and the coupling degree of the production system.
And S5, inputting the observed value of each node in the new directed graph model to the graph neural network so as to train the structure vector and the graph neural network parameters to obtain the graph neural network model. The input of the graph neural network model is a node observed value, and the output is a feature vector of the node. The specific implementation process of the step is as follows:
s5-1, forming data bits of nodes based on observed values of the nodesAnd (5) performing characterization. For example, for each sensor node
According to the window sizewDefining the sequence monitoring data as follows:
further, the univariate monitoring data is expanded tokDimension, obtaining a feature representation of the observed data:
wherein the content of the first and second substances,
are trainable parameters. And (4) inferring the size of the window by using the inference model constructed aiming at the correlation relations of different point positionswAnd (3) posterior distribution of hidden variable nodes at each time and sampling hidden variable sequences
，
And is and
. The observation node data and the hidden variable node data jointly form the data characteristics of all nodes in the directed graph model
：
S5-2, according to the adjacent relation between the nodes in the new directed graph model, using the data characteristics of the nodesExtracting feature vectors of the nodes from the structural vectors; the feature vector is represented as
：
Wherein, RELU represents the activation function of the network,
in order to train the parameters, the user may,
representative nodeiThe set of all nodes connected with the edge,
representative nodejTo nodeiImportance weight of. The calculation method is as follows:
wherein exp is an exponential function,
the matching degree between the nodes i and j is represented by the following calculation mode:
in which a function is activated
Comprises the following steps:
And S5-3, predicting the monitoring data of the sensor based on the extracted feature vector. The monitoring data is as follows:
wherein the content of the first and second substances,
representing vector dot multiplication.
And S5-4, calculating the training loss value of the neural network of the graph by using the average square error based on the prediction data. The training loss values are:
wherein, the first and the second end of the pipe are connected with each other,
representing the two-norm of the vector,Trepresenting the total length of the input sequence and the predicted sequence.
S5-5, determining the training loss value to the gradient of the model parameter and the structure vector of the graph neural network until the training condition is met to obtain the graph neural network model. For example, a training loss value is calculated
For model parameters
And a structural vector
And training the gradient.
Wherein the content of the first and second substances,
in order to obtain the learning rate of the learning,
each represents
To pair
Of the gradient of (a).
And S6, acquiring the monitoring data of the multiple sensors at the current moment in the production system to be analyzed.
And S7, inputting the monitoring data into the hidden variable reasoning network model to obtain the node observed value at the current moment.
And S8, inputting the node observation value at the current moment into the graph neural network model to obtain a feature vector. The feature vector is used for representing observation point information corresponding to the operation sensor at the current moment and can be used as the input of a system control, prediction and anomaly monitoring model based on deep learning.
Specifically, after the monitoring data of the multiple sensors at the current moment is acquired, each sensor node is subjected to monitoring
Defining window size for data analysiswGiven its sequence monitoring data as
By usingThe inference model obtains the characteristic representation of each node in the directed graph model, including observation nodes
. According to the node structure vector V M1:And point location correlation, and constructing edges in the directed graph model. Inference node characteristics of neural network of graph obtained by training
。
The implementation architecture of the whole data processing is shown in fig. 3.
Based on the above, compared with the prior art, the invention has the following advantages:
1. the multi-sensor data collaborative analysis method provided by the invention can combine the process correlation among the sensors to more fully represent the current production condition and the equipment running state, and particularly, the correlation among the flow industry multi-sensor point positions is represented by a directed graph model, and the edge set in the graph is subjected to self-adaptive optimization in a training mode. Compared with a stacking multidimensional input form adopted in a traditional data analysis method, the potential relation between observation data can be more easily captured by using the input based on the graph structure under the guidance of production prior information by using the feature extraction model, and the current production condition and the equipment running state are more fully represented.
2. According to the method, the relevance relation of three typical monitoring point positions can be modeled by introducing hidden variable nodes into the graph, so that the priori knowledge among the monitoring point positions is introduced into the model conveniently. Specifically, aiming at the correlation between three typical monitoring point locations in the process industry, the method adopts a directed graph model with hidden variables to represent the correlation, so that the prior information in production can be skillfully integrated into a graph structure, the input information of the graph neural network comprises the original observation data and the hidden variable data closely related to the production process, and the graph neural network model can better and fully represent the current production condition and the equipment running state. Meanwhile, the invention introduces three variational self-coding machine models for reasoning posterior distribution of hidden variables in the aspect of three different incidence relations, and the model training and reasoning process is simple, efficient and easy to realize.
In addition, corresponding to the above-mentioned process industry multi-sensor data collaborative analysis method, the present invention also provides a process industry multi-sensor data collaborative analysis system, as shown in fig. 4, the system includes: the system comprises a directed graph model building module 1, a structure vector endowing module 2, a first model training module 3, a directed graph model updating module 4, a second model training module 5, a monitoring data acquisition module 6, a node observation value determining module 7 and a characteristic vector determining module 8.
The directed graph model building module 1 is used for building a directed graph model by introducing hidden variables according to the correlation among the multi-sensor point positions in the production system. The correlation between the multi-sensor point positions comprises: causal relationships, homologous relationships, and homogeneous relationships.
The structure vector assigning module 2 is used for assigning structure vectors to the nodes of the directed graph model.
The first model training module 3 is used for training the hidden variable reasoning network by adopting a training data set to obtain a trained hidden variable reasoning network model. The input of the hidden variable reasoning network model is the monitoring data of the sensor, and the output is the node observed value.
The directed graph model updating module 4 is used for expanding directed edges between nodes to obtain a new directed graph model on the basis of the directed graph model.
The second model training module 5 is configured to input the observed values of each node in the new directed graph model to the graph neural network, so as to train the structure vector and the graph neural network parameters to obtain the graph neural network model. The input of the graph neural network model is node observed values, and the output is a feature vector of the nodes.
The monitoring data acquisition module 6 is used for acquiring the monitoring data of the multiple sensors at the current moment in the production system to be analyzed.
The node observation value determining module 7 is configured to input the monitoring data into the hidden variable reasoning network model to obtain a node observation value at the current time.
The feature vector determining module 8 is configured to input the node observation value at the current time to the graph neural network model to obtain a feature vector. The feature vector is used for representing observation point information corresponding to the operation sensor at the current moment.
The directed graph model building module 1 includes: the device comprises a first directed edge construction unit, a second directed edge construction unit and a third directed edge construction unit.
The first directed edge construction unit is used for respectively introducing a first hidden variable graph node and a second hidden variable graph node when the correlation among the multi-sensor point positions is a homogeneous relationship, and constructing a directed edge from the first hidden variable graph node to the second hidden variable graph node.
The second directed edge construction unit is used for introducing a third hidden variable graph node when the correlation among the multi-sensor point positions is causal, constructing directed edges from the first sensor to the third hidden variable graph node, and constructing directed edges from the third hidden variable graph node to the second sensor.
The third directed edge construction unit is used for introducing a fourth hidden variable graph node when the correlation among the multi-sensor point positions is the homologous relation, constructing a directed edge from the fourth hidden variable graph node to the first sensor, and constructing a directed edge from the fourth hidden variable graph node to the second sensor.
The first model training module 3 includes: the device comprises a model building unit and a first model training unit.
The model construction unit is used for respectively constructing a generation model and a reasoning model based on the variational cyclic neural network model structure.
The first model training unit is used for training the parameters of the generated model and the parameters of the inference model by using a variational Bayesian method and using a training data set until preset training conditions are reached.
The directed graph model update module 4 includes: the device comprises a connection weight determining unit, a connectivity index determining unit and a directed graph model updating unit.
The connection weight determining unit is used for determining the connection weight between any two graph nodes in the directed graph model.
The connectivity index determination unit is used for determining the connectivity index based on the connection weight.
And the directed graph model updating unit is used for constructing a new directed edge in the two graph nodes to form a new directed graph model when the connectivity index meets a preset condition.
The second model training module 5 includes: the device comprises a data feature generation unit, a feature vector extraction unit, a monitoring data prediction unit, a loss value training unit and a second model training unit.
The data feature generation unit is used for forming data features of the nodes based on the observed values of the nodes.
The feature vector extraction unit is used for extracting feature vectors of the nodes by using the data features and the structural vectors of the nodes according to the adjacency relation among the nodes in the new directed graph model;
and the monitoring data prediction unit is used for predicting the monitoring data of the sensor based on the extracted feature vector.
The loss value training unit is used for calculating a training loss value of the graph neural network by using the average square error based on the prediction data.
And the second model training unit is used for determining the gradient of the training loss value to the model parameters and the structure vectors of the graph neural network until the training condition is met to obtain the graph neural network model.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the description of the method part.
The principle and the embodiment of the present invention are explained by applying specific examples, and the above description of the embodiments is only used to help understanding the method and the core idea of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the foregoing, the description is not to be taken in a limiting sense.
Claims (10)
1. A process industry multi-sensor data collaborative analysis method is characterized by comprising the following steps:
introducing hidden variables according to the correlation among the multi-sensor point positions in the production system to construct a directed graph model; the correlation between the multi-sensor point positions comprises: causal, homologous, and homogeneous relationships;
assigning a structure vector to the nodes of the directed graph model;
training a hidden variable reasoning network by adopting a training data set to obtain a trained hidden variable reasoning network model; the input of the hidden variable reasoning network model is monitoring data of a sensor, and the output is a node observation value;
on the basis of the directed graph model, extending directed edges among the nodes to obtain a new directed graph model;
inputting the observed values of all nodes in the new directed graph model into a graph neural network so as to train the structure vector and the graph neural network parameters to obtain a graph neural network model; the input of the graph neural network model is a node observed value, and the output is a characteristic vector of a node;
acquiring monitoring data of multiple sensors at the current moment in a production system to be analyzed;
inputting the monitoring data into the hidden variable reasoning network model to obtain a node observation value at the current moment;
inputting the node observation value at the current moment into the graph neural network model to obtain a characteristic vector; the feature vector is used for representing observation point information corresponding to the current-time operation sensor.
2. The process industry multi-sensor data collaborative analysis method according to claim 1, wherein the introducing of hidden variables according to the correlation between multi-sensor point locations in a production system to construct a directed graph model specifically comprises:
when the correlation among the multi-sensor point positions is a homogeneous relationship, respectively introducing a first hidden variable graph node and a second hidden variable graph node, and constructing a directed edge from the first hidden variable graph node to the second hidden variable graph node;
when the correlation among the multi-sensor point positions is a causal relationship, introducing a third hidden variable graph node, constructing a directed edge from a first sensor to the third hidden variable graph node, and constructing a directed edge from the third hidden variable graph node to a second sensor;
when the correlation among the multi-sensor point positions is the homologous relation, introducing a fourth hidden variable graph node, constructing a directed edge from the fourth hidden variable graph node to the first sensor, and constructing a directed edge from the fourth hidden variable graph node to the second sensor.
3. The process industry multi-sensor data collaborative analysis method according to claim 1, wherein the training of the hidden variable inference network with the training data set to obtain the trained hidden variable inference network model specifically comprises:
respectively constructing a generation model and a reasoning model based on a variational cyclic neural network model structure;
and training the parameters of the generated model and the parameters of the inference model by using a training data set by adopting a variational Bayesian method until preset training conditions are reached.
4. The process industry multi-sensor data collaborative analysis method according to claim 1, wherein the expanding of the directed edges between the nodes to obtain a new directed graph model based on the directed graph model specifically comprises:
determining the connection weight between any two graph nodes in the directed graph model;
determining a connectivity indicator based on the connection weight;
and when the connectivity index meets a preset condition, constructing a new directed edge in the two graph nodes to form a new directed graph model.
5. The process industry multi-sensor data collaborative analysis method according to claim 1, wherein the inputting of the observed values of each node in the new directed graph model to a graph neural network to train the structure vector and the graph neural network parameters to obtain the graph neural network model specifically includes:
forming data characteristics of the nodes based on the observed values of the nodes;
extracting feature vectors of the nodes by using the data features of the nodes and the structural vectors according to the adjacency relation between the nodes in the new directed graph model;
predicting monitoring data of the sensor based on the extracted feature vectors;
calculating a training loss value of the graph neural network using the mean square error based on the prediction data;
and determining the training loss value to the gradient of the model parameters and the structure vectors of the graph neural network until the training condition is met to obtain the graph neural network model.
6. A process industry multi-sensor data collaborative analysis system, comprising:
the directed graph model building module is used for introducing hidden variables according to the correlation among the multi-sensor point positions in the production system to build a directed graph model; the correlation between the multi-sensor point positions comprises: causal, homologous, and homogeneous relationships;
a structure vector assigning module, configured to assign a structure vector to a node of the directed graph model;
the first model training module is used for training the hidden variable reasoning network by adopting a training data set to obtain a trained hidden variable reasoning network model; the input of the hidden variable reasoning network model is monitoring data of a sensor, and the output is a node observation value;
the directed graph model updating module is used for expanding directed edges among the nodes to obtain a new directed graph model on the basis of the directed graph model;
the second model training module is used for inputting the observed values of all nodes in the new directed graph model into the graph neural network so as to train the structure vector and the graph neural network parameters to obtain the graph neural network model; the input of the graph neural network model is a node observed value, and the output is a characteristic vector of a node;
the monitoring data acquisition module is used for acquiring the monitoring data of the multiple sensors at the current moment in the production system to be analyzed;
the node observation value determining module is used for inputting the monitoring data into the hidden variable reasoning network model to obtain a node observation value at the current moment;
the characteristic vector determining module is used for inputting the node observed value at the current moment into the graph neural network model to obtain a characteristic vector; the feature vector is used for representing observation point information corresponding to the operation sensor at the current moment.
7. The process industry multi-sensor data collaborative analysis system according to claim 6, wherein the directed graph model building module includes:
the first directed edge construction unit is used for respectively introducing a first hidden variable graph node and a second hidden variable graph node when the correlation among the multi-sensor point positions is a homogeneous relationship, and constructing a directed edge from the first hidden variable graph node to the second hidden variable graph node;
the second directed edge construction unit is used for introducing a third hidden variable graph node when the correlation among the multi-sensor point positions is a causal relationship, constructing a directed edge from the first sensor to the third hidden variable graph node, and constructing a directed edge from the third hidden variable graph node to the second sensor;
and the third directed edge construction unit is used for introducing a fourth hidden variable graph node when the correlation among the multi-sensor point positions is the homologous relation, constructing a directed edge from the fourth hidden variable graph node to the first sensor, and constructing a directed edge from the fourth hidden variable graph node to the second sensor.
8. The process industry multi-sensor data collaborative analysis system according to claim 6, wherein the first model training module includes:
the model construction unit is used for respectively constructing a generation model and a reasoning model based on the variational cyclic neural network model structure;
and the first model training unit is used for training the parameters of the generated model and the parameters of the inference model by using a training data set by adopting a variational Bayesian method until preset training conditions are reached.
9. The process industry multi-sensor data collaborative analysis system according to claim 6, wherein the directed graph model update module includes:
the connection weight determining unit is used for determining the connection weight between any two graph nodes in the directed graph model;
a connectivity index determination unit configured to determine a connectivity index based on the connection weight;
and the directed graph model updating unit is used for constructing a new directed edge in the two graph nodes to form a new directed graph model when the connectivity index meets a preset condition.
10. The process industry multi-sensor data collaborative analysis system according to claim 6, wherein the second model training module includes:
a data feature generation unit configured to form a data feature of a node based on the observation value of the node;
a feature vector extraction unit, configured to extract feature vectors of nodes according to adjacency relations between nodes in the new directed graph model by using the data features of the nodes and the structural vectors;
the monitoring data prediction unit is used for predicting the monitoring data of the sensor based on the extracted feature vector;
a loss value training unit for calculating a training loss value of the graph neural network using a mean square error based on prediction data;
and the second model training unit is used for determining the training loss value to the gradient of the model parameters and the structure vectors of the graph neural network until the training condition is met to obtain the graph neural network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210417451.XA CN114519405B (en) | 2022-04-21 | 2022-04-21 | Process industry multi-sensor data collaborative analysis method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210417451.XA CN114519405B (en) | 2022-04-21 | 2022-04-21 | Process industry multi-sensor data collaborative analysis method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114519405A CN114519405A (en) | 2022-05-20 |
| CN114519405B true CN114519405B (en) | 2022-07-12 |
Family
ID=81600407
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210417451.XA Active CN114519405B (en) | 2022-04-21 | 2022-04-21 | Process industry multi-sensor data collaborative analysis method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114519405B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116663725A (en) * | 2023-05-30 | 2023-08-29 | 北京网藤科技有限公司 | Intelligent process flow prediction optimization method and system for industrial Internet |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079931A (en) * | 2019-12-27 | 2020-04-28 | 浙江大学 | State space probabilistic multi-time-series prediction method based on graph neural network |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB201821197D0 (en) * | 2018-12-24 | 2019-02-06 | Native Design Ltd | Framework for handling sensor data in a smart home system |
| CN111650894A (en) * | 2020-04-02 | 2020-09-11 | 中国计量大学 | Bayesian network complex industrial process soft measurement method based on hidden variables |
| CN114167922B (en) * | 2021-11-22 | 2022-07-22 | 郑州宝诺电子科技有限公司 | Farming and pasturing intelligent analysis method and system based on multi-sensor data acquisition |
-
2022
- 2022-04-21 CN CN202210417451.XA patent/CN114519405B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079931A (en) * | 2019-12-27 | 2020-04-28 | 浙江大学 | State space probabilistic multi-time-series prediction method based on graph neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114519405A (en) | 2022-05-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Huang et al. | A generalized growing and pruning RBF (GGAP-RBF) neural network for function approximation | |
| Jalalifar et al. | Application of the adaptive neuro-fuzzy inference system for prediction of a rock engineering classification system | |
| Zhang et al. | Learning structural changes of Gaussian graphical models in controlled experiments | |
| Jin et al. | Multiple model based LPV soft sensor development with irregular/missing process output measurement | |
| Liu et al. | Cooperative stabilization of a class of LTI plants with distributed observers | |
| Wang et al. | Adaptive fuzzy decentralized control for a class of large-scale stochastic nonlinear systems | |
| Han et al. | Hierarchical extreme learning machine for feedforward neural network | |
| Guo et al. | Robust echo state networks based on correntropy induced loss function | |
| CN114519405B (en) | Process industry multi-sensor data collaborative analysis method and system | |
| CN111046961A (en) | Fault classification method based on bidirectional long-and-short-term memory unit and capsule network | |
| Olatunji et al. | Modeling permeability and PVT properties of oil and gas reservoir using hybrid model based on type-2 fuzzy logic systems | |
| Shi et al. | Membrane fouling diagnosis of membrane components based on multi-feature information fusion | |
| Money et al. | Sparse online learning with kernels using random features for estimating nonlinear dynamic graphs | |
| Lopez-Ramos et al. | Dynamic network identification from non-stationary vector autoregressive time series | |
| CN108170648B (en) | non-Gaussian process monitoring method based on known data regression | |
| Li et al. | Towards the maintenance of population diversity: A hybrid probabilistic model building genetic network programming | |
| Ye et al. | Improved SVD algorithm based on Slope One | |
| Pekaslan et al. | Noise parameter estimation for non-singleton fuzzy logic systems | |
| Luo et al. | Passivity analysis and passification of uncertain Markovian jump systems with partially known transition rates and mode-dependent interval time-varying delays | |
| US20230394304A1 (en) | Method and Apparatus for Neural Network Based on Energy-Based Latent Variable Models | |
| Liu et al. | STA-APSNFIS: STA-optimized adaptive pre-sparse neuro-fuzzy inference system for online soft sensor modeling | |
| Liu et al. | Echo state networks with double-reservoir for time-series prediction | |
| Chakraborty et al. | Brain-Inspired Spiking Neural Network for Online Unsupervised Time Series Prediction | |
| CN111027611A (en) | Fuzzy PLS modeling method based on dynamic Bayesian network | |
| CN108427867A (en) | One kind being based on Grey BP Neural Network interactions between protein Relationship Prediction method |
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 |