CN109146246A - A kind of fault detection method based on autocoder and Bayesian network - Google Patents
A kind of fault detection method based on autocoder and Bayesian network Download PDFInfo
- Publication number
- CN109146246A CN109146246A CN201810796797.9A CN201810796797A CN109146246A CN 109146246 A CN109146246 A CN 109146246A CN 201810796797 A CN201810796797 A CN 201810796797A CN 109146246 A CN109146246 A CN 109146246A
- Authority
- CN
- China
- Prior art keywords
- model
- variable
- spe
- data set
- bayesian 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.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 49
- 238000000034 method Methods 0.000 claims abstract description 66
- 230000008569 process Effects 0.000 claims abstract description 43
- 238000001311 chemical methods and process Methods 0.000 claims abstract description 28
- 238000012549 training Methods 0.000 claims abstract description 23
- 238000012824 chemical production Methods 0.000 claims abstract description 19
- 239000000126 substance Substances 0.000 claims abstract description 16
- 238000010924 continuous production Methods 0.000 claims abstract description 9
- 239000011159 matrix material Substances 0.000 claims description 19
- 238000013480 data collection Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 17
- 230000002159 abnormal effect Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 6
- 239000000203 mixture Substances 0.000 claims description 6
- 238000012546 transfer Methods 0.000 claims description 6
- 230000005856 abnormality Effects 0.000 claims description 5
- 230000007787 long-term memory Effects 0.000 claims description 4
- 238000002945 steepest descent method Methods 0.000 claims description 4
- 125000002015 acyclic group Chemical group 0.000 claims 1
- 239000012141 concentrate Substances 0.000 claims 1
- 238000013507 mapping Methods 0.000 claims 1
- 238000003745 diagnosis Methods 0.000 abstract description 9
- 238000004519 manufacturing process Methods 0.000 abstract description 5
- 239000000284 extract Substances 0.000 abstract description 4
- 238000005312 nonlinear dynamic Methods 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 11
- 239000012071 phase Substances 0.000 description 7
- 238000009826 distribution Methods 0.000 description 6
- 238000004088 simulation Methods 0.000 description 6
- 238000000513 principal component analysis Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000001149 cognitive effect Effects 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000007619 statistical method Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000003889 chemical engineering Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000034994 death Effects 0.000 description 1
- 231100000517 death Toxicity 0.000 description 1
- 238000002405 diagnostic procedure Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000012467 final product Substances 0.000 description 1
- 239000007791 liquid phase Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000000746 purification Methods 0.000 description 1
- 239000000376 reactant Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
- G06F18/24155—Bayesian classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/04—Manufacturing
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Tourism & Hospitality (AREA)
- Data Mining & Analysis (AREA)
- General Business, Economics & Management (AREA)
- Development Economics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Educational Administration (AREA)
- Bioinformatics & Computational Biology (AREA)
- Game Theory and Decision Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Manufacturing & Machinery (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The present invention proposes a kind of fault detection method based on autocoder and Bayesian network, belongs to Fault Diagnosis for Chemical Process field.This method off-line phase chooses variable data building chemical process data set and sample data set from any chemical industry continuous production process;Utilize sample data set training autocoder model and Counting statistics amount T2 and SPE detection threshold value;Bayesian network is constructed using chemical process data set and estimates conditional probability;It at this stage, obtains real time data and inputs autocoder model and obtain corresponding estimated value, calculate the corresponding T2 and SPE value of input data and be simultaneously compared with detection threshold value: if meeting condition, chemical production process is normal;If not satisfied, then calculating the contribution degree of each variable, the root primordium of failure is found by Bayesian network.The present invention automatically extracts feature from the process data of Chemical Manufacture, is efficiently applied to nonlinear dynamic chemical process, realizes the detection and quick diagnosis of chemical process failure.
Description
Technical field
The present invention relates to Fault Diagnosis for Chemical Process fields, in particular to a kind of to be based on autocoder and Bayesian network
Fault detection method.
Background technique
Chemical production process is a complex process.In order to make maximization of economic benefit, chemical production process is often required that
" peace, steady, long, full, excellent ", it is desirable that process units can long period, smoothly run.As the technology of automatic control is constantly sent out
Exhibition, the control of device is tightly coupled with state variable, when a main variable is abnormal in chemical production process
When, these exceptions can be propagated by modes such as mass transfer, heat transfers in device, fluctuated so as to cause whole device, caused device
Alarm occurs to spread unchecked, operator is caused to be difficult to correctly diagnose failure.If operator makes because of error in judgement
The decision of mistake, may cause major accident, cause huge property loss, or even cause the injures and deaths of personnel.Therefore, quickly quasi-
Really failure is diagnosed, is of great significance for the safety in production of chemical industry.
In order to which assist operators cope with unusual service condition, the research in relation to fault diagnosis has nearly had 40 years history,
Existing method for diagnosing faults is broadly divided into 3 classes: the method based on qualitative model, quantitative model-based method and Kernel-based methods
The method of data-driven.And wherein the method for Kernel-based methods data-driven has showed apparent advantage, can effectively help
Operator diagnoses fault the reason of occurring.The method of existing Kernel-based methods data-driven is dynamic bayesian network method;
Dynamic bayesian network considers the correlation of variable in time, passes through the different variables under connection different moments, it is established that
Bayesian network with time response.However, the structural generation at Dynamic Bayesian network needs for complicated chemical-process
It devotes a tremendous amount of time and manpower, and dynamic bayesian network does not consider those unmeasured variables, and some failures
It is as caused by non-measured variable.
Failure diagnostic process is divided into fault detection stage and failure cause cognitive phase.In the fault detection stage, many
Person proposes principal component analysis (PCA), and the statistical methods such as Partial Least Squares (PLS) and Fei Sheer discriminant analysis (FDA) carry out
The fault detection of chemical process.These methods are capable of the feature of extraction process data, according to these characteristic use Principles of Statistics
Define statistic and hypothesis testing.In process of production, according to the data Counting statistics amount of monitoring, when statistic has been more than a certain
A threshold value, then it is assumed that faulty generation.Traditional PCA method is only suitable for linear model, does not reflect the non-linear of data
Feature.Core principle component analysis (KPCA) although the methods of can solve nonlinear problem, need to select suitable kernel function
And parameter, the selection of these parameters affect final detection effect.In addition to this, these methods are not all by chemical process
Dynamic property take into account.In recent years, researcher begins through the method for machine learning to carry out fault detection.
For failure cause cognitive phase, process data is decomposed into feature space and residual error space by the statistical methods such as PCA,
By calculating different variables for the contribution degree of statistic, to identify the reason of failure occurs, the maximum change of contribution degree
Amount is to cause failure that basic reason occurs.Since chemical process has the coupling of height, the exception of a variable be may cause
Its dependent variable is also abnormal, and obtaining reason from contribution degree merely may not be the basic reason that failure occurs.Therefore, it needs
Further to be inferred in result herein.
Autocoder is a kind of method of unsupervised learning.It has used back-propagation algorithm, is equal to output valve defeated
Enter value.Autocoder is made of two parts: encoder and decoder.Encoder by initial data be mapped to feature space into
Row compression, decoder restore compressed data.Autocoder can automatically extract feature, and be capable of handling non-linear ask
Topic.Therefore, autocoder is widely used in image recognition, natural language processing and mechanical disorder identification.Traditional is automatic
The encoder and decoder of encoder are generally all made of multilayer feedforward neural network, and there is no the expansions for considering dynamic time.
Bayesian network is a kind of probability net.It is the graphical network based on probability inference.One Bayesian network
It is a directed acyclic graph, is constituted by representing variable node and connecting the directed edge of these nodes.Node on behalf random become
Amount, and these directed edges represent the correlation between these nodes, are expressed with conditional probability and are connected by force between these variables
Degree.Bayesian network has obtained important application in the intellectualizing system of processing uncertain information, has been successfully used to cure
Treat the fields such as diagnosis, statistical decision, expert system, study prediction.
Summary of the invention
The present invention is directed to overcome the shortcomings of existing methods place, propose a kind of based on autocoder and Bayesian network
Fault detection method.The present invention combines autocoder model and Bayesian network, can cross number of passes from Chemical Manufacture
Automatically extract feature in, be efficiently applied to nonlinear dynamic chemical process, thus realize chemical process failure detection and
Quick diagnosis.
The present invention proposes a kind of fault detection method based on autocoder and Bayesian network, which is characterized in that point
For off-line phase and on-line stage, comprising the following steps:
1) off-line phase;
The data for 1-1) collecting chemical production process, construct sample data set;
Several variable data building chemical process data sets are chosen from any chemical industry continuous production process;From chemical process
One section of normal data is chosen in data set as sample forms sample data set;
Enabling X is sample data set, wherein xijRepresent the value at i-th of moment of j-th of variable, i=1,2 ... N, j=1,2 ...
N, n are to choose variable number, and N is moment sum, then X is N × n dimension, and expression formula is as follows:
1-2) sample data set that step 1-1) is obtained is standardized;
Corresponding average and standard deviation is sought to each column of X, then by variable data each in each column and the column pair
The average value answered makees difference and divided by the corresponding standard deviation of the column, is standardized to each variable data of X;
It 1-3) constructs autocoder model and is trained, obtain the autocoder model that training finishes;Specific step
It is rapid as follows:
1-3-1) construct an autocoder model, comprising: input layer, encoder layer and decoder layer, wherein encoding
Device layer and decoder layer are constituted by LSTMs layers of shot and long term memory network;The model is denoted as "current" model;
Mode input: the set X (k) of the data composition of row k is [x in note Xk1..., xkn], then autocoder model
Each group of input data be Xm(k)=[X (k), X (k-1) ..., X (k-m)], wherein k=m+1, m+2 ..., N, all
Input data constitutes input data set and is combined into Xtrain={ Xm(m+1), Xm(m+2) ..., Xm(N) }, input data set is combined into (N-m
+ 1) × m × n three-dimensional array;
The output of model: each group of output data of model is denoted as WhereinFor every group of input data X of correspondencem(k) estimated value;
1-3-2) by input data set XtrainEvery group of input data input "current" model, the input of each batch by batch
After "current" model is trained, according to loss function, pass through the parameter of steepest descent method more new model, loss function such as formula
(2) shown in:
Wherein, the input batch of model refers to the group number of iteration input data each time, is denoted as batch, 8≤batch≤
128;XmFor the input data set of each batch,For the corresponding output data set obtained by autocoder model
It closes;
When traversal finishes input data set XtrainLater, updating "current" model is Modelold, Model is calculatedold
Loss function be denoted as Lossold;
1-3-3) repeat step 1-3-2), input data set X is traversed againtrain, update "current" model and be denoted as
Modelnew, Model is calculatednewLoss function be denoted as Lossnew
1-3-4) training stops decision condition:
If Lossnew≤Lossold, then continue training pattern, by current ModelnewIt is updated to new Modelold, will be current
LossnewIt is updated to new Lossold, return to step 1-3-3);
If Lossnew≥LossoldWhen, then model deconditioning, exports current ModeloldOneself finished is trained as final
Dynamic encoder model, enters step 1-4);
1-4) calculate detection threshold value;
By input data set XtrainEvery group of input data by batch input step 1-3) obtained training finish from
The encoder layer of dynamic encoder model, model exports each batch input data set XmCharacter pair is H=[h1, h2...,
hi..., hN-m+1], wherein hiFor input data Xm(i) feature vector obtained by autocoder model;Remember the covariance of H
Matrix is S, then shown in the calculation expression such as formula (3) of the corresponding statistic T 2 of i-th of moment all variables:
T2i=hiS-1(hi)T (3)
Shown in the corresponding statistic SPE calculation expression of i-th of moment all variables such as formula (4):
Calculate separately input data set XtrainIn the corresponding T2 and SPE value of every group of input data, take T2 in calculated result
Maximum value as T2 detection threshold value, take the maximum value of SPE in calculated result as SPE detection threshold value, remember T2 detection threshold respectively
Value is T2thre, SPE detection threshold value is SPEthre;
It 1-5) constructs Bayesian network and estimates conditional probability;
1-5-1) the chemical process data set collected using step 1-1) uses transfer entropy algorithm and convergence intersection to map and calculate
Method calculates the causality between variable, obtains adjacency matrix, obtains process by pipeline meter flow chart and process flow chart
Knowledge, and adjacency matrix is modified;Bayesian network is constructed using revised adjacency matrix;If the Bayesian network of building
There are closed loops for network, then all rings are cut off, and be the variable creation creation dummy node at the disconnected place of ring, so that Bayesian network is one
Directed acyclic graph, Bayesian network building finish;
1-5-2) conditional probability estimate: the data of chemical process data set are input to step 1-5-1) establish Bayes
In network, conditional probability between the node of Bayesian network is obtained by maximal possibility estimation;
3) on-line stage;
2-1) data acquisition: the real-time of variable is chosen from the chemical industry continuous production process of real-time data base obtaining step 1-1)
Data form real time data collection, the length of real time data collection and one group of input data equal length of autocoder model;
2-2) repeat step 1-2), the step 2-1) real time data obtained is standardized;
2-3) the model for finishing the real time data collection input step 1-3 of step 2-1) acquisition) training, model output are real-time
The corresponding estimated value of data set;
Formula (3) and formula (4) 2-4) are utilized, calculates separately to obtain the corresponding statistic T 2 of real time data collectiondAnd SPEd;
The corresponding statistic T 2 of the real time data collection for 2-5) obtaining step 2-4)dAnd SPEdRespectively withAnd SPEthre
It is compared and determines chemical production process with the presence or absence of abnormal:
If two statistic Ts 2dAnd SPEdIt is below corresponding detection threshold value, then it represents that the chemical production process is normal, weight
New return step 2-1), continue to monitor;If any statistic is higher than its corresponding detection threshold value, which exists
It is abnormal, enter step 2-6);
2-6) to the real time data of step 2-1) each variable obtained to the actual value and change of the contribution degree variable of SPE
2 norms of the reconstruction value of amount are calculated, and each variable contribution degree size is calculated, by the size histogram of all contribution degrees into
Row is shown, obtains variable contribution degree figure;Using variable contribution plot, select to the maximum variable of statistic SPE contribution degree;
2-7), will will to be set as 100% to the state of the maximum variable of statistic SPE contribution degree in Bayesian network different
Often, Bayesian network is then updated;
2-8) updated Bayesian network is checked: if the maximum variable of contribution degree is exactly the root of Bayesian network
Node, then the variable is exactly the root primordium that failure occurs;If the maximum variable of contribution degree be not the root node of Bayesian network and
It is child node, then from the corresponding father node of the child node set off in search child node, is chosen from father node and be in abnormality
The node of maximum probability is the root primordium that failure occurs.
The features of the present invention and beneficial effect are:
Invention applies the autocoders of dynamic expansion to detect failure, can automatically extract the mistake in Chemical Manufacture
The feature of number of passes evidence avoids influence of the parameter selection to final detection effect.The autocoder of dynamic expansion uses length
The autocoder of phase memory network (LSTMs) building is suitable for Nonlinear Dynamic so that the feature extracted has time response
Chemical process.In addition, the invention combines autocoder and Bayesian network, meet the diagnosis in the case of different faults, helps
Help operator's quick diagnosis.
Detailed description of the invention
Fig. 1 is the overall flow block diagram of the method for the present invention.
Fig. 2 is the Tennessee Eastman model flow figure of the embodiment of the present invention.
Fig. 3 is the distribution schematic diagram of the statistic SPE and T2 of input data set in the present embodiment.
Fig. 4 is SPE the and T2 distribution schematic diagram in the case of the failure 1 of the embodiment of the present invention.
Fig. 5 is SPE the and T2 distribution schematic diagram in the case of the failure 2 of the embodiment of the present invention.
Fig. 6 is SPE the and T2 distribution schematic diagram in the case of the failure 6 of the embodiment of the present invention.
Fig. 7 is each variable in the case of the failure 1 of the embodiment of the present invention to the contribution degree schematic diagram of statistic SPE.
Fig. 8 is each variable in the case of the failure 2 of the embodiment of the present invention to the contribution degree schematic diagram of statistic SPE.
Fig. 9 is each variable in the case of the failure 6 of the embodiment of the present invention to the contribution degree schematic diagram of statistic SPE.
Figure 10 is that the Bayesian network of the embodiment of the present invention infers schematic diagram.
Specific embodiment
The present invention proposes a kind of fault detection method based on autocoder and Bayesian network, with reference to the accompanying drawing and
It is as follows that example further description is embodied.
The present invention proposes a kind of fault detection method based on autocoder and Bayesian network, be divided into off-line phase and
On-line stage, overall flow are as shown in Figure 1, comprising the following steps:
1) off-line phase;
The data for 1-1) collecting chemical production process, construct sample data set;
Several variable data building chemical process data sets, the selection of variable are chosen from any chemical industry continuous production process
It is selected according to specific chemical process;Wherein one section of appropriate length is chosen from chemical process data set, and (general length is choosing
Take 10 to 50 times of variable number) normal data (when normal data refers to chemical production device no exceptions even running
Data) as sample form sample data set.
Enabling X is sample data set, wherein xijRepresent the value at i-th of moment of j-th of variable, i=1,2 ... N, j=1,2 ...
N, n are to choose variable number, and N is moment sum, then X is N × n dimension, and expression formula is as follows:
Wherein, each column of matrix X represent a variable in the numerical value of different moments.
1-2) sample data set that step 1-1) is obtained is standardized;
The corresponding all sampled datas (i.e. each column of X) of each variable are concentrated to ask the sample data that step 1-1) is obtained
Corresponding average and standard deviation is taken, then by variable data each in each column average value work difference corresponding with the column and divided by this
Corresponding standard deviation is arranged, each variable data of X is standardized;
It 1-3) constructs autocoder model and is trained, obtain the autocoder model that training finishes;Specific step
It is rapid as follows:
An autocoder model, including input layer, encoder layer and decoder layer 1-3-1) are constructed, wherein encoder
Layer and decoder layer are constituted by LSTMs layers of shot and long term memory network, which is denoted as "current" model.
Mode input: the set X (k) of the data composition of row k is [x in note Xk1..., xkn], then autocoder model
One group of input data Xm(k)=[X (k), X (k-1) ..., X (k-m)], wherein k=m+1, m+2 ..., N, m represent often
The input time length of group data, all input datas constitute input data set and are combined into Xtrain={ Xm(m+1), Xm(m+
..., X 2)m(N) }, input data set is combined into the three-dimensional array of (N-m) × m × n.This three-dimensional array is generated by matrix X.
Wherein the value of m and the dynamic characteristic of system are related.The input batch (barch, 8≤batch≤128) of model, batch refers to
The group number of iteration input data each time.
Model output: the export structure of model is identical as input, and each group of output data is denoted as Wherein k=m+1, m+2 ..., N;WhereinIt is defeated for every group of correspondence
Enter data Xm(k) estimated value.
1-3-2) by input data set XtrainEvery group of input data input "current" model, the input of each batch by batch
After "current" model is trained, according to the loss function of definition, pass through the parameter of steepest descent method more new model.The damage of definition
It loses shown in function such as formula (2):
Wherein, XmFor the input data set of each batch,It is corresponding defeated to be obtained by autocoder model
Data acquisition system out.When model ergod finishes input data set XtrainLater, updating "current" model is Modelold, it is calculated
ModeloldLoss function Lossold;
1-3-3) repeat step 1-3-2), input data set X is traversed againtrain, update "current" model and be denoted as
Modelnew, Model is calculatednewLoss function be denoted as Lossnew
1-3-4) training stops decision condition:
If Lossnew≤Lossold, then illustrate that test error is also declining, continue training pattern, by current ModelnewMore
It is newly new Modelold, by current LossnewIt is updated to new Lossold, return to step 1-3-3);
If Lossnew≥LossoldWhen, then illustrate that test error is begun to ramp up, model deconditioning, output is currently
ModeloldAs the autocoder model that final training finishes, 1-4 is entered step);
1-4) estimate threshold value
Detecting the statistic used is T2 and SPE, by input data set XtrainEvery group of input data by batch input
Step 1-3) the autocoder model that finishes of obtained training.The encoder layer of model exports each batch input data set
XmCharacter pair is H=[h1, h2..., hi..., hN-m+1], wherein hiFor input data Xm(i) pass through autocoder model
Obtained feature vector (Xm(i) each element X (i) in, X (i-1) ..., X (i-m)] it is sequentially inputted to autocoder mould
In type, one-dimensional characteristic vector h is obtained using LSTMs networki.The covariance matrix for remembering H is S, (the Hotelling T2 statistics of statistic T 2
Amount) calculation formula such as formula (3) shown in:
T2i=hiS-1(hi)T (3)
Wherein, i indicates the i moment;
Shown in statistic SPE (square prediction error) calculation expression such as formula (4):
Calculate separately input data set XtrainIn the corresponding T2 and SPE value of every group of input data, take T2 in calculated result
Maximum value as T2 detection threshold value, take the maximum value of SPE in calculated result as SPE detection threshold value, remember T2 detection threshold respectively
Value is T2thre, SPE detection threshold value is SPEthre;
1-5) building Bayesian network constructs and estimates conditional probability;
1-5-1) the chemical process data set collected using step 1-1) uses transfer entropy algorithm and convergence intersection to map and calculate
Method calculates the causality between variable, obtains adjacency matrix, passes through pipeline meter flow chart (P&ID) and process flow chart
(PFD) procedural knowledge is obtained, and adjacency matrix is modified.Bayesian network is constructed using revised adjacency matrix.Cause
For Bayesian network directed acyclic graph, and closed loop controller and recycle stream are had in practical chemical industry continuous production process, because
This Bayesian network established, which will appear, the case where ring.Judge whether the Bayesian network established has closed loop.If there is closed loop, this
When need to handle these rings, cut off these rings, and the variable for the disconnected place of ring creates dummy node, guarantees Bayesian network
It is a directed acyclic graph.
1-5-2) conditional probability estimate: the data of chemical process data set are input to step 1-5-1) establish Bayes
In network, conditional probability between the node of Bayesian network is obtained by maximal possibility estimation.
2) on-line stage;
2-1) data acquisition: Jave Java DataBase Connection program (JDBC driver) (linker and in real time number are used
According to library), it is real-time from the real time data composition of the selection variable of the chemical industry continuous production process of real-time data base obtaining step 1-1)
Data set, (time span is m) to one group of input data equal length of length and autocoder model that real time data integrates;
2-2) repeat step 1-2), the step 2-1) real time data obtained is standardized;
2-3) the model for finishing the real time data collection input step 1-3 of step 2-1) acquisition) training, model output are real-time
The corresponding estimated value of data set;
Formula (3) and formula (4) 2-4) are utilized, calculates separately to obtain the corresponding statistic T 2 of real time data collectiondAnd SPEd。
The corresponding statistic T 2 of the real time data collection for 2-5) obtaining step 2-4)dAnd SPEdRespectively with T2threAnd SPEthre
It is compared, determines chemical production process with the presence or absence of abnormal:
If two statistic Ts 2dAnd SPEdIt is below corresponding detection threshold value, then it represents that the chemical production process is normal, weight
New return step 2-1) continue to monitor;If any statistic is higher than its corresponding detection threshold value, then it is assumed that the chemical production process
There are exceptions, enter step 2-6);
The actual value of variable can 2-6) be used the contribution degree of SPE the real time data of step 2-1) each variable obtained
It is calculated with 2 norms of the reconstruction value of variable, calculates each variable contribution degree size, by the size column of all contribution degrees
Figure is shown, and obtains variable contribution degree figure;Using variable contribution plot, select to the maximum variable of statistic SPE contribution degree;
2-7), will will to be set as 100% to the state of the maximum variable of statistic SPE contribution degree in Bayesian network different
Often, Bayesian network is then updated;
2-8) updated Bayesian network is checked: if the maximum variable of contribution degree is exactly Bayesian network
Root node, then the variable is exactly the root primordium that failure occurs;If the maximum variable of contribution degree is not the root of Bayesian network
Node but child node choose abnormal shape then from the corresponding father node of the child node set off in search child node from father node
The maximum node of state (i.e. node be in abnormality maximum probability node) is the root primordium that failure occurs.
A specific implementation example of the invention is as follows:
Fault diagnosis is carried out to a chemical engineering simulation model using the present invention, comprising the following steps:
Entire simulation model is divided into five operating units, respectively reactor, condenser, knockout drum, stripper and compression
Machine.Reactant, which enters in reactor, to be reacted, and the substance after reaction enters condenser, and condensed substance enters separator,
The unreacted substance in part is transported to reactor by compressor and re-starts reaction, and liquid phase part enters stripper progress purification
Separation, obtains final product.
The present embodiment proposes a kind of fault detection method based on autocoder and Bayesian network, including following step
It is rapid: 1) offline part
The data for 1-1) collecting chemical production process, construct sample data set;
For simulation model using Tennessee Eastman model, whole flow process is as shown in Figure 2.Entire model has altogether
52 variables, wherein 11 variables are performance variables, 22 variables are process monitoring variables, and 19 variables are concentration dependent prisons
Variable is controlled, since concentration variable acquisition time is inconsistent, so rejecting concentration variable in input data.Performance variable and process prison
It is as shown in Table 1 and Table 2 to control variable concrete meaning difference.In normal conditions, simulation model is run 25 hours, adopts one within every 3 minutes
A sample, then sample data set X is a matrix of (500,33).
1 performance variable table of table
2 process monitoring argument table of table
1-2) sample data set that step 1-1) is obtained is standardized;
Corresponding average and standard deviation is sought to each column of X, then by variable data each in each column and the column pair
The average value answered makees difference and divided by the corresponding standard deviation of the column, is standardized to each variable data of X;
It 1-3) constructs autocoder model and is trained, obtain the autocoder model that training finishes;Specific step
It is rapid as follows:
1-3-1) construct an autocoder model, comprising: input layer, encoder layer and decoder layer, wherein encoding
Device layer and decoder layer are constituted by LSTMs layers of shot and long term memory network;The model is denoted as "current" model;
Mode input: the set X (k) of the data composition of row k is [x in note Xk1..., xkn], then autocoder model
Each group of input data be Xm(k)=[X (k), X (k-1) ..., X (k-m)], wherein k=m+1, m+2 ..., N, all
Input data constitutes input data set and is combined into Xtrain={ Xm(m+1), Xm(m+2) ..., Xm(N) }, input data set is combined into (N-m
+ 1) × m × n three-dimensional array;
The output of model: each group of output data of model is denoted as WhereinFor every group of input data X of correspondencem(k) estimated value.
In the present embodiment, m is determined as 3, then input data set is combined into the three-dimensional array of (498,3,33).
Mode input batch is (batch=64), then the input data set size of each batch is (batch, m, 33),
The output data set size of each batch is similarly (batch, m, 33);
1-3-2) by input data set XtrainEvery group of input data input "current" model, the input of each batch by batch
After "current" model is trained, according to loss function, pass through the parameter of steepest descent method more new model, loss function such as formula
(2) shown in:
Wherein, the input batch of model refers to the group number of iteration input data each time, is denoted as batch, 8≤batch≤
128;XmFor the input data set of each batch,For the corresponding output data set obtained by autocoder model
It closes;
When traversal finishes input data set XtrainLater, updating "current" model is Modelold, Model is calculatedold
Loss function be denoted as Lossold;
1-3-3) repeat step 1-3-2), input data set X is traversed againtrain, update "current" model and be denoted as
Modelnew, Model is calculatednewLoss function be denoted as Lossnew
1-3-4) training stops decision condition:
If Lossnew≤Lossold, then continue training pattern, by current ModelnewIt is updated to new Modelold, will be current
LossnewIt is updated to new Lossold, return to step 1-3-3);
If Lossnew≥LossoldWhen, then model deconditioning, exports current ModeloldOneself finished is trained as final
Dynamic encoder model, enters step 1-4);
1-4) calculate detection threshold value;
By input data set XtrainEvery group of input data by batch input step 1-3) obtained training finish from
The encoder layer of dynamic encoder model, model exports each batch input data set XmCharacter pair is H=[h1, h2...,
hi..., hN-m+1], wherein hiFor input data Xm(i) feature vector obtained by autocoder model, in the present embodiment,
The feature that the output of encoder layer is tieed up for 17;The covariance matrix for remembering H is S, then the corresponding statistic T 2 of i-th of moment all variables
Calculation expression such as formula (3) shown in:
T2i=hiS-1(hi)T (3)
Shown in statistic SPE calculation expression such as formula (4):
Calculate separately input data set XtrainIn the corresponding T2 and SPE value of every group of input data, take T2 in calculated result
Maximum value as T2 detection threshold value, take the maximum value of SPE in calculated result as SPE detection threshold value, remember T2 detection threshold respectively
Value is T2thre, SPE detection threshold value is SPEthre;
Fig. 3 is the distribution schematic diagram of the statistic SPE and T2 of input data set in the present embodiment, wherein the horizontal seat of Fig. 3 (a)
It is designated as the time, ordinate is the value of statistic SPE;Fig. 3 (b) abscissa is the time, and ordinate is the value of statistic T 2.It calculates
To threshold value T2thre=51.14 and SPEthre=171.68
It 1-5) constructs Bayesian network and estimates conditional probability;
1-5-1) the chemical process data set collected using step 1-1) uses transfer entropy algorithm and convergence intersection to map and calculate
Method calculates the causality between variable, obtains adjacency matrix, obtains process by pipeline meter flow chart and process flow chart
Knowledge, and adjacency matrix is modified;Bayesian network is constructed using revised adjacency matrix;If the Bayesian network of building
There are closed loops for network, then all rings are cut off, and be the variable creation creation dummy node at the disconnected place of ring, so that Bayesian network is one
Directed acyclic graph, Bayesian network building finish;
1-5-2) conditional probability estimate: the data of chemical process data set are input to step 1-5-1) establish Bayes
In network, conditional probability between the node of Bayesian network is obtained by maximal possibility estimation;
2) on-line stage;
2-1) data acquisition: the real-time of variable is chosen from the chemical industry continuous production process of real-time data base obtaining step 1-1)
Data form real time data collection, the length of real time data collection and one group of input data equal length of autocoder model;
2-2) repeat step 1-2), the step 2-1) real time data obtained is standardized;
2-3) the model for finishing the real time data collection input step 1-3 of step 2-1) acquisition) training, model output are real-time
The corresponding estimated value of data set;
Formula (3) and formula (4) 2-4) are utilized, calculates separately to obtain the corresponding statistic T 2 of real time data collectiondAnd SPEd;
The corresponding statistic T 2 of the real time data collection for 2-5) obtaining step 2-4)dAnd SPEdRespectively with T2threAnd SPEthre
It is compared and determines chemical production process with the presence or absence of abnormal:
If two statistic Ts 2dAnd SPEdIt is below corresponding detection threshold value, then it represents that the chemical production process is normal, weight
New return step 2-1), continue to monitor;If any statistic is higher than its corresponding detection threshold value, which exists
It is abnormal, enter step 2-6);
Simulation process one in the present embodiment shares 21 kinds of failures, chooses three kinds of failures herein and is applied in the present invention, institute
It is as shown in table 3 to choose error listing.The each failure operation 48h of simulation model, wherein introducing failure in 8h.Data constantly export
Among the model established offline, the statistic T 2 and SPE being calculated, respectively at T2threAnd SPEthreIt is compared.
Fig. 4-6 is respectively failure 1, SPE the and T2 distribution schematic diagram of failure 2 and failure 6., wherein the dotted line in every figure
Represent corresponding detection threshold value;Wherein figure (a) abscissa in Fig. 4-6 is the time, and ordinate is the value of SPE statistic;Figure
Figure (b) abscissa in 4-6 is the time, and ordinate is the value of T2 statistic.These failures can be successfully detected out.
The fault type table of 3 the present embodiment of table
2-6) to the real time data of step 2-1) each variable obtained to the actual value and change of the contribution degree variable of SPE
2 norms of the reconstruction value of amount are calculated, and each variable contribution degree size is calculated, by the size histogram of all contribution degrees into
Row is shown, obtains variable contribution degree figure;Using variable contribution plot, select to the maximum variable of statistic SPE contribution degree;
In the present embodiment, Fig. 7-9 is respectively failure 1, the SPE variable contribution plot of failure 2 and failure 6, wherein abscissa generation
Table each variable, ordinate indicate each variable to the contribution degree of SPE.It can be seen that the SPE maximum contribution variable of failure 1 (IDV1)
SPE maximum contribution variable for XMEAS20, failure 2 (IDV2) is XMV 6, and the SPE maximum contribution variable of failure 6 (IDV6) is
XMEAS1
2-7) will will to be set as 100% to the state of the maximum variable of statistic SPE contribution degree in Bayesian network different
Often, Bayesian network is then updated;
By taking failure 1 as an example, XMEAS20 is the variable of SPE maximum contribution degree, sets XMEAS20 in Bayesian network to
100% is abnormal.In network such as Figure 10, Figure 10 after update each node on behalf variable, variable altogether there are three types of state it is normal
(normal), height (high), low (low), wherein high and low is all for abnormality, XMEAS 20 is simultaneously as can be seen from Figure 10
It is not at root node, so found by XMEAS 20 to father node, XMEAS13 (low, 69%), XMEAS11 (it is low,
66%), XMEAS9 (high, 76%), XMEAS7 (low, 68%), XMEAS (high, 60%), XMEAS1 (high, 78%) are high in bracket
With low expression variable state in which, and percentage then represents variable and is in the shape probability of state.And the maximum tribute of IDV6
Variable X MEAS1 is offered in the root node at Bayes network.
2-8) updated Bayesian network is checked: if the maximum variable of contribution degree is exactly the root of Bayesian network
Node, then the variable is exactly the root primordium that failure occurs;If the maximum variable of contribution degree be not the root node of Bayesian network and
It is child node, then from the corresponding father node of the child node set off in search child node, is chosen from father node and be in abnormality
The node of maximum probability is the root primordium that failure occurs.
In the present embodiment, it can be seen that the basic reason that IDV1 occurs is XMEAS1, and IDV6 breaks down basic original
Because being XMEAS1, final failure detection result is obtained.
Claims (1)
1. a kind of fault detection method based on autocoder and Bayesian network, which is characterized in that be divided into off-line phase and
On-line stage, comprising the following steps:
1) off-line phase;
The data for 1-1) collecting chemical production process, construct sample data set;
Several variable data building chemical process data sets are chosen from any chemical industry continuous production process;From chemical process data
It concentrates and chooses one section of normal data as sample composition sample data set;
Enabling X is sample data set, wherein xijRepresent the value at i-th of moment of j-th of variable, i=1,2 ... N, j=1,2 ... n, n
To choose variable number, N is moment sum, then X is N × n dimension, and expression formula is as follows:
1-2) sample data set that step 1-1) is obtained is standardized;
Corresponding average and standard deviation is sought to each column of X, it is then that variable data each in each column is corresponding with the column
Average value makees difference and divided by the corresponding standard deviation of the column, is standardized to each variable data of X;
It 1-3) constructs autocoder model and is trained, obtain the autocoder model that training finishes;Specific steps are such as
Under:
1-3-1) construct an autocoder model, comprising: input layer, encoder layer and decoder layer, wherein encoder layer
It is constituted with decoder layer by LSTMs layers of shot and long term memory network;The model is denoted as "current" model;
Mode input: the set X (k) of the data composition of row k is [x in note Xk1..., xkn], then autocoder model is every
One group of input data is Xm(k)=[X (k), X (k-1) ..., X (k-m)], wherein k=m+1, m+2 ..., N, all inputs
Data constitute input data set and are combined into Xtrain={ Xm(m+1), Xm(m+2) ..., Xm(N) }, input data set is combined into (N-m+1)
The three-dimensional array of × m × n;
The output of model: each group of output data of model is denoted asK=m
+ 1, m+2 ..., N;WhereinFor every group of input data X of correspondencem(k) estimated value;
1-3-2) by input data set XtrainEvery group of input data input "current" model by batch, the input of each batch is current
After model is trained, according to loss function, pass through the parameter of steepest descent method more new model, loss function such as formula (2) institute
Show:
Wherein, the input batch of model refers to the group number of iteration input data each time, is denoted as batch, 8≤batch≤128;XmFor
The input data set of each batch,For the corresponding output data set obtained by autocoder model;
When traversal finishes input data set XtrainLater, updating "current" model is Modelold, Model is calculatedoldDamage
It loses function and is denoted as Lossold;
1-3-3) repeat step 1-3-2), input data set X is traversed againtrain, update "current" model and be denoted as Modelnew, meter
Calculation obtains ModelnewLoss function be denoted as Lossnew
1-3-4) training stops decision condition:
If Lossnew≤Lossold, then continue training pattern, by current ModelnewIt is updated to new Modelold, will be current
LossnewIt is updated to new Lossold, return to step 1-3-3);
If Lossnew≥LossoldWhen, then model deconditioning, exports current ModeloldThe automatic volume finished as final training
Code device model, enters step 1-4);
1-4) calculate detection threshold value;
By input data set XtrainEvery group of input data by batch input step 1-3) the automatic volume that finishes of obtained training
The encoder layer of code device model, model exports each batch input data set XmCharacter pair is H=[h1, h2...,
hi..., hN-m+1], wherein hiFor input data Xm(i) feature vector obtained by autocoder model;Remember the covariance of H
Matrix is S, then shown in the calculation expression such as formula (3) of the corresponding statistic T 2 of i-th of moment all variables:
T2i=hiS-1(hi)T (3)
Shown in the corresponding statistic SPE calculation expression of i-th of moment all variables such as formula (4):
Calculate separately input data set XtrainIn the corresponding T2 and SPE value of every group of input data, take in calculated result T2 most
Big value is used as T2 detection threshold value, takes the maximum value of SPE in calculated result as SPE detection threshold value, remembers that T2 detection threshold value is respectively
T2thre, SPE detection threshold value is SPEthre;
It 1-5) constructs Bayesian network and estimates conditional probability;
1-5-1) the chemical process data set collected using step 1-1) uses transfer entropy algorithm and convergence to intersect mapping algorithm meter
The causality between variable is calculated, adjacency matrix is obtained, procedural knowledge is obtained by pipeline meter flow chart and process flow chart,
And adjacency matrix is modified;Bayesian network is constructed using revised adjacency matrix;If the Bayesian network of building is deposited
In closed loop, then all rings are cut off, and be the variable creation creation dummy node at the disconnected place of ring, so that Bayesian network is one oriented
Acyclic figure, Bayesian network building finish;
1-5-2) conditional probability estimate: the data of chemical process data set are input to step 1-5-1) establish Bayesian network
In, conditional probability between the node of Bayesian network is obtained by maximal possibility estimation;
2) on-line stage;
2-1) data acquisition: the real time data of variable is chosen from the chemical industry continuous production process of real-time data base obtaining step 1-1)
Form real time data collection, the length of real time data collection and one group of input data equal length of autocoder model;
2-2) repeat step 1-2), the step 2-1) real time data obtained is standardized;
2-3) model for finishing the real time data collection input step 1-3 of step 2-1) acquisition) training, model export real time data
Collect corresponding estimated value;
Formula (3) and formula (4) 2-4) are utilized, calculates separately to obtain the corresponding statistic T 2 of real time data collectiondAnd SPEd;
The corresponding statistic T 2 of the real time data collection for 2-5) obtaining step 2-4)dAnd SPEdRespectively with T2threAnd SPEthreIt carries out
Relatively and determine chemical production process with the presence or absence of abnormal:
If two statistic Ts 2dAnd SPEdIt is below corresponding detection threshold value, then it represents that the chemical production process is normal, returns again
Return step 2-1), continue to monitor;If any statistic is higher than its corresponding detection threshold value, there are different for the chemical production process
Often, 2-6 is entered step);
2-6) to the real time data of step 2-1) each variable obtained to the actual value of the contribution degree variable of SPE and variable
2 norms of reconstruction value are calculated, and calculate each variable contribution degree size, the size of all contribution degrees is opened up with histogram
Show, obtains variable contribution degree figure;Using variable contribution plot, select to the maximum variable of statistic SPE contribution degree;
2-7), 100% exception will will be set as to the state of the maximum variable of statistic SPE contribution degree in Bayesian network,
Then Bayesian network is updated;
2-8) updated Bayesian network is checked: if the maximum variable of contribution degree is exactly the root section of Bayesian network
Point, then the variable is exactly the root primordium that failure occurs;If the maximum variable of contribution degree be not the root node of Bayesian network but
Child node is chosen general in abnormality then from the corresponding father node of the child node set off in search child node from father node
The maximum node of rate is the root primordium that failure occurs.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2018104714530 | 2018-05-17 | ||
CN201810471453 | 2018-05-17 |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109146246A true CN109146246A (en) | 2019-01-04 |
CN109146246B CN109146246B (en) | 2021-06-04 |
Family
ID=64801037
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810796797.9A Active CN109146246B (en) | 2018-05-17 | 2018-07-19 | Fault detection method based on automatic encoder and Bayesian network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109146246B (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110321933A (en) * | 2019-06-11 | 2019-10-11 | 武汉闻道复兴智能科技有限责任公司 | A kind of fault recognition method and device based on deep learning |
CN110378036A (en) * | 2019-07-23 | 2019-10-25 | 沈阳天眼智云信息科技有限公司 | Fault Diagnosis for Chemical Process method based on transfer entropy |
CN111241688A (en) * | 2020-01-15 | 2020-06-05 | 北京百度网讯科技有限公司 | Method and device for monitoring composite production process |
CN111626099A (en) * | 2020-04-10 | 2020-09-04 | 浙江大学 | Industrial control system multi-loop oscillation causal relationship analysis method based on improved CCM |
CN112101439A (en) * | 2020-09-09 | 2020-12-18 | 浙江大学 | High-speed wire quality defect diagnosis and tracing method based on distributed Bayesian network |
CN112311662A (en) * | 2020-10-22 | 2021-02-02 | 南京富岛信息工程有限公司 | Intelligent oilfield edge computing gateway |
CN112367293A (en) * | 2020-10-10 | 2021-02-12 | 浙江大学 | Encrypted flow abnormity detection method based on time sequence characteristics |
CN112415894A (en) * | 2020-11-20 | 2021-02-26 | 中国矿业大学 | Active learning and BN (boron nitride) -based heavy medium coal preparation process safe operation control method |
CN112949947A (en) * | 2021-04-19 | 2021-06-11 | 广东电网有限责任公司计量中心 | Power failure early warning method and system based on convolution long-term and short-term memory network |
CN113671923A (en) * | 2021-08-30 | 2021-11-19 | 东北大学 | Continuous rolling process fault detection method based on data driving |
CN113722140A (en) * | 2021-08-30 | 2021-11-30 | 中国地质大学(武汉) | Industrial alarm flooding root cause diagnosis method based on small sample learning and storage medium |
CN116458119A (en) * | 2020-11-19 | 2023-07-18 | 日本电信电话株式会社 | Estimation device, estimation method, and estimation program |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160098633A1 (en) * | 2014-10-02 | 2016-04-07 | Nec Laboratories America, Inc. | Deep learning model for structured outputs with high-order interaction |
CN106528528A (en) * | 2016-10-18 | 2017-03-22 | 哈尔滨工业大学深圳研究生院 | A text emotion analysis method and device |
CN107169117A (en) * | 2017-05-25 | 2017-09-15 | 西安工业大学 | A kind of manual draw human motion search method based on autocoder and DTW |
CN107194433A (en) * | 2017-06-14 | 2017-09-22 | 电子科技大学 | A kind of Radar range profile's target identification method based on depth autoencoder network |
-
2018
- 2018-07-19 CN CN201810796797.9A patent/CN109146246B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160098633A1 (en) * | 2014-10-02 | 2016-04-07 | Nec Laboratories America, Inc. | Deep learning model for structured outputs with high-order interaction |
CN106528528A (en) * | 2016-10-18 | 2017-03-22 | 哈尔滨工业大学深圳研究生院 | A text emotion analysis method and device |
CN107169117A (en) * | 2017-05-25 | 2017-09-15 | 西安工业大学 | A kind of manual draw human motion search method based on autocoder and DTW |
CN107194433A (en) * | 2017-06-14 | 2017-09-22 | 电子科技大学 | A kind of Radar range profile's target identification method based on depth autoencoder network |
Non-Patent Citations (1)
Title |
---|
程非凡等: "基于收敛交叉映射的化工过程扰动因果分析方法", 《化工学报》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110321933A (en) * | 2019-06-11 | 2019-10-11 | 武汉闻道复兴智能科技有限责任公司 | A kind of fault recognition method and device based on deep learning |
CN110378036A (en) * | 2019-07-23 | 2019-10-25 | 沈阳天眼智云信息科技有限公司 | Fault Diagnosis for Chemical Process method based on transfer entropy |
CN111241688A (en) * | 2020-01-15 | 2020-06-05 | 北京百度网讯科技有限公司 | Method and device for monitoring composite production process |
CN111241688B (en) * | 2020-01-15 | 2023-08-25 | 北京百度网讯科技有限公司 | Method and device for monitoring composite production process |
CN111626099A (en) * | 2020-04-10 | 2020-09-04 | 浙江大学 | Industrial control system multi-loop oscillation causal relationship analysis method based on improved CCM |
CN112101439A (en) * | 2020-09-09 | 2020-12-18 | 浙江大学 | High-speed wire quality defect diagnosis and tracing method based on distributed Bayesian network |
CN112101439B (en) * | 2020-09-09 | 2023-11-28 | 浙江大学 | High-speed wire quality defect diagnosis and tracing method based on distributed Bayesian network |
CN112367293A (en) * | 2020-10-10 | 2021-02-12 | 浙江大学 | Encrypted flow abnormity detection method based on time sequence characteristics |
CN112311662B (en) * | 2020-10-22 | 2022-08-30 | 南京富岛信息工程有限公司 | Intelligent oilfield edge computing gateway |
CN112311662A (en) * | 2020-10-22 | 2021-02-02 | 南京富岛信息工程有限公司 | Intelligent oilfield edge computing gateway |
CN116458119A (en) * | 2020-11-19 | 2023-07-18 | 日本电信电话株式会社 | Estimation device, estimation method, and estimation program |
CN112415894A (en) * | 2020-11-20 | 2021-02-26 | 中国矿业大学 | Active learning and BN (boron nitride) -based heavy medium coal preparation process safe operation control method |
CN112949947A (en) * | 2021-04-19 | 2021-06-11 | 广东电网有限责任公司计量中心 | Power failure early warning method and system based on convolution long-term and short-term memory network |
CN113722140A (en) * | 2021-08-30 | 2021-11-30 | 中国地质大学(武汉) | Industrial alarm flooding root cause diagnosis method based on small sample learning and storage medium |
CN113671923A (en) * | 2021-08-30 | 2021-11-19 | 东北大学 | Continuous rolling process fault detection method based on data driving |
CN113722140B (en) * | 2021-08-30 | 2023-08-04 | 中国地质大学(武汉) | Industrial alarm flooding source diagnosis method based on small sample learning and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109146246B (en) | 2021-06-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109146246A (en) | A kind of fault detection method based on autocoder and Bayesian network | |
Jang et al. | Adversarial autoencoder based feature learning for fault detection in industrial processes | |
CN111337768A (en) | Deep parallel fault diagnosis method and system for dissolved gas in transformer oil | |
Razavi-Far et al. | Model-based fault detection and isolation of a steam generator using neuro-fuzzy networks | |
CN103914064B (en) | Based on the commercial run method for diagnosing faults that multi-categorizer and D-S evidence merge | |
CN108921230A (en) | Method for diagnosing faults based on class mean value core pivot element analysis and BP neural network | |
Tian et al. | Identification of abnormal conditions in high-dimensional chemical process based on feature selection and deep learning | |
CN105548764A (en) | Electric power equipment fault diagnosis method | |
CN111899905B (en) | Fault diagnosis method and system based on nuclear power device | |
Charbonnier et al. | A weighted dissimilarity index to isolate faults during alarm floods | |
CN113642754A (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
CN112904810B (en) | Process industry nonlinear process monitoring method based on effective feature selection | |
CN114297918A (en) | Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning | |
CHEN et al. | Fault diagnosis in chemical process based on self-organizing map integrated with fisher discriminant analysis | |
CN109298633A (en) | Chemical production process fault monitoring method based on adaptive piecemeal Non-negative Matrix Factorization | |
CN108388234A (en) | A kind of fault monitoring method dividing changeable gauge block pca model based on correlation | |
US20240085274A1 (en) | Hybrid bearing fault prognosis with fault detection and multiple model fusion | |
Liu et al. | Model fusion and multiscale feature learning for fault diagnosis of industrial processes | |
CN112132394A (en) | Power plant circulating water pump prediction state assessment method and system | |
Xu et al. | Global attention mechanism based deep learning for remaining useful life prediction of aero-engine | |
Tang et al. | Dual attention bidirectional generative adversarial network for dynamic uncertainty process monitoring and diagnosis | |
CN111241629B (en) | Intelligent prediction method for performance change trend of hydraulic pump of airplane based on data driving | |
CN114298220B (en) | Fault classification method based on context attention dynamic feature extractor | |
Zhang et al. | A flexible monitoring framework via dynamic-multilayer graph convolution network | |
CN114239402A (en) | Nuclear power circulating water pump fault diagnosis method and system based on optimized capsule network |
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 |