CN109190702A - Sparse Bayesian network incremental learning method based on continuous type industrial data - Google Patents
Sparse Bayesian network incremental learning method based on continuous type industrial data Download PDFInfo
- Publication number
- CN109190702A CN109190702A CN201811009985.9A CN201811009985A CN109190702A CN 109190702 A CN109190702 A CN 109190702A CN 201811009985 A CN201811009985 A CN 201811009985A CN 109190702 A CN109190702 A CN 109190702A
- Authority
- CN
- China
- Prior art keywords
- sample
- bayesian network
- data
- regularization
- continuous type
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 239000011159 matrix material Substances 0.000 claims abstract description 43
- 239000000203 mixture Substances 0.000 claims abstract description 12
- 238000012549 training Methods 0.000 claims abstract description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- 230000002159 abnormal effect Effects 0.000 claims description 6
- 238000000354 decomposition reaction Methods 0.000 claims description 2
- 238000009826 distribution Methods 0.000 claims description 2
- 230000001105 regulatory effect Effects 0.000 claims description 2
- 238000005516 engineering process Methods 0.000 description 6
- 238000005457 optimization Methods 0.000 description 5
- 239000003795 chemical substances by application Substances 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000001276 controlling effect Effects 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 235000015170 shellfish Nutrition 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 239000003245 coal Substances 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000010438 heat treatment Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 230000007257 malfunction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Algebra (AREA)
- Evolutionary Biology (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention discloses a kind of sparse Bayesian network incremental learning method based on continuous type industrial data.The sparse Bayesian network incremental learning method based on continuous type industrial data includes: step 1, data prediction: training sample is subjected to the pretreatment such as regularization, guarantee mean value be 0, variance 1；Step 2, Structure learning: being based on sliding window, and the incremental learning of structure is carried out in the form of lot sample sheet；Step 3, parameter learning: the coefficient of relationship matrix and regularization sample obtained using Structure learning obtains the mean value and variance of the Gaussian mixtures of each feature under current state, i.e. continuous type Bayesian network parameters；Step 4, network updates: to the sample newly to arrive, being pre-processed using the regularization parameter of population sample, forms the update of this progress of the lot sample of same scale structure and parameter using sliding window.
Description
Technical field
The invention belongs to computer application technologies, are specifically related to a kind of sparse shellfish based on continuous type industrial data
This network incremental learning method of leaf.
Background technique
With the development of China's industrial technology, the industrial equipment of more and more modernizations is put into actual production, is
State and society creates huge economic value.But with the update of industrial equipment, although having liberated a large amount of artificial
Labour, but also the monitoring for a whole set of industrial flow brings no small difficulty.Although currently realizing many large scale industries to set
Standby automation, but the normal operation of equipment constantly regulates and controls there is still a need for people.Due to the complexity of equipment, work as equipment
There is unit exception situation, needs to learn operational process by heart and just can be carried out fault tracing in the heart, veteran worker.Therefore logical
Complicated industrial equipment data are crossed, from study in data to the dependence equipment various components inside, and to abnormal number
Precise positioning is carried out according to source, is the direction highly studied.
Bayesian network is the completely model about variable and relationship between variables, general between variable by establishing
Rate model provides the dependence between variable and excavates implicit causality from data, therefore can be used to answer each
The probability problem of kind relevant variable.Also, Bayesian network, which possesses, extremely strong to be had logic, intelligibly handles uncertain problem
Ability, be also based on uncertain information and make inferences, therefore it has a wide range of applications in the field of data mining.
One Bayesian network mainly includes two parts: network structure and network parameter, and wherein network structure is one
Directed acyclic graph indicates the dependence between variable, and network parameter indicates probability distribution of the variable relative to its father node.
There is the method for comparative maturity for the Bayesian Network Learning of discrete variable, common method has the side based on constraint
Method and method based on score function carry out greedy or heuristic search to map space on this basis, acquisition most
Parameter learning is carried out in excellent structure, safeguards a conditional probability table for each variable.But these methods are more for feature
Data need a large amount of time overhead.And in industrial data, the parameter of an equipment is often more, and data Chang Weilian
Ideotype data.Meanwhile for certain special installations, the dependence between device parameter often becomes with time change
Change.Based on the above fact, the Bayesian network incremental learning method based on continuous data that we have proposed a kind of.
Summary of the invention
Above-mentioned the deficiencies in the prior art are directed to, the purpose of the present invention is to provide a kind of based on continuous type industrial data
Sparse Bayesian network incremental learning method, it is intended to which current operating conditions are established according to large industry equipment running state parameter
Dependence and dependent probability model between lower parameters, the causality between intuitive presentation parameter, and can be into
One step carries out anomaly source retrospect to abnormal data, effectively reduces the regulation manually for device parameter.
To reach above-mentioned purpose, the present invention adopts the following technical scheme: a kind of sparse shellfish based on continuous type industrial data
This network incremental learning method of leaf includes: step 1, and data prediction: training sample is carried out the pretreatment such as regularization, is guaranteed
Mean value is 0, variance 1；Step 2, Structure learning: being based on sliding window, and the increment type of structure is carried out in the form of lot sample sheet
Study；Step 3, parameter learning: the coefficient of relationship matrix and regularization sample obtained using Structure learning obtains current state
The mean value and variance of the Gaussian mixtures of each lower feature, i.e. continuous type Bayesian network parameters；Step 4, network updates:
It to the sample newly to arrive, is pre-processed using the regularization parameter of population sample, forms same scale using sliding window
The update of this progress of lot sample structure and parameter.
Preferably, step 1 specifically comprises the following steps；Step 1.1: rejecting the abnormal data in initial data；Step
1.2: original sample being carried out to overall regularization, and records the mean value of population sample and the regularization parameter information of variance.
Preferably, step 2 specifically comprises the following steps: step 2.1:, will be original big first, in accordance with the size of sliding window
The training lot sample sheet that scale sample decomposition is partially overlapped with each other at several；Step 2.2: initialization is related with sample size and dimension
The parameters such as penalty term, the degree of rarefication of regulated and control network structure；Step 2.3: sparse Bayesian network algorithm is utilized, successively to each
A node carries out the search of father node, and breadth first search method is added and avoids generating loop.By iteration several times, father's section
Point and its corresponding relationship coefficient matrix tend towards stability, and obtain the network structure under current data；Step 2.4: mobile sliding window,
Structure fine tuning is carried out using the network structure generated in step 2.2 and step 2.3 and coefficient of relationship matrix as original state.
Preferably, the specific calculating process of step 3 is as follows: step 3.1: the coefficient of relationship matrix obtained according to Structure learning obtains
Each node is obtained for the mean value of the Gaussian mixtures of father node；Step 3.2: calculating the association side of any father node combination
Difference；Step 3.3: using the coefficient of relationship of father node combination as weight, the covariance that adds up obtains the variance of Gaussian mixtures.
Preferably, step 4 specifically comprises the following steps: step 4.1: using the mean value and variance of population sample to new samples
Carry out regularization pretreatment；Step 4.2: new samples and part bulk sample originally being formed by new window according to sequential relationship, original
The update of structure and parameter is carried out on network foundation.
Compared to the prior art, technical solution provided by the invention has the following beneficial effects:
The sparse Bayesian network Increment Learning Algorithm based on continuous data is different from most of based on discrete type
The Bayesian Network Learning algorithm of data, the present invention substitute simple conditional probability table using Gaussian mixtures, can be more preferable
It must adapt to the industrial data of real world complexity.Also, common Bayesian Network Learning algorithm requires first to find candidate parent section
Point can omit correct father node in search process, and the present invention does not need to find candidate parent nodes, thus a kind of before eliminating
The drawbacks of method.Finally, incremental learning method proposed by the present invention, can not only accelerate convergence speed on big-sample data
Degree, additionally it is possible to which there are the flow datas of part concept drift for processing, as the time constantly automatically updates structure.Based on the above content,
The present invention is able to solve the problem of industrial data volume is excessive and flow data updates, so more preferable that meet continuous type magnanimity industry
Data causality analysis and the demand traced extremely.
Detailed description of the invention
The drawings described herein are used to provide a further understanding of the present invention, constitutes a part of the invention, this hair
Bright illustrative embodiments and their description are used to explain the present invention, and are not constituted improper limitations of the present invention.In the accompanying drawings:
The sparse Bayesian network incremental learning technology ensemble stream based on continuous type industrial data that Fig. 1 present invention is implemented
Cheng Tu；
The sparse Bayesian network incremental learning technology based on continuous type industrial data that Fig. 2 present invention is implemented flows in detail
Cheng Tu；
What the sparse Bayesian network incremental learning technology based on continuous type industrial data that Fig. 3 present invention is implemented obtained
Bayesian network structure figure.
Specific embodiment
In order to be clearer and more clear technical problems, technical solutions and advantages to be solved, tie below
Drawings and examples are closed, the present invention will be described in further detail.It should be appreciated that specific embodiment described herein is only used
To explain the present invention, it is not intended to limit the present invention.
In claims of the present invention, specification and above-mentioned attached drawing, unless otherwise specifically limited, such as using term " the
One ", " second " or " third " etc. are provided to distinguish different objects, be not use to describe a particular order.
In claims of the present invention, specification and above-mentioned attached drawing, unless otherwise specifically limited, for the noun of locality, such as
Using term " center ", " transverse direction ", " longitudinal direction ", "horizontal", " vertical ", "top", "bottom", "inner", "outside", "upper", "lower",
The indicating positions such as "front", "rear", "left", "right", " clockwise ", " counterclockwise " or positional relationship are orientation based on the figure
And positional relationship, and be merely for convenience of the narration present invention and simplify description, rather than the device or member of indication or suggestion meaning
Part must have a particular orientation or be constructed and operated in a specific orientation, so can not be interpreted as limiting of the invention specific
Protection scope.
In claims of the present invention, specification and above-mentioned attached drawing, unless otherwise specifically limited, such as using term " Gu
Connect " or " being fixedly connected ", it should broadly understood, i.e., between the two without any connection side of displacement relation and relative rotation relationship
Formula, that is to say, that including be unremovably fixedly connected with, be removably secured connection, be linked together and by other devices or
Element is fixedly connected.
In claims of the present invention, specification and above-mentioned attached drawing, such as using term " includes ", " having " and they
Deformation, it is intended that " including but not limited to ".
As depicted in figs. 1 and 2, a kind of sparse Bayesian network increment based on continuous type industrial data provided by the invention
Formula learning art, comprising the following steps:
Step 1, data prediction: carrying out the pretreatment such as regularization for training sample, guarantee mean value be 0, variance 1.Tool
Body includes the following steps:
Step 1.1: the abnormal data in initial data being excluded according to priori knowledge.Abnormal data refers to that equipment is non-
Data under data under normal condition, such as shutdown status or malfunction obtain sample data matrix rawD, size
For N × p；
Step 1.2: Regularization is carried out to original sample, it is ensured that every one-dimensional data is in bulk sample this on the basis of all mean values
It is 1 for 0 and variance, pretreated data is denoted as D, size is also N × p, and mean value and variance for recording bulk sample sheet etc. is just
Then change parameter information.
Step 2, Structure learning: being based on sliding window, and the incremental learning of structure is carried out in the form of lot sample sheet.Specifically
Include the following steps:
Step 2.1: in SBN algorithm, assuming initially that variable number is p, indicate obtained with the matrix D AG of a p × p
Directed acyclic graph, element DAGijThere are directed edge i → j between variable i, j for=1 representative, and directed edge is otherwise not present.Secondly,
Average information is provided for training process with the matrix P of p × p dimension, matrix P is represented and (closed between node with the presence or absence of path
It is closure), element Pij=1 represents variable i to variable j there are a paths, on the contrary then do not have.Since training needs, finally
It can learn to obtain the matrix B of (p-1) × p, record coefficient of relationship of each node with respect to its father node.Due to any
One node can not become the father node (there is no from ring) of oneself, so coefficient of relationship matrix B equivalence is by the coefficient of p × p
Square matrix diagonal entry is deleted, and lower triangle is spliced upwards.The window size batchSize of lot sample sheet is finally set；
Step 2.2: bayesian network structure learning, the core entirely learnt are optimization following formula:
s.t.βji×Pij=0, i, j=1,2, K, p, i ≠ j
Wherein, xiIt is the corresponding sample vector of variable i and x/iRepresent the later sample matrix of removal variable i, βiIt is relationship
The i-th column of coefficient matrix B are the corresponding column of variable i.The optimization aim of above formula is minimized between true value and Coefficient Fitting value
Error and L1The sum of canonical penalty term, wherein λ1For controlling the nonzero value number in matrix B, i.e., control network structure is dilute
Dredge property.λ1Bigger, the nonzero value in matrix B is fewer, and network structure is more sparse.And it needs to maintain β in an iterative processji×Pij
Perseverance is 0, to guarantee the acyclicity of network structure, wherein PijIt is the connectivity of variable i and variable j at this stage, connectivity is sentenced
It is disconnected to be solved using BFS (BreadthFirst Search, breadth first search) method.
Above formula is converted using lagrange formula are as follows:
In above formulaThe sample eliminated after variable i is represented, is added to two penalty term λ in the formula1、λ2,
Middle λ1It is still L1Regularization penalty term, for controlling sparsity；And λ2So that | βji×Pij| zero is leveled off to, to avoid oriented
Ring is formed in figure.By proving, λ is worked as in discovery2When meeting the following conditions, it can be ensured that not will form ring in the training process:
Step 2.3: given λ1、λ2Value after, BCD (BlockCoordinate Descent, BCD) can be passed through
Algorithm carrys out calculated relationship coefficient matrix B.BCD algorithm is a kind of piecemeal optimization method, and for matrix B, BCD algorithm fixes remaining institute
Some column, successively update βi, that is, successively optimizing fi(βi) until meeting the preset condition of convergence.Specifically, with L2Model
Formula variation is used as the condition of convergence less than 0.001.And for each fi(βi) optimization, can using be similar to LASSO
The form (being shown below) of (LeastAbsolute Shrinkage and Selection Operator, LASSO) optimization,
It is iterated using shooting algorithm:
Finally, the numerical value and symbol of each coefficient, continuous convergence of approximation item are calculated separately by recursive calculation following formula
Part:
The numerical value of coefficient, for judging whether node and its father node have relationship:
The symbol of coefficient, for judging the positive negativity of relationship between node and its father node:
This process is repeated for each subsequent window, due to the window study before having passed through a to coefficient
Matrix, it is only necessary to can converge to a new coefficient matrix by the iteration of seldom number.Total iteration time when such as initial training
It is tens of time or so to reach convergence, and only needing iteration one twice in the incrementally updating stage can restrain.Then, it is only necessary to
Numerical value is added in matrix B and is all 0 diagonal line, and sets 1 for nonzero value, can construct the adjoining formed by 0,1
Matrix D AG.
Step 3, parameter learning: the coefficient of relationship matrix and regularization sample obtained using Structure learning obtains current shape
The mean value and variance of the Gaussian mixtures of each feature, i.e. continuous type Bayesian network parameters under state.It specifically includes as follows
Step:
Step 3.1: the directed acyclic graph DAG and coefficient of relationship matrix B arrived using the acquistion of step 2 middle school is calculated each
A node Y is relative to his father's nodes X1,K,XkGaussian mixtures P (Y | X1,K,Xk):N(β0+β1X1L+βkXk,σ2) ginseng
Number.For the mean value of Gaussian mixtures, father node and its corresponding parameter are found by Y respective column in coefficient matrix B, to obtain
Obtain mean value β0+β1X1L+βkXk；
Step 3.2: for the variances sigma of Gaussian mixtures2, by estimating that the parameter of following likelihood function obtains:
In above formula, X indicates current and investigates feature, and x indicates to investigate the corresponding sampling feature vectors of feature, UkIndicate the kth of X
The sample values vector of a father node character pair, [m] indicate the numerical value of corresponding m-th of element of vector.The target of parameter learning
It is learning parameter θX|U={ β0,β1,K,βk,σ2, wherein { β0,β1,K,βkObtained by step 3.1, and Gaussian mixtures
Variances sigma2, by seeking local derviation to possibility predication function and to be set to zero available:
Wherein formula:
CovD[X；X]=ED[X；Y]-ED[X]ED[Y]
The mixed Gaussian variances sigma of each characteristic node is obtained according to above-mentioned formula2.In actual moving process, parameter
Study does not need to carry out constantly, when needing to carry out dependent probability analysis and causal reasoning, carries out parameter for specific node
Study, can effectively save computing resource.
Step 4, network updates: to the sample newly to arrive, being pre-processed, is utilized using the regularization parameter of population sample
Sliding window forms the update of this progress of the lot sample of same scale structure and parameter.Specifically comprise the following steps:
Step 4.1: for the sample rawS newly to arrive, the mean value obtained when using step 1.2 population sample regularization and
Variance generates pretreated sample S, and S is put into the end of sample；
Step 4.2: sliding window is one mobile, form new batch data.Structure several times is carried out under new data
The iteration of habit can converge to the coefficient of relationship matrix B of stable state on the basis of prototype structuret+1, and generate corresponding adjoining
Matrix D AGt+1。
Next with specific embodiment, the present invention is further elaborated:
In order to clearly demonstrate implementation process, supplied as shown in figure 3, choosing 9 power generations of a 1000MW coal heating unit
Thermal parameter is illustrated for annual 1 minute 2017 for the historical data being spaced.
1, data prediction
1.1, noise data is rejected
There are two types of situations for noise data in initial data, first is that full attribute is 0 or the thick-and-thin variable category of numerical value
Property, second is that the shutdown time data that generator power is 0 records.In order to illustrate conveniently, this example extracts on the basis of initial data
Part sample and attribute, only illustrate implementation process, the part sample data after cancelling noise is as shown in table 1.
Part sample data after 1 example unit cancelling noise of table
1.2, data regularization
To the data after cancelling noise carry out mean value be 0, the regularization that variance is 1, obtain that the results are shown in Table 2.
2 example unit sample data regularization result of table
2, Structure learning
This example has chosen 1000 datas record as training sample, therefore, it is necessary to be arranged 9 × 9 matrix D AG and
8 × 9 coefficient of relationship matrix B.After training, obtained bayesian network structure figure such as Fig. 3.In order to preferably express pass
It is coefficient matrix B, the matrix that the diagonal entry that addition numerical value is all 0 in matrix B is extended to 9 × 9 is as shown in table 3,
Diagonal entry (extensible element) is indicated with "-" in table, is not had these elements in former coefficient of relationship matrix, is merely to illustrate section
The coefficient of relationship of point and its father node.
The coefficient of relationship matrix that 3 example set structure of table learns (after extension)
3, parameter learning
3.1, mixed Gaussian mean value computation
The mean value of mixed Gaussian is directly obtained by coefficient of relationship matrix B.By taking the 1st node (feature) as an example, square has been corresponded to
0th column of battle array, therefore its mixed Gaussian mean value is indicated as X0=0.3157X1-0.2702X2-0.3049X4+0.1792X5, with
This analogizes linear expression of other the available nodes relative to its father node.
3.2, mixed Gaussian variance calculates
Mixed Gaussian variance calculates the formula for using step 3.2, is directed to each node, calculates separately its variance.9
The mixed Gaussian variance of node (feature) is as shown in table 4.
The mixed Gaussian variance that 4 example unit parameter of table learns
4, network updates
4.1, flow data sample regularization
For the sample rawS newly to arrive, regularization is carried out to it, obtains this regularization of bulk sample mean value and standard deviation such as table 5
It is shown.
5 example unit new samples regularization mean value of table, standard deviation
4.2, incrementally updating
In this example, sliding window size is 1000.When new data arrives, sliding window is moved along a sample
This.Due to the particularity of algorithm itself, it is also required to recalculate each coefficient in incrementally updating.By new sample
After this recalculate, network structure is not adjusted, and (table 6 be the relationship after extending to the adjustment of coefficient of relationship matrix B as shown in table 6
Coefficient matrix).
The coefficient of relationship matrix that 6 example unit of table updates after new samples arrival through network (after extension)
As can be seen from Table 6, the variation of coefficient of relationship matrix is smaller, this is because single sample is relative to this very little of bulk sample,
So being capable of the preferable stability that must keep network structure.Although unobvious for network structure update, work as flow data in a steady stream
Constantly enter, structure can slowly be updated towards target direction.
The preferred embodiment of the present invention has shown and described in above description, as previously described, it should be understood that the present invention is not office
Be limited to form disclosed herein, should not be regarded as an exclusion of other examples, and can be used for various other combinations, modification and
Environment, and can be changed within that scope of the inventive concept describe herein by the above teachings or related fields of technology or knowledge
It is dynamic.And changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, then it all should be appended by the present invention
In scope of protection of the claims.
Claims (5)
1. a kind of sparse Bayesian network incremental learning method based on continuous type industrial data characterized by comprising
Step 1, data prediction: carrying out the pretreatment such as regularization for training sample, guarantee mean value be 0, variance 1；
Step 2, Structure learning: being based on sliding window, and the incremental learning of structure is carried out in the form of lot sample sheet；
Step 3, parameter learning: the coefficient of relationship matrix and regularization sample obtained using Structure learning is obtained under current state
The mean value and variance of the Gaussian mixtures of each feature, i.e. continuous type Bayesian network parameters；
Step 4, network updates: to the sample newly to arrive, being pre-processed using the regularization parameter of population sample, utilizes sliding
Window forms the update of this progress of the lot sample of same scale structure and parameter.
2. a kind of sparse Bayesian network incremental learning method based on continuous type industrial data as described in claim 1,
It is characterized by: step 1 specifically comprises the following steps；
Step 1.1: rejecting the abnormal data in initial data；
Step 1.2: original sample is carried out to overall regularization, and the regularization parameters such as mean value and variance for recording population sample
Information.
3. a kind of sparse Bayesian network incremental learning method based on continuous type industrial data as described in claim 1,
It is characterized by: step 2 specifically comprises the following steps:
Step 2.1: first, in accordance with the size of sliding window, original extensive sample decomposition being partially overlapped with each other at several
Training lot sample sheet；
Step 2.2: the parameters such as initialization penalty term related with sample size and dimension, the degree of rarefication of regulated and control network structure；
Step 2.3: sparse Bayesian network (Sparse Bayesian Network, SBN) algorithm is utilized, successively to each
Node carries out the search of father node, and breadth first search method is added and avoids generating loop.By iteration several times, father node
And its corresponding relationship coefficient matrix tends towards stability, and obtains the network structure under current data；
Step 2.4: mobile sliding window, using the network structure generated in step 2.2 and step 2.3 and coefficient of relationship matrix as
Original state carries out structure fine tuning.
4. a kind of sparse Bayesian network incremental learning method based on continuous type industrial data as described in claim 1,
It is characterized by: the specific calculating process of step 3 is as follows:
Step 3.1: the coefficient of relationship matrix obtained according to Structure learning obtains the mixed Gaussian that each node is directed to father node
The mean value of distribution；
Step 3.2: calculating the covariance of any father node combination；
Step 3.3: using the coefficient of relationship of father node combination as weight, the covariance that adds up obtains the variance of Gaussian mixtures.
5. a kind of sparse Bayesian network incremental learning method based on continuous type industrial data as described in claim 1,
It is characterized by: step 4 specifically comprises the following steps:
Step 4.1: regularization pretreatment being carried out to new samples using the mean value and variance of population sample；
Step 4.2: new samples and part bulk sample originally being formed by new window according to sequential relationship, are carried out on the basis of primitive network
The update of structure and parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811009985.9A CN109190702A (en) | 2018-08-31 | 2018-08-31 | Sparse Bayesian network incremental learning method based on continuous type industrial data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811009985.9A CN109190702A (en) | 2018-08-31 | 2018-08-31 | Sparse Bayesian network incremental learning method based on continuous type industrial data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109190702A true CN109190702A (en) | 2019-01-11 |
Family
ID=64917630
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811009985.9A Pending CN109190702A (en) | 2018-08-31 | 2018-08-31 | Sparse Bayesian network incremental learning method based on continuous type industrial data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109190702A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110928922A (en) * | 2019-11-27 | 2020-03-27 | 开普云信息科技股份有限公司 | Public policy analysis model deployment method and system based on big data mining |
| WO2020191722A1 (en) * | 2019-03-28 | 2020-10-01 | 日本电气株式会社 | Method and system for determining causal relationship, and computer program product |
-
2018
- 2018-08-31 CN CN201811009985.9A patent/CN109190702A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020191722A1 (en) * | 2019-03-28 | 2020-10-01 | 日本电气株式会社 | Method and system for determining causal relationship, and computer program product |
| US11537910B2 (en) | 2019-03-28 | 2022-12-27 | Nec Corporation | Method, system, and computer program product for determining causality |
| CN110928922A (en) * | 2019-11-27 | 2020-03-27 | 开普云信息科技股份有限公司 | Public policy analysis model deployment method and system based on big data mining |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| He et al. | Variational quantum compiling with double Q-learning | |
| Ortiz et al. | Dimensional synthesis of mechanisms using differential evolution with auto-adaptive control parameters | |
| CN107357757B (en) | Algebraic application problem automatic solver based on deep reinforcement learning | |
| CN109190702A (en) | Sparse Bayesian network incremental learning method based on continuous type industrial data | |
| CN117077671B (en) | Interactive data generation method and system | |
| Lu et al. | A new hybrid algorithm for bankruptcy prediction using switching particle swarm optimization and support vector machines | |
| Rodríguez-Fdez et al. | An instance selection algorithm for regression and its application in variance reduction | |
| CN110502739A (en) | The building of the machine learning model of structuring input | |
| CN108536844B (en) | Text-enhanced network representation learning method | |
| Li et al. | A novel adaptive weight algorithm based on decomposition and two-part update strategy for many-objective optimization | |
| CN113988418A (en) | Visualization method for energy load prediction | |
| Liu et al. | Gradient‐Sensitive Optimization for Convolutional Neural Networks | |
| CN111311000B (en) | User consumption behavior prediction model training method, device, equipment and storage medium | |
| CN116415177A (en) | Classifier parameter identification method based on extreme learning machine | |
| Nitti | Hybrid probabilistic logic programming | |
| CN111813837B (en) | Method for intelligently detecting data quality | |
| CN113867724A (en) | Method and system for automatically generating GUI (graphical user interface) code, server and medium | |
| Goodwin et al. | A vector quantization approach to scenario generation for stochastic NMPC | |
| McFall | An artificial neural network method for solving boundary value problems with arbitrary irregular boundaries | |
| Briol et al. | Rejoinder: Probabilistic integration: A role in statistical computation? | |
| Dou et al. | Research on Power Network Regulation Mechanism Based on Knowledge Mapping | |
| Sun et al. | Remote supervision relation extraction method of power safety regulations knowledge graph based on ResPCNN-ATT | |
| US20240220688A1 (en) | Numerical Simulation Method By Deep Learning And Associated Recurrent Neural Network | |
| Nishino et al. | A sparse parameter learning method for probabilistic logic programs | |
| CN111506962B (en) | Complex system reliability calculation method based on BN and UGF |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190111 |
|
| RJ01 | Rejection of invention patent application after publication |