CN109254219A - A kind of distribution transforming transfer learning method for diagnosing faults considering multiple factors Situation Evolution - Google Patents
A kind of distribution transforming transfer learning method for diagnosing faults considering multiple factors Situation Evolution Download PDFInfo
- Publication number
- CN109254219A CN109254219A CN201811396356.6A CN201811396356A CN109254219A CN 109254219 A CN109254219 A CN 109254219A CN 201811396356 A CN201811396356 A CN 201811396356A CN 109254219 A CN109254219 A CN 109254219A
- Authority
- CN
- China
- Prior art keywords
- distribution transforming
- state
- index
- failure
- target
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention provides a kind of distribution transforming transfer learning fault diagnosis model, comprising: (1) is divided into dynamic state quantity, Quasi dynamic quantity of state and static state amount to the quantity of state for influencing distribution transforming operating status, constructs distribution transforming evaluation of running status index system；(2) binary quantization is carried out to index quantity of state, excavates its incidence relation between failure using Apriori algorithm, extracts the key index quantity of state of induction transformer fault；(3) Tanimoto coefficient is introduced, by effective assist trouble Data Migration to target distribution transforming；(4) it is iterated solution using weight of the transfer learning algorithm TrAdaBoost to target faults data and assist trouble data, distribution transforming fault diagnosis model is obtained, to carry out the fault diagnosis of target distribution transforming.The present invention migrates the fault message for assisting distribution transforming to target distribution transforming, preferably solves distribution transforming monomer fault data and gives distribution transforming fault diagnosis bring problem less.
Description
Technical field
The invention belongs to distribution transformer fault diagnosis field, more specifically, a kind of consider multiple factors Situation Evolution
Distribution transforming transfer learning method for diagnosing faults.
Background technique
In distribution, distribution transformer (distribution transforming) substantial amounts guarantee that its safety is the base of stabilization of power grids reliability service
Plinth is ensureing power supply reliability, is realizing risk to the timely investigation of the accurate perception of its state, the Accurate Diagnosis of failure, risk
Early warning, reduction contingency occurrence probability etc. are significant.
In recent years, it with the rapid development of the technologies such as big data, data mining, is obtained in terms of transformer fault diagnosis
It is widely applied.Research is concentrated mainly on the pass between intelligent extraction, fault type and the influence factor of fault characteristic value
The excavation of connection property, the multiprecision arithmetic design of fault diagnosis, the efficient Rapid Detection of fault type, unstructured data are examined in failure
Application in disconnected etc..Such research fault diagnosis object be mainly transmitting transformer (refer to the main transformer of 110kV or more,
The defeated change of hereinafter referred), in practice, it is defeated change and distribution transforming operating condition, monitoring means, in terms of exist it is larger
Difference.Defeated accommodation Chang Buhui overlond running, operating condition is good, so generally influence of the external environment to its failure is not considered,
And due to involving great expense, the replacement cycle it is long, to its internal state amount monitoring be usually number comprehensive, for fault diagnosis
It is abundant according to amount；And distribution transforming substantial amounts, quality are irregular, update fast, data monitoring amount is not comprehensive, weather and negative
The operating conditions such as load rate averagely will cause larger impact to distribution transforming healthy water.By above-mentioned analysis, the failure of defeated change and distribution transforming is lured
Because having differences, monitoring state amount is different, and the difference of data rich degree is obvious, cannot be simple by the method for diagnosing faults of defeated change
It is transplanted to distribution transforming.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the present invention is intended to provide a kind of consideration multiple factors situation is drilled
The distribution transforming transfer learning method for diagnosing faults of change solves distribution transforming fault diagnosis and asks for the few problem of distribution transforming monomer fault data
Topic.
In order to achieve the above object, the invention discloses a kind of distribution transforming transfer learning events for considering multiple factors Situation Evolution
Hinder diagnostic method, includes the following steps:
1) to influence distribution transforming operating status index quantity of state be divided into dynamic indicator quantity of state, Quasi dynamic index quantity of state and
Static State Index quantity of state constructs distribution transforming evaluation of running status index system on this basis；
2) binary quantization is carried out to the index quantity of state in state evaluation index system, calculates this using Apriori algorithm
Incidence relation between a little index quantity of states and distribution transforming failure, index quantity of state stronger for incidence relation are regarded as luring
Lead the key index quantity of state of distribution transforming failure；
3) it obtains the target distribution transforming for needing to carry out fault diagnosis and possesses the history distribution transforming of failure logging under all kinds of failures
Key index state quantity data, utilize Tanimoto coefficient calculate target distribution transforming and history distribution transforming key index quantity of state number
According to similarity, select the higher history distribution transforming of similarity as auxiliary distribution transforming；
4) using transfer learning algorithm TrAdaBoost to the key index state quantity data of target distribution transforming and auxiliary distribution transforming
The weight of key index state quantity data solved, and then distribution transforming fault diagnosis model is obtained, to carry out target distribution transforming
Fault diagnosis.
Further, dynamic indicator quantity of state, Quasi dynamic index quantity of state and Static State Index quantity of state in the step 1)
It is respectively as follows:
A. dynamic indicator quantity of state refers to the data that can be acquired in real time, and such as meteorological data electrically acquires data.Such index
The quantity of state update cycle is shorter, and general 15min or 1h can update once, therefore real-time is preferable；In addition, such index shape
State amount can effecting reaction distribution transforming state dynamic change, be the important assessment source of distribution transforming fault diagnosis；
B. Quasi dynamic index quantity of state refers to the data by periodically or non-periodically obtaining, test and inspection number such as transformer
According to.The collection period of such index quantity of state is much higher than dynamic state quantity, therefore negligible amounts；Simultaneously as the scale of construction of distribution transforming
It is huge, the replacement cycle is short, therefore such index quantity of state difficulty of comprehensive acquisition is larger.Quasi dynamic index quantity of state is assessment
Distribution transforming local defect referring especially to foundation；
C. Static State Index quantity of state refers to the information for not needing to obtain by system interaction, such as account information, history electricity consumption.
Such index quantity of state is once placed on record and is just no longer modified, and is the data source for reacting distribution transforming state for time cumulative effect.
Further, binary quantization method is carried out to the index quantity of state in state evaluation index system in the step 2)
Are as follows:
B. the index quantity of state attribute set of distribution transforming is set
P={ p1,p2,...,pM} (1)
Wherein, P indicates that the index quantity of state attribute and failure collection of distribution transforming, element include influencing distribution transforming operating status
Dynamic indicator quantity of state, Quasi dynamic index quantity of state and Static State Index quantity of state；p1,p2,...,pMIndicate index quantity of state attribute
Element in set P indicates some index quantity of state in dynamic indicator quantity of state, or indicates in Quasi dynamic index quantity of state
Some index quantity of state, or indicate some index quantity of state in Static State Index quantity of state；M indicates index quantity of state number；
B. choose certain distribution transforming i, set its failure affairs library as
Di={ d1,d2,...,dN} (2)
Wherein, DiIndicate distribution transforming i failure affairs library, element includes the failure logging of distribution transforming i；d1,d2,...,dNIt indicates
Distribution transforming failure affairs library DiIn element, indicate certain failure logging of distribution transforming i, the element in every failure logging corresponds in P
The value and fault type of each element；The failure logging item number of N expression distribution transforming i；
C. " good " and " bad " are divided into according to superiority and inferiority to the element in the index quantity of state attribute of distribution transforming and failure collection P, it is right
DiIn any one failure logging, if in the record certain index quantity of state be " good ", 1 is denoted as, if certain index in the record
Quantity of state is " bad ", then is denoted as 0.
Further, it is calculated between these index quantity of states and distribution transforming failure in the step 2) using Apriori algorithm
Incidence relation, index quantity of state stronger for incidence relation, be regarded as induction distribution transforming failure key index quantity of state
Method are as follows:
A. for DiIn any one failure logging, set the index quantity of state attribute of its distribution transforming as association rule mining
In former piece X, failure is the consequent Y in association rule mining, if certain index quantity of state in failure logging is 1, then it is assumed that
The index quantity of state exists with the fault type of this failure logging to be associated with, if certain index quantity of state in failure logging is 0,
Think that the index quantity of state, there is no being associated with, chooses suitable support threshold with the fault type of this failure logging, extracts
Support is greater than the correlation rule of support threshold
Wherein, the support calculation formula of correlation rule X → Y is
Wherein, s (X → Y) indicates the support of correlation rule X → Y；The support of σ (X → Y) expression correlation rule X → Y
It counts；The failure logging item number of N expression distribution transforming i；
B. it chooses suitable confidence threshold value, extracts the correlation rule that confidence level is greater than confidence threshold value, correlation rule X →
The confidence calculations formula of Y is
Wherein, c (X → Y) indicates the confidence level of correlation rule X → Y；The support of σ (X → Y) expression correlation rule X → Y
It counts；σ (X) indicates the support counting of former piece X in correlation rule.
C. choose while meeting the correlation rule that support is greater than confidence threshold value greater than support threshold, confidence level；
D. all to distribution transforming i at the same meet support greater than support threshold, confidence level be greater than confidence threshold value association
Index state in rule measures union, constitutes the key index state duration set I of distribution transforming ii；
E. I is takeniUnion, constitute induction transformer fault key index quantity of state I
Wherein, Z indication transformer number of units；IiIndicate the key index state duration set extracted from i-th distribution transforming；I table
Show the index index state duration set for distribution transforming fault diagnosis.
Further, the key index of target distribution transforming and history distribution transforming is calculated in the step 3) using Tanimoto coefficient
The similarity of state quantity data selects the higher history distribution transforming of similarity as the method for auxiliary distribution transforming are as follows:
A. the failure distribution proportion for defining target distribution transforming and history distribution transforming is as follows:
In formula, R indicates fault type number；Indicate the failure r in target distribution transforming in the faulty middle proportion of institute；Table
Show the failure r in history distribution transforming in the faulty middle proportion of institute；PaIndicate the defective proportion set of target distribution transforming；PbExpression is gone through
The defective proportion set of history distribution transforming.
B. set the key index state duration set of target distribution transforming and history distribution transforming asWith
In formula,Indicate the per unit value of the key index quantity of state of target distribution transforming；Indicate the key index of history distribution transforming
The per unit value of quantity of state；mrIndicate the quantity of failure r in target distribution transforming；nrIndicate the quantity of failure r in history distribution transforming.
C. rightWithIn vector average processing:
In formula,It indicatesSet after handling averagely；It indicatesSet after handling averagely.
D. Tanimoto coefficient is introduced, the similarity of target distribution transforming and history distribution transforming failure r is obtained:
In formula,It indicates target distribution transforming and assists the similarity of distribution transforming failure r.
E. combination failure distribution proportion obtains the faulty comprehensive similarity of target distribution transforming and history distribution transforming institute
In formula, TRIndicate the faulty comprehensive similarity of target distribution transforming and history distribution transforming institute.
F. mobilance threshold value δ is defined, history distribution transforming is screened, if TRThis history distribution transforming is then considered as auxiliary and matched by >=δ
Become, key index state quantity data will be used for next target distribution transforming fault diagnosis；Conversely, if TR< δ then abandons this and goes through
History distribution transforming and the key index state quantity data corresponding to it.
Further, utilize transfer learning algorithm TrAdaBoost to the key index shape of target distribution transforming in the step 4)
The weight of state amount data and the key index state quantity data of auxiliary distribution transforming is solved, and then obtains distribution transforming fault diagnosis mould
Type, the specific steps are that:
A. the failure logging of target distribution transforming is set as Ta, TaGather as target training, the failure logging of setting auxiliary distribution transforming
For Tb, TbAs supplemental training set；
B. T is setaSample size is m, TbSample size is n, merges training set T=Ta∪Tb, the number of iterations Iter, base
This classification algorithm of neural network, wherein
In formula, x is the per unit value of distribution transforming key index quantity of state, and y is fault type；
C. weight vectors are initializedWherein
D. initiation parameter
E. start iteration, iteration t=1,2 ..., Iter
F. fault diagnosis model is exported, the output of the model is the fault diagnosis result of target distribution transforming
Further, iterative process in the step e are as follows:
A. weight normalizes, and enables
B. neural network algorithm is called, according to T, ptObtain the weak diagnostor h of failuret:X→Y；
C. the weak diagnostor h of failure is calculatedtIn TaError rate above:
In formula, ht(xi) presentation class device is to xiObtained study mark；
D., the weak diagnostor weight parameter α of failure is settAnd target weight adjusting parameter βt
E. weight updates
Compared with the prior art, the present invention has the following beneficial effects:
Distribution transforming transfer learning fault diagnosis model established by the present invention will assist distribution transforming by way of transfer learning
Fault message is migrated to target distribution transforming, and it is few difficult to distribution transforming fault diagnosis bring preferably to solve distribution transforming monomer fault data
Topic.
Detailed description of the invention
Fig. 1 is the distribution transforming evaluation of running status index schematic diagram of the embodiment of the present invention；
Fig. 2 is that the distribution transforming Quasi dynamic quantity of state of the embodiment of the present invention and the correlation rule diagram of failure are intended to；
Fig. 3 is the auxiliary data of the embodiment of the present invention and the relational graph schematic diagram of diagnostic accuracy；
Fig. 4 is the mobilance threshold value of the embodiment of the present invention and the relational graph schematic diagram of diagnostic accuracy；
Fig. 5 be the embodiment of the present invention different data amount under M1 diagnosis accuracy schematic diagram；
Fig. 6 is influence schematic diagram of the number of iterations to M1 diagnosis accuracy of the embodiment of the present invention；
Fig. 7 is the flow chart for the distribution transforming transfer learning method for diagnosing faults that the present invention considers multiple factors Situation Evolution.
Specific embodiment
Below with reference to the drawings and specific embodiments in the present invention, the technical solution in the present invention is carried out clear, complete
Ground description.
As shown in fig. 7, the present invention provides a kind of distribution transforming transfer learning fault diagnosis side for considering multiple factors Situation Evolution
Method, comprising the following steps:
1) to influence distribution transforming operating status index quantity of state be divided into dynamic indicator quantity of state, Quasi dynamic index quantity of state and
Static State Index quantity of state constructs distribution transforming evaluation of running status index system on this basis；
2) binary quantization is carried out to the index quantity of state in state evaluation index system, calculates this using Apriori algorithm
Incidence relation between a little index quantity of states and distribution transforming failure, index quantity of state stronger for incidence relation are regarded as luring
Lead the key index quantity of state of distribution transforming failure；
3) it obtains the target distribution transforming for needing to carry out fault diagnosis and possesses the history distribution transforming of failure logging under all kinds of failures
Key index state quantity data, utilize Tanimoto coefficient calculate target distribution transforming and history distribution transforming key index quantity of state number
According to similarity, select the higher history distribution transforming of similarity as auxiliary distribution transforming；
4) using transfer learning algorithm TrAdaBoost to the key index state quantity data of target distribution transforming and auxiliary distribution transforming
The weight of key index state quantity data solved, and then distribution transforming fault diagnosis model is obtained, to carry out target distribution transforming
Fault diagnosis
Specifically, dynamic state quantity, Quasi dynamic quantity of state and static state amount are respectively as follows: in the step 1)
A. dynamic state quantity refers to the data that can be acquired in real time, and such as meteorological data electrically acquires data.Such quantity of state is more
The new period is shorter, and general 15min or 1h can update once, therefore real-time is preferable；In addition, such quantity of state can be effectively anti-
The dynamic change for answering distribution transforming state is the important assessment source of distribution transforming fault diagnosis；
B. Quasi dynamic quantity of state refers to the data by periodically or non-periodically obtaining, such as the test and inspection data of transformer.
The collection period of such quantity of state is much higher than dynamic state quantity, therefore negligible amounts；Simultaneously as the scale of construction of distribution transforming is huge, more
Change that the period is short, therefore such quantity of state difficulty of comprehensive acquisition is larger.Quasi dynamic quantity of state is assessment distribution transforming local defect
Referring especially to foundation；
C. static state amount refers to the information for not needing to obtain by system interaction, such as account information, history electricity consumption.Such
Quantity of state is once placed on record and is just no longer modified, and is the data source for reacting distribution transforming state for time cumulative effect.
Specifically, in the present embodiment, distribution transforming evaluation of running status index system is as shown in Figure 1.
Specifically, carrying out binary quantization method to index quantity of state in the step 2) are as follows:
C. the quantity of state attribute set of distribution transforming is set
P={ p1,p2,...,pM} (1)
Wherein, P indicates that the quantity of state attribute and failure collection of distribution transforming, element include the dynamic for influencing distribution transforming operating status
Quantity of state, Quasi dynamic quantity of state and static state amount；p1,p2,...,pMIt indicates the element in quantity of state attribute set P, indicates dynamic
Some quantity of state in state quantity of state, or indicate some quantity of state in Quasi dynamic quantity of state, or indicate in static state amount
Some quantity of state；M indicates quantity of state number；
B. choose certain distribution transforming i, set its failure affairs library as
Di={ d1,d2,...,dN} (2)
Wherein, DiIndicate distribution transforming i failure affairs library, element includes the failure logging of distribution transforming i；d1,d2,...,dNIt indicates
Distribution transforming failure affairs library DiIn element, indicate certain failure logging of distribution transforming i, the element in every failure logging corresponds in P
The value and fault type of each element；The failure logging item number of N expression distribution transforming i；
C. " good " and " bad " are divided into according to superiority and inferiority to the element in the quantity of state attribute of distribution transforming and failure collection P, to DiIn
Any one failure logging, if in the record certain quantity of state be " good ", be denoted as 1, if in the record certain quantity of state be " bad ",
Then it is denoted as 0.
Clearly to indicate P and DiBetween relationship, the relationship drawn between the two is as shown in table 1:
Table 1P and DiBetween relationship
Specifically, excavating its incidence relation between failure using Apriori algorithm in the step 2), induction is extracted
The method of the Key state of transformer fault are as follows:
A. for DiIn any one failure logging, set the quantity of state attribute of its distribution transforming as in association rule mining
Former piece X, failure are the consequent Y in association rule mining, if certain quantity of state in failure logging is 1, then it is assumed that the quantity of state
Exist with the fault type of this failure logging and be associated with, if certain quantity of state in failure logging is 0, then it is assumed that the quantity of state and this
There is no associations for the fault type of failure logging, choose suitable support threshold, extract support and are greater than support threshold
Correlation rule
Wherein, the support calculation formula of correlation rule X → Y is
Wherein, s (X → Y) indicates the support of correlation rule X → Y；The support of σ (X → Y) expression correlation rule X → Y
It counts；The failure logging item number of N expression distribution transforming i；
For popular statement step, with DiIn a few fault datas for, illustrate extract support be greater than support
The method of the correlation rule of threshold value:
Assuming that
p1p2p3p4p5... fault type
d1=1,0,1,1,0 ..., Y1}
d2=1,1,0,1,0 ..., Y1}
d3=0,0,1,1,1 ..., Y2}
......
That is the 1st article of fault data and the 2nd article of fault data are Y1Failure, the 3rd article of fault data are Y2Failure,
Assuming that the failure logging item number N=10 of distribution transforming i, support threshold 0.18 are then Y in distribution transforming i1When failure,
p1Occur 2 times for 1, σ (p1→Y1)=2, s (p1→Y1)=2/10=0.2 is greater than support threshold 0.18,
p4Occur 2 times for 1, σ (p4→Y1)=2, s (p4→Y1)=2/10=0.2 is greater than support threshold 0.18,
p1And p4Occur 2 times for 1 simultaneously, σ (p1,p4→Y1)=2, s (p1,p4→Y1)=2/10=0.2 is greater than and supports
Threshold value 0.18 is spent,
Therefore, the correlation rule that support is greater than support threshold has:
p1→Y1, p4→Y1, p1,p4→Y1
B. it chooses suitable confidence threshold value, extracts the correlation rule that confidence level is greater than confidence threshold value, correlation rule X →
The confidence calculations formula of Y is
Wherein, c (X → Y) indicates the confidence level of correlation rule X → Y；The support of σ (X → Y) expression correlation rule X → Y
It counts；σ (X) indicates the support counting of former piece X in correlation rule.
For popular statement step (2.6), with DiIn a few fault datas for, illustrate extract confidence level be greater than confidence
Spend the method for the correlation rule of threshold value:
Assuming that
p1p2p3p4p5... fault type
d1=1,0,1,1,0 ..., Y1}
d2=1,1,0,1,0 ..., Y1}
d3=0,0,1,1,1 ..., Y2}
......
That is the 1st article of fault data and the 2nd article of fault data are Y1Failure, the 3rd article of fault data are Y2Failure,
Assuming that confidence threshold value is 0.8, it is apparent from,
σ(p1)=2, c (p1→Y1)=2/2=1 is greater than confidence threshold value 0.8,
σ(p4)=3, c (p4→Y1)=2/3=0.67 is less than confidence threshold value 0.8,
σ(p1,p4)=2, c (p1,p4→Y1)=2/2=1 is greater than confidence threshold value 0.8,
Therefore, the correlation rule that confidence level is greater than confidence threshold value has:
p1→Y1, p1,p4→Y1
C. choose while meeting the correlation rule that support is greater than confidence threshold value greater than support threshold, confidence level；
D. all to distribution transforming i at the same meet support greater than support threshold, confidence level be greater than confidence threshold value association
State in rule measures union, constitutes the key state duration set I of distribution transforming ii；
E. I is takeniUnion, constitute induction transformer fault Key state I
Wherein, Z indication transformer number of units；IiIndicate the key state duration set extracted from i-th distribution transforming；I indicates to use
In the index state duration set of distribution transforming fault diagnosis.
Specifically, in the present embodiment, support threshold is set as 0.1, and confidence threshold value is set as 0.6, Quasi dynamic shape
The correlation rule of state amount and fault type is as shown in Figure 2.In figure, abscissa indicates the Quasi dynamic amount of distribution transforming, and ordinate expression is matched
The fault type of change, vertical coordinate indicate level of confidence, and the height of cylindrical body indicates the confidence level of correlation rule, on cylindrical body
Digital representation correlation rule support.Gray plane indicates confidence threshold value in figure, and only cylinder, should by the plane cutting
Correlation rule can just be considered as strong association, and corresponding quantity of state can just be considered as influencing the Key state of transformer fault.It is logical
Screening is crossed, in the present embodiment, removes wind-force, precipitation, absorptance, rated capacity and all kinds of losses, remaining quantity of state and distribution transforming
Failure has stronger incidence relation, needs to consider the influence of these Key states in the fault diagnosis of distribution transforming.
Specifically, introducing Tanimoto coefficient in the step 3), effective assist trouble Data Migration to target is matched
The method of change are as follows:
A. it defines target distribution transforming and assists the failure distribution proportion of distribution transforming as follows:
In formula, R indicates fault type number；Indicate the failure r in target distribution transforming in the faulty middle proportion of institute；Table
Show the failure r in auxiliary distribution transforming in the faulty middle proportion of institute；PaIndicate the defective proportion set of target distribution transforming；PbIndicate auxiliary
Help the defective proportion set of distribution transforming.
B. set state duration set asWith
In formula,Indicate the per unit value of the Key state of target distribution transforming；Indicate the Key state of auxiliary distribution transforming
Per unit value；mrIndicate the quantity of failure r in target distribution transforming；nrIndicate the quantity of failure r in auxiliary distribution transforming.
C. rightWithIn vector average processing:
In formula,It indicatesSet after handling averagely；It indicatesSet after handling averagely.
D. Tanimoto coefficient is introduced, target distribution transforming is obtained and assists the similarity of distribution transforming failure r:
In formula,It indicates target distribution transforming and assists the similarity of distribution transforming failure r.
E. combination failure distribution proportion obtains the faulty comprehensive similarity of target distribution transforming and auxiliary distribution transforming institute
In formula, TRIndicate the faulty comprehensive similarity of target distribution transforming and auxiliary distribution transforming institute.
F. mobilance threshold value δ is defined, auxiliary distribution transforming is screened, if TR>=δ, then by the fault data of this auxiliary distribution transforming
As information to be migrated；Conversely, if TR< δ then abandons the fault data of this auxiliary distribution transforming.
Specifically, in the present embodiment, mobilance threshold value δ is selected as 0.6, chooses a distribution transforming and matches as target to be diagnosed
Become, comprehensive similarity inspection is carried out to the fault data of target distribution transforming and other 14 distribution transformings.Through examining, the failure of 8 distribution transformings
Data and target distribution transforming similarity are higher, and as auxiliary distribution transforming, number is followed successively by f1-f8, remaining 6 distribution transforming number consecutively
For s1-s6.
Specifically, using transfer learning algorithm TrAdaBoost to target faults data and assist trouble in the step 4)
The weight of data is iterated solution, the method for obtaining distribution transforming fault diagnosis device are as follows:
A. the failure logging of target distribution transforming is set as Ta, TaGather as target training, setting auxiliary distribution transforming passes through similarity
Failure logging after inspection is Tb, TbAs supplemental training set；
B. T is setaSample size is m, TbSample size is n, merges training set T=Ta∪Tb, the number of iterations Iter, base
This classification algorithm of neural network, wherein
In formula, x is the per unit value of distribution transforming Key state, and y is fault type；
C. weight vectors are initializedWherein
D. initiation parameter
E. start iteration, iteration t=1,2 ..., Iter
F. the strong diagnostor of failure is exported
Specifically, in the present embodiment, to illustrate the invention in diagnostor (being denoted as M1) accuracy, by its with only lead to
Look over so as to check mark switching data training failure modes diagnostor (being denoted as M1_0) compare.After obtaining diagnostor, using target
Input of the distribution transforming quantity of state as diagnostor calculates diagnostic result.The diagnosis comparison of two kinds of diagnostors is as shown in table 2.From table
As can be seen that M1 is higher in the diagnostic result precision of various fault types, stability is also higher；And M1_0 is only in overheating fault
And it is more accurate in discharge fault diagnosis, and it is lower to the diagnostic accuracy of other types failure.This is because target distribution transforming
Fault data amount is very little, and the fault diagnosis device M1_0 generalization ability only trained with it is weaker, it is difficult to according to new input state amount
Correct failure modes are obtained as a result, even using training data as the input of fault diagnosis device, it is also difficult to obtain satisfied meter
Calculate result；And M1 has carried out the primary screening of fault message using comprehensive similarity by means of the fault message of other distribution transformings, and
Constantly be iterated the weight of fault message by transfer learning, effective information utilization rate is higher.
The diagnosis of table 2M1 and M1_0 compare
Influence for verifying assist trouble data volume to M1 fault diagnosis precision, the auxiliary distribution transforming failure of different number is believed
It ceases and is participated in the fault diagnosis device training of target distribution transforming as auxiliary data, Fig. 3 shows diagnosis comparative situation.It can from figure
To find out, if auxiliary distribution transforming is chosen in f1-f8, with the increase of auxiliary distribution transforming, the diagnosis of various types failure is accurate
Increased trend is presented in degree on the whole；If diagnostic accuracy can be and using the fault data of s1-s6 as auxiliary data
Decline, the generalization ability of diagnosis are weakened.This illustrates that the fault diagnosis of distribution transforming is not only related with the quantity of auxiliary data, equally
It is decided by the similarity degree of auxiliary data and target data, further demonstrates in distribution transforming transfer learning fault diagnosis device training
Before, data are screened by comprehensive similarity necessity.
As seen through the above analysis, the selection of distribution transforming fault diagnosis precision and auxiliary data has close relationship,
High quality, the auxiliary data of multi-quantity are conducive to improve the precision of diagnosis.In the selection of auxiliary data, need to define mobilance
Therefore threshold value δ, the data for taking comprehensive similarity to be greater than mobilance threshold value, select different δ to auxiliary data as auxiliary data
It is chosen, studies the fault diagnosis device precision under different δ, calculated result is as shown in Figure 4.
For analysis target data amount and influence of the auxiliary data amount to distribution transforming fault diagnosis model, different data amount is had chosen
M1 is trained with the fault data of ratio, the precision of fault diagnosis is as shown in Figure 5.It can be seen from the figure that target data
With auxiliary data quantity ratio be 1 when, the diagnostic accuracy of M1 is worst；When ratio is less than 1, the diagnostic accuracy of M1 is with ratio
Increase and be reduced rapidly, this is because the generalization ability of distribution transforming fault diagnosis device constantly weakens with the reduction of auxiliary data, leads
Cause the precision of diagnostor worse and worse；When ratio is greater than 1, the reduction of auxiliary data may consequently contribute to improve diagnostic accuracy instead,
But amplitude is little, this is because having differences property between auxiliary data and target data, when auxiliary data negligible amounts, auxiliary
The influence of data is weakened, and the precision of fault diagnosis device can increase, but in general, the quantity of auxiliary data generally compares
Target data is more, so this situation has theoretical significance, it is little to the directive significance of practice.Meanwhile working as auxiliary data
One timing of quantity, target data quantity is more, and diagnostic accuracy is higher, this is because when target data quantity is more, by auxiliary
Help the influence of data fewer, robustness is stronger, and diagnostic accuracy can increase.
The mentioned method of the present invention is by constantly adjusting the weight of target data and auxiliary data constantly to distribution transforming fault diagnosis
Model is modified, and studies influence of the number of iterations to M1 fault diagnosis precision below.In this example, M1 target data takes 30
Group, and auxiliary data takes 50 groups, 80 groups, 110 groups, 140 groups, calculates the fault diagnosis precision under different the number of iterations, such as Fig. 6 institute
Show.It can be seen from the figure that the diagnostor convergence rate under various quantity auxiliary datas is not much different at iteration initial stage；And with
The increase of the number of iterations, the few diagnostor of auxiliary data restrain first, attainable diagnostic accuracy is lower, more than auxiliary data
Diagnostor convergence is slower, but can reach better diagnostic accuracy.
The present invention is before carrying out fault diagnosis, first with fuzzy Apriori algorithm to the key of induction distribution transforming failure
Quantity of state is excavated, it is necessary to the necessity of Key state excavation is verified, calculated result is as shown in table 3, and in example, M1 mesh
Mark data take 30 groups, and auxiliary data takes 80 groups, 110 groups, 140 groups respectively.
3 Key state of table excavates the influence to diagnostic accuracy
Specifically, the iterative process are as follows:
A. weight normalizes, and enables
B. neural network algorithm is called, according to T, ptObtain the weak diagnostor h of failuret:X→Y；
C. the weak diagnostor h of failure is calculatedtIn TaError rate above:
In formula, ht(xi) presentation class device is to xiObtained study mark；
D., the weak diagnostor weight parameter α of failure is settAnd target weight adjusting parameter βt
E. weight updates
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (7)
1. a kind of distribution transforming transfer learning method for diagnosing faults for considering multiple factors Situation Evolution, it is characterised in that including following step
It is rapid:
1) dynamic indicator quantity of state, Quasi dynamic index quantity of state and static state are divided into the index quantity of state for influencing distribution transforming operating status
Index quantity of state constructs distribution transforming evaluation of running status index system on this basis；
2) binary quantization is carried out to the index quantity of state in state evaluation index system, calculates these using Apriori algorithm and refers to
The incidence relation between quantity of state and distribution transforming failure is marked, index quantity of state stronger for incidence relation is regarded as induction and matches
The key index quantity of state of accident barrier；
3) it obtains the target distribution transforming for needing to carry out fault diagnosis and possesses pass of the history distribution transforming of failure logging under all kinds of failures
Key index state quantity data calculates the key index state quantity data of target distribution transforming and history distribution transforming using Tanimoto coefficient
Similarity selects the higher history distribution transforming of similarity as auxiliary distribution transforming；
4) using transfer learning algorithm TrAdaBoost to the pass of the key index state quantity data of target distribution transforming and auxiliary distribution transforming
The weight of key index state quantity data is solved, and then obtains distribution transforming fault diagnosis model, to carry out the event of target distribution transforming
Barrier diagnosis.
2. distribution transforming transfer learning fault diagnosis model according to claim 1, which is characterized in that dynamic in the step 1)
Index quantity of state, Quasi dynamic index quantity of state and Static State Index quantity of state are respectively as follows:
A. dynamic indicator quantity of state refers to the data that can be acquired in real time, including meteorological data, electrically acquires data, such index shape
State amount can effecting reaction distribution transforming state dynamic change, be the important assessment source of distribution transforming fault diagnosis；
B. Quasi dynamic index quantity of state refers to the data by periodically or non-periodically obtaining, test and inspection number including transformer；
C. Static State Index quantity of state refers to the information for not needing to obtain by system interaction, including account information, history electricity consumption.
3. distribution transforming transfer learning fault diagnosis model according to claim 1, which is characterized in that shape in the step 2)
Index quantity of state in state assessment indicator system carries out binary quantization method are as follows:
A. the index quantity of state attribute set of distribution transforming is set
P={ p1,p2,...,pM} (1)
Wherein, P indicates that the index quantity of state attribute and failure collection of distribution transforming, element include the dynamic for influencing distribution transforming operating status
Index quantity of state, Quasi dynamic index quantity of state and Static State Index quantity of state；p1,p2,...,pMIndicate index quantity of state attribute set
Element in P indicates some index quantity of state in dynamic indicator quantity of state, or indicates some in Quasi dynamic index quantity of state
Index quantity of state, or indicate some index quantity of state in Static State Index quantity of state；M indicates index quantity of state number；
B. choose certain distribution transforming i, set its failure affairs library as
Di={ d1,d2,...,dN} (2)
Wherein, DiIndicate distribution transforming i failure affairs library, element includes the failure logging of distribution transforming i；d1,d2,...,dNIt indicates to match accident
Hinder affairs library DiIn element, indicate certain failure logging of distribution transforming i, the element in every failure logging corresponds to each element in P
Value and fault type；The failure logging item number of N expression distribution transforming i；
C. " good " and " bad " are divided into according to superiority and inferiority to the element in the index quantity of state attribute of distribution transforming and failure collection P, to DiIn
Any one failure logging is denoted as 1, if certain index quantity of state in the record if certain index quantity of state is " good " in the record
For " bad ", then 0 is denoted as.
4. distribution transforming transfer learning fault diagnosis model according to claim 1, which is characterized in that utilized in the step 2)
Apriori algorithm calculates the incidence relation between these index quantity of states and distribution transforming failure, index stronger for incidence relation
Quantity of state, the method for being regarded as the key index quantity of state of induction distribution transforming failure are as follows:
A. for DiIn any one failure logging, set the index quantity of state attribute of its distribution transforming as in association rule mining
Former piece X, failure are the consequent Y in association rule mining, if certain index quantity of state in failure logging is 1, then it is assumed that this refers to
Mark quantity of state exists with the fault type of this failure logging to be associated with, if certain index quantity of state in failure logging is 0, then it is assumed that
The index quantity of state, there is no being associated with, chooses suitable support threshold with the fault type of this failure logging, extracts and supports
Degree is greater than the correlation rule of support threshold
Wherein, the support calculation formula of correlation rule X → Y is
Wherein, s (X → Y) indicates the support of correlation rule X → Y；The support counting of σ (X → Y) expression correlation rule X → Y；
The failure logging item number of N expression distribution transforming i；
B. suitable confidence threshold value is chosen, the correlation rule that confidence level is greater than confidence threshold value is extracted, correlation rule X → Y's
Confidence calculations formula is
Wherein, c (X → Y) indicates the confidence level of correlation rule X → Y；The support counting of σ (X → Y) expression correlation rule X → Y；
σ (X) indicates the support counting of former piece X in correlation rule.
C. choose while meeting the correlation rule that support is greater than confidence threshold value greater than support threshold, confidence level；
D. all to distribution transforming i at the same meet support greater than support threshold, confidence level be greater than confidence threshold value correlation rule
In index state measure union, constitute distribution transforming i key index state duration set Ii；
E. I is takeniUnion, constitute induction transformer fault key index quantity of state I
Wherein, Z indication transformer number of units；IiIndicate the key index state duration set extracted from i-th distribution transforming；I expression is used for
The index index state duration set of distribution transforming fault diagnosis.
5. distribution transforming transfer learning fault diagnosis model according to claim 1, which is characterized in that utilized in the step 3)
Tanimoto coefficient calculates the similarity of the key index state quantity data of target distribution transforming and history distribution transforming, selects similarity higher
History distribution transforming as auxiliary distribution transforming method are as follows:
A. the failure distribution proportion for defining target distribution transforming and history distribution transforming is as follows:
In formula, R indicates fault type number；Indicate the failure r in target distribution transforming in the faulty middle proportion of institute；Expression is gone through
Failure r in history distribution transforming is in the faulty middle proportion of institute；PaIndicate the defective proportion set of target distribution transforming；PbIndicate that history is matched
The defective proportion set of change.
B. set the key index state duration set of target distribution transforming and history distribution transforming asWith
In formula, fi aIndicate the per unit value of the key index quantity of state of target distribution transforming；fi bIndicate the key index state of history distribution transforming
The per unit value of amount；mrIndicate the quantity of failure r in target distribution transforming；nrIndicate the quantity of failure r in history distribution transforming.
C. rightWithIn vector average processing:
In formula,It indicatesSet after handling averagely；It indicatesSet after handling averagely.
D. Tanimoto coefficient is introduced, the similarity of target distribution transforming and history distribution transforming failure r is obtained:
In formula,It indicates target distribution transforming and assists the similarity of distribution transforming failure r.
E. combination failure distribution proportion obtains the faulty comprehensive similarity of target distribution transforming and history distribution transforming institute
In formula, TRIndicate the faulty comprehensive similarity of target distribution transforming and history distribution transforming institute.
F. mobilance threshold value δ is defined, history distribution transforming is screened, if TRThis history distribution transforming is then considered as auxiliary distribution transforming by >=δ,
Key index state quantity data will be used for next target distribution transforming fault diagnosis；Conversely, if TR< δ then abandons this history and matches
Change and the key index state quantity data corresponding to it.
6. distribution transforming transfer learning fault diagnosis model according to claim 1, which is characterized in that utilized in the step 4)
Transfer learning algorithm TrAdaBoost is to the key index state quantity data of target distribution transforming and the key index state of auxiliary distribution transforming
The weight of amount data is solved, and then obtains distribution transforming fault diagnosis model, the specific steps are that:
A. the failure logging of target distribution transforming is set as Ta, TaAs target training gather, set assist distribution transforming failure logging as
Tb, TbAs supplemental training set；
B. T is setaSample size is m, TbSample size is n, merges training set T=Ta∪Tb, the number of iterations Iter, basic classification
Algorithm of neural network, wherein
In formula, x is the per unit value of distribution transforming key index quantity of state, and y is fault type；
C. weight vectors are initializedWherein
D. initiation parameter
E. start iteration, iteration t=1,2 ..., Iter
F. fault diagnosis model is exported, the output of the model is the fault diagnosis result of target distribution transforming
7. distribution transforming transfer learning fault diagnosis model according to claim 6, which is characterized in that iteration in the step e
Process are as follows:
A. weight normalizes, and enables
B. neural network algorithm is called, according to T, ptObtain the weak diagnostor h of failuret:X→Y；
C. the weak diagnostor h of failure is calculatedtIn TaError rate above:
In formula, ht(xi) presentation class device is to xiObtained study mark；
D., the weak diagnostor weight parameter α of failure is settAnd target weight adjusting parameter βt
E. weight updates
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811396356.6A CN109254219B (en) | 2018-11-22 | 2018-11-22 | A kind of distribution transforming transfer learning method for diagnosing faults considering multiple factors Situation Evolution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811396356.6A CN109254219B (en) | 2018-11-22 | 2018-11-22 | A kind of distribution transforming transfer learning method for diagnosing faults considering multiple factors Situation Evolution |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109254219A true CN109254219A (en) | 2019-01-22 |
| CN109254219B CN109254219B (en) | 2019-10-22 |
Family
ID=65042065
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811396356.6A Active CN109254219B (en) | 2018-11-22 | 2018-11-22 | A kind of distribution transforming transfer learning method for diagnosing faults considering multiple factors Situation Evolution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109254219B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110275879A (en) * | 2019-05-16 | 2019-09-24 | 浙江浙能技术研究院有限公司 | A method of Trouble Match and early warning are carried out based on fault data state matrix |
| CN110378872A (en) * | 2019-06-10 | 2019-10-25 | 河海大学 | A kind of multi-source adaptive equalization transfer learning method towards crack image detection |
| CN110780146A (en) * | 2019-12-10 | 2020-02-11 | 武汉大学 | Transformer fault identification and positioning diagnosis method based on multi-stage transfer learning |
| CN112036451A (en) * | 2020-08-13 | 2020-12-04 | 广东电网有限责任公司 | Novel insulation defect identification method based on transfer learning, computer equipment and storage medium |
| CN112083337A (en) * | 2020-10-22 | 2020-12-15 | 重庆大学 | Power battery health prediction method oriented to predictive operation and maintenance |
| CN118211124A (en) * | 2024-05-22 | 2024-06-18 | 国网山东省电力公司汶上县供电公司 | Transformer fault prediction method and system |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103871003A (en) * | 2014-03-31 | 2014-06-18 | 国家电网公司 | Power distribution network fault diagnosis method utilizing historical fault data |
| CN104267270A (en) * | 2014-08-06 | 2015-01-07 | 中国南方电网有限责任公司超高压输电公司检修试验中心 | Transformer key parameter extraction method based on vector similarity |
| CN105300693A (en) * | 2015-09-25 | 2016-02-03 | 东南大学 | Bearing fault diagnosis method based on transfer learning |
| CN105891629A (en) * | 2016-03-31 | 2016-08-24 | 广西电网有限责任公司电力科学研究院 | Transformer equipment fault identification method |
| CN106019084A (en) * | 2016-06-16 | 2016-10-12 | 上海交通大学 | Power distribution and utilization data association-based medium-voltage power grid line fracture fault diagnosis method |
| CN106526352A (en) * | 2016-09-30 | 2017-03-22 | 中国电力科学研究院 | Method and system for determining power transformer fault types |
| CN108629528A (en) * | 2018-06-20 | 2018-10-09 | 国网江苏省电力有限公司电力科学研究院 | Quality of Transformer problem analysis method based on Apriori algorithm |
-
2018
- 2018-11-22 CN CN201811396356.6A patent/CN109254219B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103871003A (en) * | 2014-03-31 | 2014-06-18 | 国家电网公司 | Power distribution network fault diagnosis method utilizing historical fault data |
| CN104267270A (en) * | 2014-08-06 | 2015-01-07 | 中国南方电网有限责任公司超高压输电公司检修试验中心 | Transformer key parameter extraction method based on vector similarity |
| CN105300693A (en) * | 2015-09-25 | 2016-02-03 | 东南大学 | Bearing fault diagnosis method based on transfer learning |
| CN105891629A (en) * | 2016-03-31 | 2016-08-24 | 广西电网有限责任公司电力科学研究院 | Transformer equipment fault identification method |
| CN106019084A (en) * | 2016-06-16 | 2016-10-12 | 上海交通大学 | Power distribution and utilization data association-based medium-voltage power grid line fracture fault diagnosis method |
| CN106526352A (en) * | 2016-09-30 | 2017-03-22 | 中国电力科学研究院 | Method and system for determining power transformer fault types |
| CN108629528A (en) * | 2018-06-20 | 2018-10-09 | 国网江苏省电力有限公司电力科学研究院 | Quality of Transformer problem analysis method based on Apriori algorithm |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110275879A (en) * | 2019-05-16 | 2019-09-24 | 浙江浙能技术研究院有限公司 | A method of Trouble Match and early warning are carried out based on fault data state matrix |
| CN110378872A (en) * | 2019-06-10 | 2019-10-25 | 河海大学 | A kind of multi-source adaptive equalization transfer learning method towards crack image detection |
| CN110780146A (en) * | 2019-12-10 | 2020-02-11 | 武汉大学 | Transformer fault identification and positioning diagnosis method based on multi-stage transfer learning |
| CN112036451A (en) * | 2020-08-13 | 2020-12-04 | 广东电网有限责任公司 | Novel insulation defect identification method based on transfer learning, computer equipment and storage medium |
| CN112083337A (en) * | 2020-10-22 | 2020-12-15 | 重庆大学 | Power battery health prediction method oriented to predictive operation and maintenance |
| CN118211124A (en) * | 2024-05-22 | 2024-06-18 | 国网山东省电力公司汶上县供电公司 | Transformer fault prediction method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109254219B (en) | 2019-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109254219B (en) | A kind of distribution transforming transfer learning method for diagnosing faults considering multiple factors Situation Evolution | |
| CN111985561B (en) | Fault diagnosis method and system for intelligent electric meter and electronic device | |
| CN107463633B (en) | A kind of real time data rejecting outliers method based on EEMD- neural network | |
| US11740619B2 (en) | Malfunction early-warning method for production logistics delivery equipment | |
| CN110929918B (en) | 10kV feeder fault prediction method based on CNN and LightGBM | |
| CN106779505A (en) | A kind of transmission line malfunction method for early warning driven based on big data and system | |
| CN109103874B (en) | Power distribution network reliability evaluation method considering load characteristics and distributed power access | |
| CN112016251B (en) | Nuclear power device fault diagnosis method and system | |
| CN105302848A (en) | Evaluation value calibration method of equipment intelligent early warning system | |
| CN109858140B (en) | Fault diagnosis method for water chilling unit based on information entropy discrete Bayesian network | |
| CN105956788A (en) | Dynamic management control method for cost of power transmission and transformation project | |
| CN108931709A (en) | A kind of distribution network fault location method of high serious forgiveness | |
| CN106569030A (en) | Alarm threshold optimizing method and device in electric energy metering abnormity diagnosis | |
| CN105608842A (en) | Nuclear reactor fuel failure online monitoring alarm device | |
| CN106845526A (en) | A kind of relevant parameter Fault Classification based on the analysis of big data Fusion of Clustering | |
| CN112529327A (en) | Method for constructing fire risk prediction grade model of buildings in commercial areas | |
| CN104268375A (en) | Ship electric power station fault diagnosing method based on knowledge petri network | |
| CN107677903A (en) | A kind of clustering method of Transformer's Condition Monitoring data | |
| CN108520267A (en) | A kind of hydrological telemetering data exception detection method based on space-time characteristic | |
| CN114519923A (en) | Intelligent diagnosis and early warning method and system for power plant | |
| Gu et al. | Singular value diagnosis in dam safety monitoring effect values | |
| CN105930900A (en) | Method and system for predicting hybrid wind power generation | |
| Sun et al. | An ensemble system to predict the spatiotemporal distribution of energy security weaknesses in transmission networks | |
| CN116432123A (en) | Electric energy meter fault early warning method based on CART decision tree algorithm | |
| CN110956281A (en) | Power equipment abnormity detection alarm system based on Log analysis |
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 |