CN109308522A - A kind of GIS failure prediction method based on Recognition with Recurrent Neural Network - Google Patents
A kind of GIS failure prediction method based on Recognition with Recurrent Neural Network Download PDFInfo
- Publication number
- CN109308522A CN109308522A CN201811018590.5A CN201811018590A CN109308522A CN 109308522 A CN109308522 A CN 109308522A CN 201811018590 A CN201811018590 A CN 201811018590A CN 109308522 A CN109308522 A CN 109308522A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- gis
- failure
- time
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The present invention relates to a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network.Including data processing module and Recognition with Recurrent Neural Network identification module.In data processing module, first the GIS historical parameter data of longer period of time collected is handled, reuses the method for mathematical function assignment to construct training and test sample to training circulation neural model.In Recognition with Recurrent Neural Network identification module, building circulation neural model structure first, then outlier detection is carried out using circulation neural model, it is labelled for sample data, sample data after reusing determining label is trained circulation neural model, model parameter is adjusted, using revised circulation neural model as GIS fault prediction model；Recently enter test data, the probability and fault type that prediction output GIS future may break down.
Description
Technical field
The present invention relates to a kind of GIS failure prediction methods of Recognition with Recurrent Neural Network, belong to preassembled transformer station technology neck
Domain.
Background technique
In recent years, along with the rising of the raising of voltage class and electricity consumption, requirement of the people for power supply reliability has
It is further to improve.Gas insulation switch cabinet GIS (Gas Insulated Switchgear) is small, reliable because of its occupied area
Property high and shock resistance it is strong the features such as in distribution link in occupation of an important link.Once GIS breaks down, it will
Cause large-scale power grid to be paralysed, causes significant impact to national daily life and necessary industrial production.Failure GIS's
Maintenance not only needs interruption maintenance and removes external shell to repair, and can also spend longer maintenance time, it will to its people
Economy further results in great loss.Can effectively GIS failure be checked by carrying out failure predication in time, in GIS failure
Danger is released before jeopardizing power grid.In the case, how to carry out failure predication before GIS device failure occurs becomes
The demand of safe distribution of electric power.GIS failure occur before quickly and accurately to fault type and probability of happening judged for
Power off time is reduced, improving maintenance efficiency and utilization rate of equipment and installations has vital meaning.
From tradition modeling or low volume data training it is different, deep learning can be excavated by the training of mass data with
Labyrinth inside learning data, i.e. feature extraction.Concrete application can be abstracted as using initial data, signal as low layer
Secondary feature is sent into neural network, the process that the information that people are wanted is indicated as high level.The main knot of deep learning
Structure includes convolutional neural networks (CNN), depth confidence network (DBN) and Recognition with Recurrent Neural Network (RNN) etc..Pass through sliding with CNN
Window retains partial history input difference, and the Recognition with Recurrent Neural Network (RNN) used herein can retain all history inputs, lead to
The association between abstract history input is crossed, there is very high classification accuracy.Also, RNN network universality is good, to various defeated
Entering signal has good recognition effect, is widely used on solving the problems, such as time series correlation.RNN at present
Network has been successfully applied in the fields such as natural language processing, computer vision, disease forecasting, achieves good effect.
Failure predication is similar to Language Processing and disease forecasting, and the sensor signal of various equipment is mostly time series, current time
Signal and historical data have it is indivisible contact, therefore can satisfy the demand of failure predication just using RNN network.
Summary of the invention
The technology of the present invention solves the problems, such as: for the missing of existing GIS failure predication technology, proposing a kind of based on circulation mind
GIS failure prediction method through network, fills up the vacancy of the prior art.
The technology of the present invention solution: a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network, including at data
Manage module and Recognition with Recurrent Neural Network identification module.Training is established using the method for mathematical function assignment in data processing module
Sample, in Recognition with Recurrent Neural Network identification module, building recycles neural model structure first, then using circulation neural model
Outlier detection is carried out, is that training sample data are labelled, reuses the training sample of data processing module building to circulation
Neural model is trained, and is adjusted model parameter, is determined GIS fault prediction model.Finally enter data into circulation neural model
Failure predication is carried out to GIS.Specifically includes the following steps:
1) acquire vibration on GIS sensor in the long period, air pressure, three kinds of signals of electric current data, as history number
According to.
2) in data processing module, trained and test sample matrix is established using the method for mathematical function assignment.
3) in Recognition with Recurrent Neural Network identification module, it is first determined circulation neural model structure reuses the neural mould of circulation
Type carries out outlier detection, is that sample data is labelled, then inputs training sample and adjusts model parameter, finally to following
Ring neural model input test data, obtain prediction result.
Above-mentioned steps 1) in, collect GIS sensor longer period of time, i.e., 183 days or more GIS vibration, air pressure and
Current data is summarized.
Above-mentioned steps 2) in data processing module, comprising the following steps:
21) historical data of the vibration of extraction, air pressure and current signal is sampled as unit of day；
22) the training sample data of training circulation neural model are waited for by the method construct of mathematical function assignment, and are used
Same method constructs test data；
Above-mentioned steps 22) in, the method for sample matrix is constructed with mathematical function assignment are as follows: acquire same signal
The fault waveform of two different sensors extracts several samplings in two fault waveform function a (1, t) and a (2, t) respectively
Point is combined into the matrix b of one five dimension.Matrix has 180 column to respectively represent 180 samples.Matrix is as follows:
Wherein b is input matrix, and a (1, t) and a (2, t) are respectively two fault waveform functions, and t is the cross of historical data
Coordinate, that is, number of days.
Above-mentioned steps 3) in specifically includes the following steps:
31) circulation neural model specific structure, including implicit number of layers and every layer of neuron number are determined, and is put up
Recycle neural model；The circulation neural model of building, which uses, has 1 input layer, 4 hidden layers and 1 output layer, Mei Geyin
It is 10 in layer containing the neuron number for including；
32) outlier detection is carried out using the circulation neural model put up, is that training sample data are labelled；
33) it is trained by using the training sample data input circulation neural model after labelled, makes it
Self-optimizing model parameter, using the circulation neural model after the completion of training as GIS fault prediction model；
34) it using current GIS signal data as input data, inputs in GIS fault prediction model, it is pre- by GIS failure
It surveys model to learn input data, finally obtains the probability and fault-signal class of the possible hidden failure of GIS under current time
Type；
The method of outlier detection are as follows: detect that fault-signal starts to become when fault-signal has small variation
The point A of change, referred to as fault-signal starting point then detect the point B that failure really occurs, and are that the period of T=B-A is corresponding
Sample stick the label of [0.1-1].
Above-mentioned outlier detection step are as follows:
321) abnormal point moment v is rule of thumb chosen first, and v value is partially late in time；
322) enter data into circulation neural model and carry out training for the first time, reserve v value and fault trend starting point A it
Between a period of time allowance M data be not put into circulation neural model be trained.It is assumed that the value chosen is v1；
323) by failure from the output valve P assignment of abnormal moment v1, and the variation of P value and the variation of physical fault are enabled
Trend is identical；
324) data that moment B occurs from abnormal point to failure are assigned to using function P (t), wherein fault model time domain
T=B-v1；
Wherein, independent variable t indicates the time since failure trend, changes to T from 0, and as t=T, failure is sent out
It is raw, therefore T is the time occurred from discovery fault trend to failure；
325) it is put into same data test after the completion of training for the first time and observes output valve；
326) fault trend before the neural v1 for having been detected by first time selection at this time of circulation in △ T time section, enables
N is the data number that fault trend is detected in △ T time section, in △ T time section, faulty trend in the unit time
Data numberMeetWhen (u therein is proportionality coefficient and u ∈ (0,1)), v2=v1- △ T is enabled, v2 has compared at this time
V1 is closer to fault trend starting point A；
327) label data is trained network again, repeats the above process, untilOr data in the M period
All training finishes；
328) in unit time faulty trend data numberOr data all train when, stop update v
Value, at the time of v value at this time has been infinitely close to fault trend starting point A.
Above-mentioned steps 324) in, final data carry out assignment using formula (1), due in the abnormal point v guessed before
Be it is partially late, so P value wants larger, formula (1) is modified by formula (2):
Wherein independent variable t indicates the time since failure trend, and at the time of B indicates that failure occurs, v is indicated most
The v value updated eventually.
Above-mentioned steps 328) in, judge the method with fault trend starting point are as follows: by the output valve Pd at time point undetermined
It compares with output average value when equipment health status, is judged according to formula (3):
Wherein Pd is point to be located output valve, and Pk is k moment health status output valve, and s is proportionality coefficient, and s is set as
Value slightly larger than 1, n are data number.
Above-mentioned steps 33) in, the training process of neural model is recycled, the specific steps are that:
331) initialize: random initializtion recycles neural model parameter, including three weight matrix U, W, V and two are partially
Set matrix b and c；
332) forward-propagating: training sample data are inputted into circulation neural model, by Positive Propagation Algorithm, are obtained just
The predicted value that neural model is recycled under beginning model parameter, adjusts model parameter as difference for the label with training sample.Specifically
Steps are as follows:
3321) the hidden state h of model when calculating moment t(t), h(t)It can be by inputting x(t)With the hidden state of last moment
h(t-1)It obtains, formula is as follows:
h(t)=σ (Ux(t)+Wh(t-1)+b) (5)
Wherein activation primitive σ is generally tanh, and bias matrix b is the bias of linear relationship, and weight matrix U, W are circulations
The linear relationship parameter of neural model.x(t)Represent the input of the training sample in moment t, h(t-1)Represent t-1 moment model
Hidden state；
3322) the calculated hidden state h of above-mentioned formula is used(t)Come the output o of model when calculating moment t(t), formula
It is as follows:
o(t)=Vh(t)+c (4)
Wherein weight matrix V and bias matrix c is circulation neural model parameter.
333) backpropagation: to circulation neural model carry out backpropagation calculating, by before output and sample mark
Label, which compare, calculates error, to further be repaired using gradient descent method iteration for model parameter according to error
Just, adjustment circulation neural model parameter, including three weight matrix U, W, V and two bias matrixes b and c.Specific steps are such as
Under:
3331) defining final loss is L, andWherein L(t)For loss function, t represents the moment, and τ is then
Represent the final moment；
3332) weight matrix V and bias matrix c is calculated, specific formula is as follows:
3333) weight matrix W, U and bias matrix b are calculated, formula difference is as follows:
Wherein δ(t)The gradient of the hidden state of the position t is represented, function diag expression takes matrix diagonals element,Represent t
The prediction at moment exports, y(t)Represent the reality output of t moment sample.
334) iterate determining final argument: after the model parameter after being adjusted, the determination of final argument is specific
Steps are as follows:
3341) the circulation neural model after adjusting parameter is re-entered using identical training sample；
3342) error between comparison output result and sample label；
3343) if error is met the requirements, it is determined that model parameter；
It is unsatisfactory for requiring 3344) if difference is still larger, repeatedly step 332) and step 333) adjust parameter
It is whole, until error reaches requirement；
3345) determine that final model parameter includes: weight matrix U, W, V and bias matrix b and c, and by following at this time
Ring neural model is as GIS fault prediction model.
Above-mentioned steps 34) in final prediction output form it is as follows:
Wherein three row a, b, c of matrix are used to occurring as a result, also representing current GIS failure simultaneously for intermediate scheme identification
Probability, the value of a, b, c is more easy to happen closer to 1 expression failure, and i represents the state of GIS at different times and at that time
The result of GIS failure predication.When failure occurs, every kind is as follows containing abnormal Signal coding:
| Contain abnormal signal | abc |
| Current anomaly | 001 |
| Pressure abnormity | 010 |
| Abnormal vibration | 100 |
Advantageous effects of the invention:
A kind of GIS failure prediction method based on Recognition with Recurrent Neural Network proposed by the invention, contains vibration, air pressure
The failure caused with the exception of three kinds of unlike signals of electric current.It is made full use of using the method for outlier detection and is easy to acquirement
A large amount of unlabeled exemplars, and it is labelled to its, there is the phenomenon that exemplar difficulty is high, spends human and material resources to previous acquisition
It is resolved.Also, for noise signal adjoint in sensor actual application, can be had using Recognition with Recurrent Neural Network
The noise signal of sensor when effect identification operates normally, and noise signal is filtered makes it not and influence failure predication result.
Using Recognition with Recurrent Neural Network for the capability of fitting of time series data, in conjunction with sensor current demand signal and historical data for
The failure of GIS device is predicted, may be implemented to make full use of measurement data, improves prediction accuracy.Entire method tool
There are stronger adaptability and generalization ability, there is certain social value and realistic meaning.
Detailed description of the invention:
Fig. 1 is implementation flow chart of the present invention；
Fig. 2 is circulation neural model training process figure.
Specific embodiment:
The invention will be further described below in conjunction with the accompanying drawings.Following embodiment is only used for clearly illustrating this hair
Bright technical solution, and not intended to limit the protection scope of the present invention.
As shown in Figure 1, a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network, comprising the following steps:
Step 1: the data of three kinds of vibration, air pressure, electric current signals on acquisition GIS sensor in the long period, as going through
History data.
Step 2: in data processing module, trained and test sample matrix is established using the method for mathematical function assignment,
Wherein, data processing module includes the following steps；
1) historical data of the vibration of extraction, air pressure and current signal is sampled as unit of day；
2) the training sample data of training circulation neural model are waited for by the method construct of mathematical function assignment, and are used
Identical method construct test data.The method of mathematical function assignment is as follows:
The fault waveform for acquiring two different sensors of same signal, in two fault waveform function a (1, t) and a
The matrix b that several groups of samples synthesize one five dimension is extracted in (2, t) respectively.Matrix has 180 column to respectively represent 180 samples.
Matrix is as follows:
Wherein b is input matrix, and a (1, t) and a (2, t) are respectively two fault waveform functions, and t is the cross of historical data
Coordinate, that is, number of days.
Step 3 is in Recognition with Recurrent Neural Network identification module, it is first determined circulation neural model structure reuses circulation mind
Outlier detection is carried out through model, is that sample data is labelled, inputs training sample then to adjust model parameter.Specifically
The following steps are included:
1) circulation neural model specific structure, including implicit number of layers and every layer of neuron number are determined, and puts up and follows
Ring neural model.The circulation neural model of building, which uses, has 1 input layer, 4 hidden layers and 1 output layer, each implicit
The neuron number for including in layer is 10.
2) outlier detection is carried out using the circulation neural model put up, is that training sample data are labelled, label
Respectively correspond vibration signal failure, air pressure signal failure and current signal failure and normal.Abnormal point detecting method are as follows: in event
Barrier signal has the point A for detecting that fault-signal starts variation when small variation, referred to as fault trend starting point, connects
The detection point B that really occurs of failure, the label of [0.1-1] is sticked for the period corresponding sample of T=B-A.Abnormal point
Survey method and step are as follows:
21) abnormal point moment v is rule of thumb chosen first, and v value is partially late in time；
22) it enters data into circulation neural model and carries out training for the first time, reserve between v value and fault trend starting point A
A period of time allowance M data be not put into circulation neural model be trained.It is assumed that the value chosen is v1；
23) by failure from the output valve P assignment of abnormal moment v1, and the variation of P value and the variation of physical fault is enabled to become
Gesture is identical；
24) data that moment B occurs from abnormal point to failure are assigned to using function P (t), wherein fault model time domain T
=B-v1；
Wherein, independent variable t indicates the time since failure trend, changes to T from 0, and as t=T, failure is sent out
It is raw, therefore T is the time occurred from discovery fault trend to failure.Also, formula (1) is modified by formula (2):
Wherein independent variable t indicates the time since failure trend, and at the time of B indicates that failure occurs, v is indicated most
The v value updated eventually；
25) it is put into same data test after the completion of training for the first time and observes output valve；
26) fault trend before the neural v1 for having been detected by first time selection at this time of circulation in △ T time section, enables n
For the data number for detecting fault trend in △ T time section, in △ T time section, the number of faulty trend in the unit time
According to numberMeetWhen (u therein is proportionality coefficient and u ∈ (0,1)), v2=v1- △ T is enabled, v2 has compared v1 at this time
Closer to fault trend starting point A；
27) label data is trained network again, repeats the above process, untilOr data are complete in the M period
Portion's training finishes；
28) in unit time faulty trend data numberOr data all train when, stop update v value,
At the time of v value at this time has been infinitely close to fault trend starting point A.The method for judging that there is fault trend starting point are as follows:
The output valve Pd at time point undetermined is compared with output average value when equipment health status, is judged according to formula (3):
Wherein Pd is point to be located output valve, and Pk is k moment health status output valve, and s is proportionality coefficient, and s is set as
Value slightly larger than 1, n are data number.
3) it is trained by using the training sample data input circulation neural model after labelled, makes it
Self-optimizing model parameter, using the circulation neural model after the completion of training as GIS fault prediction model, the specific steps are as follows:
31) initialize: random initializtion recycles neural model parameter, including three weight matrix U, W, V and two biasings
Matrix b and c；
32) forward-propagating: inputting circulation neural model for training sample data, by Positive Propagation Algorithm, obtains initial
The predicted value that neural model is recycled under model parameter, adjusts model parameter as difference for the label with training sample.Specific step
It is rapid as follows:
321) hidden state of model when calculating moment t(t),(t)It can be by inputting x(t)With the hidden state of last moment(t-1)
It obtains, formula is as follows:
(t)=σ (Ux(t)+W(t-1)+b) (5)
Wherein activation primitive σ is generally tanh, and bias matrix b is the bias of linear relationship, and weight matrix U, W are circulations
The linear relationship parameter of neural model.x(t)The input of the training sample in moment t is represented,(t-1)Represent the hidden of t-1 moment model
Hiding state；
322) the calculated hidden state of above-mentioned formula is used(t)Come the output o of model when calculating moment t(t), formula is such as
Under:
o(t)=V(t)+c (4)
Wherein weight matrix V and bias matrix c is circulation neural model parameter.
33) backpropagation: to circulation neural model carry out backpropagation calculating, by before output and sample mark
Label, which compare, calculates error, to further be repaired using gradient descent method iteration for model parameter according to error
Just, adjustment circulation neural model parameter, including three weight matrix U, W, V and two bias matrixes b and c.Specific steps are such as
Under:
331) defining final loss is L, andWherein L(t)For loss function, t represents the moment, and τ is then
Represent the final moment.
332) weight matrix V and bias matrix c is calculated, specific formula is as follows:
333) weight matrix W, U and bias matrix b are calculated, formula difference is as follows:
Wherein δ(t)The gradient of the hidden state of the position t is represented, function diag expression takes matrix diagonals element,Represent t
The prediction at moment exports, y(t)Represent the reality output of t moment sample.
34) iterate determining final argument: iterate determining final argument: the model parameter after being adjusted
Afterwards, specific step is as follows for the determination of final argument:
341) the circulation neural model after adjusting parameter is re-entered using identical training sample；
342) error between comparison output result and sample label；
343) if error is met the requirements, it is determined that model parameter；
It is unsatisfactory for requiring 344) if difference is still larger, repeatedly step 332) and step 333) are adjusted parameter,
Until error reaches requirement；
345) determine that final model parameter includes: weight matrix U, W, V and bias matrix b and c, and by following at this time
Ring neural model is as GIS fault prediction model.
4) it using current GIS signal data as input data, inputs in GIS fault prediction model, it is pre- by GIS failure
It surveys model to learn input data, finally obtains the probability and fault-signal class of the possible hidden failure of GIS under current time
Type.Final prediction output form is as follows:
Wherein three row a, b, c of matrix are used to occurring as a result, also representing current GIS failure simultaneously for intermediate scheme identification
Probability, the value of a, b, c is more easy to happen closer to 1 expression failure, and i represents the state of GIS at different times and at that time
The result of GIS failure predication.When failure occurs, every kind is as follows containing abnormal Signal coding:
Step 4 pair recycles neural model input data, obtains prediction result.
Embodiment 1
Certain power plant, including vibration signal, current signal and air pressure signal are derived from using input signal.
The data that preceding 5 chronomeres occur for failure carry out assignment, after the data after assignment are put into network training, weight
All data are newly put into network test, obtained result is as follows:
| Time | 131 | 132 | 133 | 134 | 135 | 136 | 137 | 138 | 139 | 140 |
| a | 0.100 | 0.035 | 0.032 | 0.043 | 0.272 | 0.401 | 0.544 | 0.999 | 0.924 | 0.910 |
| c | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| b | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
The output mean value of health status is 0.0294, takes proportionality coefficient s=1.3, it can be seen that the time is 131~135
Interior, five data of GIS failure predication value a then mark the time in 131~140 data there are three formula (3) are met again
Label, are put into network training, all data are put into network test again, as a result as follows:
| Time | 126 | 127 | 128 | 129 | 130 | 131 | 132 | 133 | 134 | 135 |
| a | 0.070 | 0.369 | 0.200 | 0.062 | 0.704 | 0.354 | 0.297 | 0.378 | 0.451 | 0.559 |
| b | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| c | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
By data inputting MATLAB and draw image.As t=120, failure does not occur, but small signal intensity is opened
Begin, failure occurs when t=140, and the signal containing exception information is vibration signal.
Above example is provided just for the sake of the description purpose of the present invention, and is not intended to limit the scope of the invention.This hair
Bright range is defined by the following claims.The various equivalent alterations and modifications that spirit and principles of the present invention are made are not departed from,
It should all cover within the scope of the present invention.
Claims (10)
1. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network, which is characterized in that utilize data processing module and circulation
Neural network recognization module；Comprising the following specific steps
1) acquire vibration on GIS sensor in the long period, air pressure, three kinds of signals of electric current data, as historical data；
2) in data processing module, trained and test sample matrix is established using the method for mathematical function assignment；
3) in Recognition with Recurrent Neural Network identification module, it is first determined circulation neural model structure, reuse circulation neural model into
Row outlier detection, is that sample data is labelled, then inputs training sample and is adjusted to model parameter, finally to following
Ring neural model input test data, obtain prediction result.
2. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 1, it is characterised in that: step
It is rapid 1) in, collect GIS sensor longer period of time, i.e., GIS vibration in 183 days or more, air pressure and current data are converged
Always.
3. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 1, it is characterised in that: step
It is rapid 2) in, the method for using mathematical function assignment is established trained with test sample matrix, comprising the following steps:
21) historical data of the vibration of extraction, air pressure and current signal is sampled as unit of day；
22) the training sample data of training circulation neural model are waited for by the method construct of mathematical function assignment, and using same
Method construct test data.
4. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 3, it is characterised in that: step
It is rapid 22) in, the method for sample matrix is constructed with mathematical function assignment are as follows: acquire two different sensors of same signal
Fault waveform extracts several groups of samples respectively in two fault waveform function a (1, t) and a (2, t) and synthesizes one five dimension
Matrix b；Matrix has 180 column to respectively represent 180 samples；Matrix is as follows:
Wherein b is input matrix, and a (1, t) and a (2, t) are respectively two fault waveform functions, and t is the abscissa of historical data
That is number of days.
5. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 1, it is characterised in that: step
It is rapid 3) in comprising the following specific steps
31) circulation neural model specific structure, including implicit number of layers and every layer of neuron number are determined, and puts up circulation mind
Through model；The circulation neural model of building, which uses, has 1 input layer, 4 hidden layers and 1 output layer, in each hidden layer
The neuron number for including is 10；
32) outlier detection is carried out using the circulation neural model put up, is that training sample data are labelled；
33) it is trained by using the training sample data input circulation neural model after labelled, keeps it automatic
Model parameter is adjusted, using the circulation neural model after the completion of training as GIS fault prediction model；
34) it using current GIS signal data as input data, inputs in GIS fault prediction model, passes through GIS failure predication mould
Type learns input data, finally obtains the probability and fault signal type of the possible hidden failure of GIS under current time.
6. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 5, it is characterised in that: step
It is rapid 32) in, the method for outlier detection are as follows: detect that fault-signal starts to change when fault-signal has small variation
Point A, it is the period corresponding sample of T=B-A that referred to as fault trend starting point, which then detects the point B that failure really occurs,
Originally the label of [0.1-1] is sticked；
Outlier detection step are as follows:
321) abnormal point moment v is rule of thumb chosen first, and v value is partially late in time；
322) it enters data into circulation neural model and carries out training for the first time, reserve one between v value and fault trend starting point A
The data of section argin M are not put into circulation neural model and are trained；It is assumed that the value chosen is v1；
323) by failure from the output valve P assignment of abnormal moment v1, and variation tendency phase of the variation of P value with physical fault is enabled
Together；
324) data that moment B occurs from abnormal point to failure are assigned to using function P (t), wherein fault model time domain T=B-
v1；
Wherein, independent variable t indicates the time since failure trend, changes to T from 0, and as t=T, failure occurs, therefore
T is the time occurred from discovery fault trend to failure；
325) it is put into same data test after the completion of training for the first time and observes output valve；
326) fault trend before the neural v1 for having been detected by first time selection at this time of circulation in Δ T time section, enabling n is Δ
The data number that fault trend is detected in T time section, in Δ T time section, the data number of faulty trend in the unit time
MeshMeetWhen, u therein is proportionality coefficient and u ∈ (0,1), enables v2=v1- Δ T, v2 is more closer than v1 at this time
Fault trend starting point A；
327) label data is trained network again, repeats the above process, untilOr data are all instructed in the M period
White silk finishes；
328) in unit time faulty trend data numberOr data all train when, stop update v value, at this time
V value at the time of be infinitely close to fault trend starting point A.
7. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 6, it is characterised in that: step
It is rapid 324) in, final data carry out assignment using formula (1), due to be in the abnormal point v guessed before it is partially late, so P
Value wants larger, is modified by formula (2) to formula (1):
Wherein independent variable t indicates the time since failure trend, and at the time of B indicates that failure occurs, v indicates final updated
V value.
8. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 6, it is characterised in that: step
It is rapid 326) in, judge have fault trend starting point method are as follows: by the output valve Pd and equipment health status at time point undetermined
When output average value compare, judged according to formula (3)；
Wherein Pd is point to be located output valve, and Pk is k moment health status output valve, and s is proportionality coefficient, and s is set as bigger
In 1 value, n is data number.
9. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 5, it is characterised in that: step
It is rapid 33) in, recycle the training process of neural model, the specific steps are that:
331) initialize: random initializtion recycles neural model parameter, including three weight matrix bias matrixes of U, W, V and two
B and c；
332) forward-propagating: training sample data are inputted into circulation neural model, initial model is obtained by Positive Propagation Algorithm
The predicted value that neural model is recycled under parameter, adjusts model parameter as difference for the label with training sample；Specific steps are such as
Under:
3321) the hidden state h of model when calculating moment t(t), h(t)By inputting x(t)With the hidden state h of last moment(t-1)?
It arrives, formula is as follows:
h(t)=σ (Ux(t)+Wh(t-1)+b) (5)
Wherein activation primitive σ is tanh, and bias matrix b is the bias of linear relationship, and weight matrix U, W are circulation neural models
Linear relationship parameter；x(t)Represent the input of the training sample in moment t, h(t-1)Represent the hidden state of t-1 moment model；
3322) the calculated hidden state h of above-mentioned formula is used(t)Come the output o of model when calculating moment t(t), formula is as follows:
o(t)=Vh(t)+c (4)
Wherein weight matrix V and bias matrix c is circulation neural model parameter；
333) backpropagation: to circulation neural model carry out backpropagation calculating, by before output and sample label into
Row comparing calculation goes out error, to further be corrected using gradient descent method iteration for model parameter according to error, adjusts
Whole circulation neural model parameter, including three weight matrix U, W, V and two bias matrixes b and c；Specific step is as follows:
3331) defining final loss is L, andWherein L(t)For loss function, t represents the moment, and τ is then represented most
Terminal hour is carved；
3332) weight matrix V and bias matrix c is calculated, specific formula is as follows:
3333) weight matrix W, U and bias matrix b are calculated, formula difference is as follows:
Wherein δ(t)The gradient of the hidden state of the position t is represented, function diag expression takes matrix diagonals element,Represent t moment
Prediction output, y(t)Represent the reality output of t moment sample；
334) iterate determining final argument: after the model parameter after being adjusted, the determination specific steps of final argument
It is as follows:
3341) the circulation neural model after adjusting parameter is re-entered using identical training sample；
3342) error between comparison output result and sample label；
3343) if error is met the requirements, it is determined that model parameter；
It is unsatisfactory for requiring 3344) if difference is still larger, repeatedly step 332) and step 333) are adjusted parameter, until
Error reaches requirement；
3345) determine that final model parameter includes: weight matrix U, W, V and bias matrix b and c, and circulation at this time is refreshing
Through model as GIS fault prediction model.
10. a kind of GIS failure prediction method based on Recognition with Recurrent Neural Network according to claim 5, it is characterised in that: step
It is rapid 34) in final prediction output form it is as follows:
Wherein three row a, b, c of matrix are used to that intermediate scheme identifies as a result, also representing the general of current GIS failure generation simultaneously
The value of rate, a, b, c is more easy to happen closer to 1 expression failure, and i represents the state of GIS at different times and at that time GIS failure
The result of prediction；When failure occurs, every kind is as follows containing abnormal Signal coding:
Contain abnormal signal: abc；
Current anomaly: 001；
Pressure abnormity: 010；
Abnormal vibration: 100.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811018590.5A CN109308522B (en) | 2018-09-03 | 2018-09-03 | GIS fault prediction method based on recurrent neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811018590.5A CN109308522B (en) | 2018-09-03 | 2018-09-03 | GIS fault prediction method based on recurrent neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109308522A true CN109308522A (en) | 2019-02-05 |
| CN109308522B CN109308522B (en) | 2022-02-22 |
Family
ID=65224525
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811018590.5A Active CN109308522B (en) | 2018-09-03 | 2018-09-03 | GIS fault prediction method based on recurrent neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109308522B (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109872014A (en) * | 2019-03-29 | 2019-06-11 | 永安行科技股份有限公司 | Hydrogen energy source moped day usage amount and fault condition prediction technique |
| CN110210516A (en) * | 2019-04-25 | 2019-09-06 | 合刃科技（武汉）有限公司 | Lathe abnormality recognition method, system and terminal device based on deep learning |
| CN110647924A (en) * | 2019-09-05 | 2020-01-03 | 国网河北省电力有限公司电力科学研究院 | GIS equipment state evaluation method based on support vector description and K-nearest neighbor algorithm |
| CN110737732A (en) * | 2019-10-25 | 2020-01-31 | 广西交通科学研究院有限公司 | electromechanical equipment fault early warning method |
| CN111160454A (en) * | 2019-12-27 | 2020-05-15 | 北京旋极信息技术股份有限公司 | Method and device for detecting speed change signal |
| CN111210231A (en) * | 2019-12-31 | 2020-05-29 | 联想(北京)有限公司 | Information processing method and information processing apparatus |
| CN111865863A (en) * | 2020-07-20 | 2020-10-30 | 山东大学 | RNN neural network-based OFDM signal detection method |
| CN112633550A (en) * | 2020-11-23 | 2021-04-09 | 成都唐源电气股份有限公司 | RNN-based catenary fault trend prediction method, equipment and storage medium |
| CN112650204A (en) * | 2020-12-30 | 2021-04-13 | 中南大学 | Intelligent track unmanned vehicle fault gene identification method and system |
| CN112670996A (en) * | 2020-12-03 | 2021-04-16 | 浙江大学 | CNN-based GPU parallel power flow out-of-limit discrimination method |
| CN112836577A (en) * | 2020-12-30 | 2021-05-25 | 中南大学 | Intelligent traffic unmanned vehicle fault gene diagnosis method and system |
| CN112883639A (en) * | 2021-02-03 | 2021-06-01 | 国网浙江省电力有限公司宁波供电公司 | GIS equipment service life prediction device and method based on machine learning |
| CN112926752A (en) * | 2021-01-26 | 2021-06-08 | 西华大学 | Intelligent power transmission network fault line detection and fault recovery method considering energy storage system |
| CN113866572A (en) * | 2021-09-06 | 2021-12-31 | 西安理工大学 | Direct-current fault arc detection and positioning method under condition of access of multiple power electronic devices |
| CN117372880A (en) * | 2023-12-07 | 2024-01-09 | 天津市祥途测绘科技有限公司 | Road engineering supervision system and method based on remote sensing image |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107769972A (en) * | 2017-10-25 | 2018-03-06 | 武汉大学 | A kind of power telecom network equipment fault Forecasting Methodology based on improved LSTM |
| CN108241298A (en) * | 2018-01-09 | 2018-07-03 | 南京航空航天大学 | A kind of aerogenerator method for diagnosing faults based on FWA-RNN models |
| CN108268935A (en) * | 2018-01-11 | 2018-07-10 | 浙江工业大学 | A kind of PM2.5 concentration values Forecasting Methodology and system based on sequential Recognition with Recurrent Neural Network |
-
2018
- 2018-09-03 CN CN201811018590.5A patent/CN109308522B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107769972A (en) * | 2017-10-25 | 2018-03-06 | 武汉大学 | A kind of power telecom network equipment fault Forecasting Methodology based on improved LSTM |
| CN108241298A (en) * | 2018-01-09 | 2018-07-03 | 南京航空航天大学 | A kind of aerogenerator method for diagnosing faults based on FWA-RNN models |
| CN108268935A (en) * | 2018-01-11 | 2018-07-10 | 浙江工业大学 | A kind of PM2.5 concentration values Forecasting Methodology and system based on sequential Recognition with Recurrent Neural Network |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109872014A (en) * | 2019-03-29 | 2019-06-11 | 永安行科技股份有限公司 | Hydrogen energy source moped day usage amount and fault condition prediction technique |
| CN110210516A (en) * | 2019-04-25 | 2019-09-06 | 合刃科技（武汉）有限公司 | Lathe abnormality recognition method, system and terminal device based on deep learning |
| CN110647924B (en) * | 2019-09-05 | 2022-04-22 | 国网河北省电力有限公司电力科学研究院 | GIS equipment state evaluation method based on support vector description and K-nearest neighbor algorithm |
| CN110647924A (en) * | 2019-09-05 | 2020-01-03 | 国网河北省电力有限公司电力科学研究院 | GIS equipment state evaluation method based on support vector description and K-nearest neighbor algorithm |
| CN110737732A (en) * | 2019-10-25 | 2020-01-31 | 广西交通科学研究院有限公司 | electromechanical equipment fault early warning method |
| CN111160454A (en) * | 2019-12-27 | 2020-05-15 | 北京旋极信息技术股份有限公司 | Method and device for detecting speed change signal |
| CN111160454B (en) * | 2019-12-27 | 2023-12-22 | 北京旋极信息技术股份有限公司 | Quick change signal detection method and device |
| CN111210231A (en) * | 2019-12-31 | 2020-05-29 | 联想(北京)有限公司 | Information processing method and information processing apparatus |
| CN111865863A (en) * | 2020-07-20 | 2020-10-30 | 山东大学 | RNN neural network-based OFDM signal detection method |
| CN112633550A (en) * | 2020-11-23 | 2021-04-09 | 成都唐源电气股份有限公司 | RNN-based catenary fault trend prediction method, equipment and storage medium |
| CN112633550B (en) * | 2020-11-23 | 2023-07-18 | 成都唐源电气股份有限公司 | RNN-based contact network fault trend prediction method, equipment and storage medium |
| CN112670996A (en) * | 2020-12-03 | 2021-04-16 | 浙江大学 | CNN-based GPU parallel power flow out-of-limit discrimination method |
| CN112670996B (en) * | 2020-12-03 | 2022-05-17 | 浙江大学 | CNN-based GPU parallel power flow out-of-limit discrimination method |
| CN112836577A (en) * | 2020-12-30 | 2021-05-25 | 中南大学 | Intelligent traffic unmanned vehicle fault gene diagnosis method and system |
| CN112650204A (en) * | 2020-12-30 | 2021-04-13 | 中南大学 | Intelligent track unmanned vehicle fault gene identification method and system |
| CN112650204B (en) * | 2020-12-30 | 2021-12-24 | 中南大学 | Intelligent track unmanned vehicle fault gene identification method and system |
| CN112836577B (en) * | 2020-12-30 | 2024-02-20 | 中南大学 | Intelligent traffic unmanned vehicle fault gene diagnosis method and system |
| CN112926752A (en) * | 2021-01-26 | 2021-06-08 | 西华大学 | Intelligent power transmission network fault line detection and fault recovery method considering energy storage system |
| CN112926752B (en) * | 2021-01-26 | 2022-07-01 | 西华大学 | Intelligent power transmission network fault line detection and fault recovery method considering energy storage system |
| CN112883639A (en) * | 2021-02-03 | 2021-06-01 | 国网浙江省电力有限公司宁波供电公司 | GIS equipment service life prediction device and method based on machine learning |
| CN113866572A (en) * | 2021-09-06 | 2021-12-31 | 西安理工大学 | Direct-current fault arc detection and positioning method under condition of access of multiple power electronic devices |
| CN117372880A (en) * | 2023-12-07 | 2024-01-09 | 天津市祥途测绘科技有限公司 | Road engineering supervision system and method based on remote sensing image |
| CN117372880B (en) * | 2023-12-07 | 2024-02-09 | 天津市祥途测绘科技有限公司 | Road engineering supervision system and method based on remote sensing image |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109308522B (en) | 2022-02-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109308522A (en) | A kind of GIS failure prediction method based on Recognition with Recurrent Neural Network | |
| Zhou et al. | A novel UKF-RBF method based on adaptive noise factor for fault diagnosis in pumping unit | |
| Kosek et al. | Ensemble regression model-based anomaly detection for cyber-physical intrusion detection in smart grids | |
| CN111474919B (en) | Aeroengine control system sensor fault diagnosis method based on AANN network group | |
| CN110163075A (en) | A kind of multi-information fusion method for diagnosing faults based on Weight Training | |
| CN104237777A (en) | Support vector machine high-voltage circuit breaker fault diagnosis method based on core principal component analysis | |
| CN109596942A (en) | A kind of voltage sag reason recognition methods based on depth confidence network | |
| CN115455746B (en) | Nuclear power device operation monitoring data anomaly detection and correction integrated method | |
| CN116205265A (en) | Power grid fault diagnosis method and device based on deep neural network | |
| Gu et al. | An improved sensor fault diagnosis scheme based on TA-LSSVM and ECOC-SVM | |
| Alippi et al. | An HMM-based change detection method for intelligent embedded sensors | |
| Yu et al. | A continuous learning monitoring strategy for multi-condition of nuclear power plant | |
| CN115081673A (en) | Oil and gas pipeline abnormity prediction method and device, electronic equipment and medium | |
| CN113009077B (en) | Gas detection method, gas detection device, electronic equipment and storage medium | |
| CN110532629A (en) | A kind of method for predicting residual useful life of multi-modal degenerative process | |
| CN105823801A (en) | Deep belief network characteristic extraction-based electronic nose drift compensation method | |
| CN117407770A (en) | High-voltage switch cabinet fault mode classification and prediction method based on neural network | |
| CN117409890A (en) | Transformer fault self-identification method and system based on two-way long and short-time memory | |
| CN116015274B (en) | Intelligent switch based on wireless control | |
| CN116597635A (en) | Wireless communication intelligent gas meter controller and control method thereof | |
| CN113536607B (en) | Transformer substation signal transmission system evaluation method and system | |
| CN116402777A (en) | Power equipment detection method and system based on machine vision | |
| Singh et al. | Development of ANN-based virtual fault detector for Wheatstone bridge-oriented transducers | |
| CN114021470A (en) | Relay storage life prediction method based on AMFO algorithm and SVM algorithm | |
| CN110336280B (en) | Power system cascading failure analysis method based on dictionary set acceleration |
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 |