CN113466681B - Breaker service life prediction method based on small sample learning - Google Patents
Breaker service life prediction method based on small sample learning Download PDFInfo
- Publication number
- CN113466681B CN113466681B CN202110598719.XA CN202110598719A CN113466681B CN 113466681 B CN113466681 B CN 113466681B CN 202110598719 A CN202110598719 A CN 202110598719A CN 113466681 B CN113466681 B CN 113466681B
- Authority
- CN
- China
- Prior art keywords
- data
- circuit breaker
- prediction
- waveform
- breaker
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000004364 calculation method Methods 0.000 claims abstract description 6
- 238000012549 training Methods 0.000 claims description 38
- 238000012360 testing method Methods 0.000 claims description 32
- 238000012545 processing Methods 0.000 claims description 14
- 238000010586 diagram Methods 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 9
- 238000005299 abrasion Methods 0.000 claims description 7
- 238000002474 experimental method Methods 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 4
- 239000002131 composite material Substances 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 2
- 230000036541 health Effects 0.000 abstract description 10
- 238000003745 diagnosis Methods 0.000 abstract description 8
- 238000012423 maintenance Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 7
- 238000009826 distribution Methods 0.000 description 6
- 238000005457 optimization Methods 0.000 description 4
- 238000005520 cutting process Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000000638 solvent extraction Methods 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000003416 augmentation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003211 malignant effect Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000003746 surface roughness Effects 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
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
- G01R31/327—Testing of circuit interrupters, switches or circuit-breakers
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Testing Electric Properties And Detecting Electric Faults (AREA)
Abstract
The invention discloses a breaker service life prediction method based on small sample learning, and belongs to the technical field of breaker equipment. According to the breaker service life prediction method based on small sample learning, the small sample learning method is utilized, the required sample size can be generated according to a small amount of data, a large number of fault diagnosis sample libraries are not needed, a prediction model can be built, the prediction efficiency is high, and the application cost is low; and further, according to the vibration waveform of the circuit breaker, the health condition of the circuit breaker is analyzed, and the service life of the circuit breaker is predicted. According to the invention, a Markov chain prediction method is not adopted, complex probability calculation of each state is not needed, the prediction accuracy can be effectively improved, the actual prediction accuracy and the actual prediction efficiency can meet the requirements of field application, and the scheme is simple, effective and practical. Furthermore, the invention improves the efficiency of the overhaul work of the circuit breaker, and operation and maintenance staff can grasp the operation state in time, thereby ensuring the safety and reliability of power supply.
Description
Technical Field
The invention relates to a breaker service life prediction method based on small sample learning, and belongs to the technical field of breaker equipment.
Background
The low-voltage circuit breaker is one of extremely important components in the whole low-voltage distribution system, and is widely applied to the transportation and distribution of electric energy, the control and protection of electric equipment, but due to various reasons such as circuit breaker design, manufacture, material quality and operation, malignant faults of the equipment occur, the safe and stable operation of a power grid is seriously influenced, and the grasping of the health condition of the circuit breaker is important to the guarantee of the safe operation of a power system.
The mechanical complexity and various uncertainties of circuit breakers can cause their performance to degrade during use, rendering the circuit breaker ineffective. The current research method mainly comprises the steps of equivalent service life of the circuit breaker and service life of an electric contact, including a surface roughness method, a contact resistance method, a shower arc method, an effective contact distance method arcing parameter statistical analysis method, a mass loss method spectrum analysis method and the like, wherein the service life of the circuit breaker can be predicted to a certain extent, but the improvement and the perfection are needed.
Chinese patent (publication No. CN110287543 a) discloses a life prediction method for a relay protection device, which includes: based on importance and available criteria, indexes of incorrect action times, fault times, CPU temperature and working voltage are used as core indexes for evaluating the health state of the protection device. And defining the relative degradation degree of the obtained data, obtaining an initial state probability distribution vector through the relative degradation degree and intersecting cloud drops of the cloud model, obtaining a state transition probability matrix according to the no-back effect of the Markov chain, finally obtaining the state probability distribution state of each year of the protection device through the state probability distribution vector and the state transition probability matrix, and comparing the state probability distribution with a standard reliability criterion so as to predict the service life of the protection device.
However, the method needs more variables to be collected, and adopts a Markov chain prediction method, all matters have independence in statistics, so that the future state is independent of all past states, meanwhile, various probabilities of each state change are needed to be known, and one state calculation is inaccurate and errors can be caused to the whole prediction result; the related matrix operation process is very complex, the diagnosis efficiency is low, and the requirements of field application are difficult to meet, so the method has the limitations, and the actual diagnosis accuracy and the diagnosis efficiency are low.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a breaker service life prediction method based on small sample learning, which can generate required sample size according to a small amount of data, can build a prediction model without a large amount of fault diagnosis sample libraries and has high prediction efficiency; further, according to the vibration waveform of the circuit breaker, the health condition of the circuit breaker is analyzed, and the service life of the circuit breaker is predicted; the scheme is simple, effective and feasible, and can meet the requirements of field application.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
a breaker life prediction method based on small sample learning,
The method comprises the following steps:
the method comprises the steps that firstly, data generated by opening and closing of a circuit breaker are collected to form a data set;
Secondly, dividing the data set in the first step into a training set and a testing set;
A training set for training the model; a test set for testing the model;
Thirdly, carrying out enhancement processing on the training set and the test set data in the second step;
fourth, according to the enhancement data in the third step, a small sample learning method is utilized to establish a prediction model;
And fifthly, analyzing the health condition of the circuit breaker and predicting the service life of the circuit breaker by using the prediction model in the fourth step according to the vibration waveform generated by detecting the opening and closing of the circuit breaker, and making early warning and reminding.
According to the invention, through continuous exploration and test, the small sample learning method is utilized, the required sample size can be generated according to a small amount of data, a prediction model can be built without a large amount of fault diagnosis sample libraries, the prediction efficiency is high, and the application cost is low; and further, according to the vibration waveform of the circuit breaker, the health condition of the circuit breaker is analyzed, and the service life of the circuit breaker is predicted.
According to the method, a Markov chain prediction method is not adopted, complex probability calculation of each state is not needed, error probability is effectively reduced, prediction accuracy can be effectively improved, actual prediction accuracy and prediction efficiency can meet requirements of field application, and the scheme is simple, effective and practical.
Furthermore, the invention can realize the accurate prediction of the service life of the circuit breaker, improves the efficiency of the overhaul work of the circuit breaker, ensures that operation maintenance staff can grasp the running state in time, prevents the circuit breaker from having safety accidents and ensures the safety and reliability of power supply.
As a preferred technical measure:
in the first step, the specific steps of data acquisition are as follows:
Step 1: constructing a breaker experiment acquisition data platform;
Step 2: data acquisition is carried out on vibration waveforms generated by opening and closing the circuit breaker, and time domain data T data of the vibration waveforms are recorded;
Step 3: recording the switching times omega and the switching current I of the circuit breaker while carrying out the step 2;
step 4: substituting the discrete data of different opening and closing currents I and opening and closing times N of the circuit breaker recorded in the step 3 into the formula (1) to obtain a relational expression of the opening and closing currents I, the opening and closing times omega and the theoretical electric wear quantity E loss:
Wherein E loss represents theoretical electric abrasion quantity, alpha represents an open-close current index, and omega represents the number of open-close times allowed by the circuit breaker when the open-close current is I β;
step 5: and (3) calculating unknown parameters in the formula (1) to obtain a specific expression.
As a preferred technical measure:
the second step, the training set comprises a support set S and a query set Q;
the test set includes a support set S 'and a query set Q';
the specific calculation steps of the sample types number required by the support set S and the query set Q' are as follows:
Step 1: the theoretical service life of the circuit breaker is the maximum opening and closing times omega max;
step 2: according to the prediction precision required by the circuit breaker, setting a minimum resolution r, and calculating the formula of the number x of the test sample types as follows:
The data set may be generally as per 7:3 are divided into a training set and a testing set, which are basically consistent, are all acquired data, and are classified and named according to different roles.
As a preferred technical measure:
the training set comprises a plurality of labels and category data;
The support set comprises labels and category data;
the query set includes only category data, no tag information, and data in the same tag space as the support set.
As a preferred technical measure:
The third step, the enhancement processing of the data specifically comprises the following contents:
step 1: according to the voltage, current and vibration data acquired by sensing, calculating the contact loss of the circuit breaker by the formula (1);
step 2: converting the recorded vibration waveform data into frequency domain data through Fast Fourier Transform (FFT);
step 3: by utilizing the mapping relation between the contact loss and the vibration waveform frequency domain data, the recorded existing data sets are linearly combined to amplify the data, and after the data sets are amplified, the overall prediction performance and generalization capability of the model can be improved;
Step 4: normalizing the data set row amplified in the step 3, and converting each group of data into an array of H multiplied by W.
As a preferred technical measure:
the conversion of the vibration waveform data includes the following:
the n vibration waveform data is F i (t) (i=1, 2,..n), the corresponding frequency domain waveform data is F i (ω) (i=1, 2,..n), the time domain waveform and the frequency domain waveform of the noise are G (t), G (ω), and the corresponding electric wear amount is Q i (U, I) (i=1, 2,..n) each time the corresponding electric wear amount is calculated, then
The sampled raw vibration waveform is noted as h (f, g) =f (t) +g (t);
The noisy frequency domain plot waveform is denoted H (F, G) =f (ω) +g (ω);
The preprocessed waveform is denoted h (f, 0) =f (t);
The post-pretreatment frequency domain plot is denoted h (0, g) =g (t);
The composite waveform is recorded as
The Fourier transform shows that F and F, G and G, H and H are inverse Fourier transforms, and H and Q have mapping relations.
As a preferred technical measure:
The data set is enhanced (amplified) in the following manner:
Mode 1: same Q value, clipping waveform "
After cutting waveform data of h (f, g) and h (f, 0) for k times, vibration waveform data of 2k sections are formed, and k times of the vibration waveform data is enlarged compared with an original data volume sample under the same Q value, wherein k is E (2, 3);
Obtaining a spectrogram through Fourier transformation of the time domain waveform data set after data enhancement, and finally converting the spectrogram into a matrix of H-W, which can be used as a model input;
Q is the electric abrasion quantity, and when the data is not enhanced, one Q value corresponds to one sample; after data enhancement, one Q value corresponds to multiple samples. .
As a preferred technical measure:
The manner of enhancement of the dataset:
mode 2: different Q values, superimposed waveforms "
If j different Q values exist, selecting any two Q values, overlapping two sections of waveforms, cutting waveform data of h (f, g) and h (f, 0) k times, and expanding a data setDoubling;
Obtaining a spectrogram through Fourier transformation of the time domain waveform data set after data enhancement, and finally converting the spectrogram into a matrix of H-W, which can be used as a model input;
Q is the electric abrasion quantity, and when the data is not enhanced, one Q value corresponds to one sample; after data enhancement, one Q value corresponds to multiple samples.
As a preferred technical measure: ,
The fourth step, the establishment of a prediction model:
Step 1: taking the data with the size of H multiplied by W obtained in the third step as input, taking the size of a convolution kernel as k, the size of a step length as s, filling the size as l, and carrying out two-dimensional convolution according to a formula (3) to obtain feature graphs H out and W out;
step 2: carrying out batch normalization processing and linear rectification on the characteristic diagram H out×Wout;
Step 3: according to hardware computing capability and resources, selecting and repeating the steps 1 and 2 for 16 times, 32 times, 64 times, 128 times, and the like, and processing the data through a convolutional neural network to obtain two feature sets, wherein the two feature sets are respectively And Z q, calculating the inner product V p according to equation (4):
the T is the transposed symbol.
As a preferred technical measure:
In the fifth step, the prediction process of the prediction model specifically includes:
Step 1: initializing parameters phi 0 of the prediction model according to the prediction model determined in the fourth step;
step 2: sampling a training task m, assigning a network parameter phi 0 to a network unique to the task m to obtain I.e. initial/>
Step 3: based on once optimizedCalculating the loss of task m using the query set of the training set and according to equation (5)
L is a shorthand of loss function, which is used to measure the degree of inconsistency between the predicted value f (x) and the true value Y of the model, and is a non-negative real value function, which is usually expressed by L (Y, f (x)), and the smaller the loss function, the better the robustness of the model; the lower case L represents the loss of single update, the upper case L represents the global loss, and in the optimization problem, when the minimum global loss is ensured, we consider that the optimal solution is obtained; n represents the total number of updates of the task.
Step 4: calculated according to the formula (6)Pair/>Is a gradient of (2);
Step 5: updating phi 0 to phi 1 according to the formula (7), wherein epsilon is the learning rate;
Step 6: 1 task n is sampled, and a parameter phi 1 is assigned to the task n to obtain I.e./>
Step 7: based on step 6, training data of task n is used to calculate learning rate epsilon n of task nPerforming one-time optimization, updating/>Calculating the loss of the task n according to the formula (8) by using the query set of the training set;
step 8: calculated according to the formula (9) Pair/>Is a gradient of (2);
Step 9: based on the optimized step 7 Phi 1 is updated to obtain phi according to the formula (10) 2
Step 10: repeatedly updating the parameters according to the formula (11);
step 11: adjusting parameters by using a learning optimizer, and evaluating effects by using a query set of the test set;
step 12: and inputting vibration waveform data of the circuit breaker to be tested into the regulated prediction model to obtain the health condition of the circuit breaker, finish the prediction of the service life of the circuit breaker and make early warning and reminding.
In the practical application of the circuit breaker to be tested after training of the prediction model, the circuit breaker to be tested has no label information (the label information is a classification label of how long the circuit breaker can be used presumably, which is a result to be predicted by the model), the prediction model predicts the service life of the circuit breaker through vibration waveforms, namely predicts to obtain the label information, and the query set has no label information, so that the performance and generalization capability of the model are evaluated by comparing the equivalent practical test result with the prediction result of the prediction model.
Compared with the prior art, the invention has the following beneficial effects:
according to the invention, through continuous exploration and test, the small sample learning method is utilized, the required sample size can be generated according to a small amount of data, a prediction model can be built without a large amount of fault diagnosis sample libraries, the prediction efficiency is high, and the application cost is low; and further, according to the vibration waveform of the circuit breaker, the health condition of the circuit breaker is analyzed, and the service life of the circuit breaker is predicted.
According to the invention, a Markov chain prediction method is not adopted, complex probability calculation of each state is not needed, the prediction accuracy can be effectively improved, the actual prediction accuracy and the actual prediction efficiency can meet the requirements of field application, and the scheme is simple, effective and practical.
Furthermore, the invention can realize the accurate prediction of the service life of the circuit breaker, improves the efficiency of the overhaul work of the circuit breaker, ensures that operation maintenance staff can grasp the running state in time, prevents the circuit breaker from having safety accidents and ensures the safety and reliability of power supply.
Drawings
FIG. 1 is a schematic diagram of the construction principle of the acquisition system of the present invention;
FIG. 2 is a flow chart of the data set acquisition process of the present invention;
FIG. 3 is a block diagram of a dataset partitioning of the present invention;
FIG. 4 is a block diagram of a specific partitioning of training and testing sets of the present invention;
FIG. 5 is a flow chart of the training process of the present invention;
FIG. 6 is a diagram of learning optimizer parameter updates in accordance with the present invention;
fig. 7 is a flow chart of waveform data processing according to the present invention.
Description of the drawings:
(a) An original waveform; (b) noisy frequency domain plot; (c) a preprocessed waveform; (d) a post-pretreatment frequency domain plot; (e) synthesizing a waveform diagram.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
On the contrary, the invention is intended to cover any alternatives, modifications, equivalents, and variations as may be included within the spirit and scope of the invention as defined by the appended claims. Further, in the following detailed description of the present invention, certain specific details are set forth in order to provide a better understanding of the present invention. The present invention will be fully understood by those skilled in the art without the details described herein.
As shown in fig. 1-7, a breaker life prediction method based on small sample learning,
The method comprises the following steps:
the method comprises the steps that firstly, data generated by opening and closing of a circuit breaker are collected to form a data set;
Secondly, dividing the data set in the first step into a training set and a testing set;
A training set for training the model; the test set is used for estimating the generalization capability of the model in practical application;
Thirdly, carrying out enhancement processing on the training set and the test set data in the second step;
fourth, according to the enhancement data in the third step, a small sample learning method is utilized to establish a prediction model;
And fifthly, analyzing the health condition of the circuit breaker and predicting the service life of the circuit breaker by using the prediction model in the fourth step according to the vibration waveform generated by detecting the opening and closing of the circuit breaker, and making early warning and reminding.
According to the invention, through continuous exploration and test, the small sample learning method is utilized, the required sample size can be generated according to a small amount of data, a large number of fault diagnosis sample libraries are not needed, a prediction model can be built, and the application cost is low; further, according to the vibration waveform of the circuit breaker, the health condition of the circuit breaker is analyzed, and the service life of the circuit breaker is predicted; the scheme is simple, effective and feasible.
One specific embodiment of the dataset of the present invention:
Stage one data set acquisition:
step 1: an experimental acquisition data platform is built according to the schematic diagram of the system of fig. 1.
Step 2: the vibration waveform generated by opening and closing the circuit breaker is subjected to data acquisition according to the data acquisition processing flow chart of fig. 2, and the time domain data T data and the frequency domain data F data are recorded.
Step 3: and (2) simultaneously recording the switching times omega and the switching current I of the circuit breaker when the step (2) is carried out.
Step 4: substituting the discrete data of different opening and closing currents I and opening and closing times omega of the circuit breaker recorded in the step 3 into the formula (1) to obtain a relational expression of the opening and closing currents I, the opening and closing times omega and the theoretical electric wear quantity E loss
Wherein E loss represents the theoretical electric abrasion quantity, alpha represents the switching current index, and omega represents the switching times allowed by the circuit breaker when the switching current is I β.
Step 5: and (3) solving parameters in the equation (1) through a computer to obtain a specific expression.
Stage two data set partitioning:
step 1: the theoretical service life of the circuit breaker is assumed to be the maximum opening and closing times omega max.
Step 2: according to the prediction precision required by the circuit breaker, setting a minimum resolution r, and calculating the formula of the number x of the test sample types as follows: .
Step3: the prediction samples obtained according to equation (2) will be divided into x classes,
Step 4: according to the data set division of fig. 3, the data set is divided into two major parts of a training set and a test set.
The dataset is generally as per 7:3 are divided into a training set and a testing set, which are essentially the same and are all acquired data, and are classified and named according to different roles. In a strict sense, the data set is divided into a training set, a verification set and a test set, wherein the training set is used for training a model, the verification set is used for model selection and parameter adjustment, the test set is used for estimating the generalization capability of the model in practical application, and the verification set and the test set are considered to have the same function in general cases, but only one observation is carried out on the model, and the generalization capability of the model is estimated.
Further, the training set includes a support set S and a query set Q, and the test set includes a support set S 'and a query set Q'.
The training set comprises a plurality of labels and category data, wherein the support set comprises the labels and category data, and the query set is data without label information and in the same label space with the support set.
Step 5: the specific division of step 4 refers to fig. 5, where x represents the number of sample types used in the training process, M represents the number of sample types used in the testing process, and k represents the number of samples of each type in the testing process.
Enhancement and processing of phase three data sets
Step 1: and according to the data F data collected in the stage one, artificially enhancing the collected data according to the fitted formula (1). According to the data enhancement method, according to the voltage, current and vibration data acquired by sensing, on one hand, the contact loss of the circuit breaker is calculated through the existing formula (1), on the other hand, recorded vibration waveform data are converted into frequency domain data through FFT (fast Fourier transform fast Fourier transform), after the fact that the contact loss and the vibration waveform frequency domain data have approximate linear relations is verified through experiments, a data set can be amplified through linear combination according to the existing data records, and after the data set is amplified, the overall prediction performance and generalization capability of a model can be improved.
Step 2: normalizing the data in the step 1, converting each group of data into an H multiplied by W array,
One specific embodiment of the predictive model of the present invention:
step 1: training according to the training process of fig. 5, taking the data with the size of H multiplied by W obtained by the third step as input, taking the size of a convolution kernel as k, the size of a step as s, filling the size as l, and carrying out two-dimensional convolution according to a formula (3) to obtain feature graphs H out and W out;
step 2: carrying out batch normalization processing and linear rectification on the characteristic diagram H out×Wout;
Step 3: according to hardware computing capability and resources, the number of times of repeating the steps 1 and 2 is 16, 32, 64, 128 and the like, and after the data is processed by a convolutional neural network, two feature sets can be obtained, namely And Z q, calculating the inner product V according to equation (4) p
One specific embodiment of the predictive model of the present invention:
Step 1: the network parameters phi 0 are initialized according to the network model defined in step 1 in the model.
Step 2: sampling a training task m, assigning a network parameter phi 0 to a network unique to the task m to obtainI.e. initial/>
Step 3: based on once optimizedCalculating the loss of task m using the query set and according to equation (5)
L is a shorthand for loss function, which is used to measure the degree of inconsistency between the predicted value f (x) and the true value Y of the model, and is a non-negative real value function, usually denoted by L (Y, f (x)), the smaller the loss function, the better the robustness of the model. The lower case L represents the loss of a single update, the upper case L represents the global loss, and in the optimization problem, when the minimum global loss is ensured, we consider that the optimal solution is obtained.
Step 4: calculated according to the formula (6)Pair/>Is a gradient of (a).
Step 5: and updating phi 0 to phi 1 according to the formula (7), wherein epsilon is the learning rate.
Step 6: 1 task n is sampled, and a parameter phi 1 is assigned to the task n to obtainI.e./>
Step 7: based on step 6, training data of task n is used to calculate learning rate epsilon n of task nPerforming one-time optimization, updating/>The loss of task n is calculated using the query set and according to equation (8).
Step 8: calculated according to the formula (9)Pair/>Is a gradient of (a).
Step 9: based on the optimized step 7Phi 1 is updated to obtain phi according to the formula (10) 2
Step 10: the update parameters are repeated according to equation (11).
Step 11: the learning optimizer according to fig. 6 adjusts the parameters, using the query set to evaluate the effect.
In practical application after training of the model, the model predicts the service life of the circuit breaker through a vibration wave pattern, namely, predicts the label information, and the query set is free of the label information, so that the performance and generalization capability of the model can be evaluated by equivalent practical test results.
As shown in fig. 7, one embodiment of waveform data enhancement (augmentation) of the present invention:
Let n vibration waveform data recorded by the experiment be F i (t) (i=1, 2..n), the corresponding frequency domain waveform data is F i (ω) (i=1, 2,..n), the time domain waveform and the frequency domain waveform of the noise are G (t), G (ω), each experiment calculates that the corresponding electrical wear amount is Q i (U, I) (i=1, 2,., n), then
The sampled raw vibration waveform as shown in fig. 7 (a) can be denoted as h (f, g) =f (t) +g (t);
the noisy frequency domain plot waveform shown in fig. 7 (b) can be noted as H (F, G) =f (ω) +g (ω);
The preprocessed waveform as shown in fig. 7 (c) may be denoted as h (f, 0) =f (t);
the post-pretreatment frequency domain plot as shown in fig. 7 (d) can be denoted as h (0, g) =g (t);
The composite waveform as shown in FIG. 7 (e) can be written as
The Fourier transform shows that F and F, G and G, H and H are inverse Fourier transforms, and H and Q have mapping relations.
Model training data is enhanced in two ways:
Mode 1: same Q value, clipping waveform "
By cutting the waveform data of h (f, g), h (f, 0) k times, vibration waveform data of 2k sections are formed, and k times of the waveform data is enlarged compared with the original data volume sample under the same Q value, wherein k is E (2, 3).
Mode 2: different Q values, superimposed waveforms "
If j different Q values exist, any two Q values are selected, and after two-section waveform superposition is carried out, the data set can be expanded on the basis of the mode 1Multiple times.
The enhanced time domain waveform data set is subjected to Fourier transformation to obtain a spectrogram, and finally the spectrogram is converted into a matrix of H-W, so that the matrix can be used as a model input.
Finally, it should be noted that: the above embodiments are only for illustrating the technical aspects of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the above embodiments, it should be understood by those of ordinary skill in the art that: modifications and equivalents may be made to the specific embodiments of the invention without departing from the spirit and scope of the invention, which is intended to be covered by the claims.
Claims (3)
1. A breaker life prediction method based on small sample learning is characterized in that,
The method comprises the following steps:
the method comprises the steps that firstly, data generated by opening and closing of a circuit breaker are collected to form a data set;
Secondly, dividing the data set in the first step into a training set and a testing set;
A training set for training the model; a test set for testing the model;
Thirdly, carrying out enhancement processing on the training set and the test set data in the second step;
fourth, according to the enhancement data in the third step, a small sample learning method is utilized to establish a prediction model;
fifthly, predicting the service life of the circuit breaker by using the prediction model in the fourth step according to the vibration waveform generated by detecting the opening and closing of the circuit breaker;
The specific steps of data acquisition are as follows:
Step 1: constructing a breaker experiment acquisition data platform;
step 2: data acquisition is carried out on vibration waveforms generated by opening and closing of the circuit breaker, and time domain data of the vibration waveforms are recorded ；
Step3: while step 2 is carried out, the opening and closing times of the circuit breaker are recordedAnd switching current/>；
Step 4: for the different switching currents recorded in step 3And the number of times of opening and closing of the circuit breaker/>The discrete data of (1) are substituted into the equation (1) to obtain the switching current/>Number of times of opening and closing/>Theoretical amount of electrical wear/>Is defined by the relation:
（1）
wherein, Indicating the on-off current index,/>Indicating the switching current as/>When the circuit breaker is in the open-close state, the number of times allowed by the circuit breaker is counted;
step 5: calculating the theoretical electric wear in the formula (1) to obtain a specific value;
the enhancement processing of the data specifically comprises the following contents:
Step 1: calculating theoretical electric wear through the formula (1) according to current and vibration waveform data acquired by sensing;
step 2: converting the recorded vibration waveform data into frequency domain data through fast Fourier transform;
Step 3: amplifying data by linearly combining the recorded existing data sets by using a mapping relation existing between the theoretical electric wear amount and frequency domain data of the vibration waveform;
step 4: normalizing the data set amplified in the step 3, and converting each group of data into An array of (a) is provided;
the conversion of the vibration waveform data includes the following:
The secondary vibration waveform data is/> The corresponding frequency domain waveform data is/>The time domain waveform and the frequency domain waveform of the noise are/>, respectively，/>Each time the corresponding electric abrasion loss is calculated as/>The sampled raw vibration waveform is noted/>;
Noisy frequency domain graph waveform notation;
The preprocessed waveforms are noted as;
The pretreated frequency domain diagram is recorded as;
The composite waveform is recorded as；
The data set is enhanced in the following manner:
By aligning ，/>Waveform data processing of/>After the secondary clipping, will form/>The vibration waveform data of the segment is enlarged/>, compared with the original data volume sample, under the same Q valueMultiple of (1)/>；
Obtaining a spectrogram through Fourier transformation of the time domain waveform data set after data enhancement, and finally converting the spectrogram into a matrix of H-W as a model input;
Q is the electric abrasion quantity, and when the data is not enhanced, one Q value corresponds to one sample; after data enhancement, one Q value corresponds to a plurality of samples;
The manner of enhancement of the dataset:
If there is Selecting any two Q values from different Q values, overlapping two sections of waveforms, and performing the process of matching/>，/>Waveform data processing of/>After the secondary cropping, the dataset is enlarged/>Doubling;
Obtaining a spectrogram through Fourier transformation of the time domain waveform data set after data enhancement, and finally converting the spectrogram into a matrix of H-W as a model input;
Q is the electric abrasion quantity, and when the data is not enhanced, one Q value corresponds to one sample; after data enhancement, one Q value corresponds to a plurality of samples;
Building a prediction model:
step 1: the size obtained by the third step is Takes as input the data of the convolution kernel of sizeStep size is/>Filling size is/> Performing two-dimensional convolution according to formula (3) to obtain a feature map/>And；
（3）
Step 2: map the characteristic mapCarrying out batch normalization treatment and linear rectification;
step 3: according to the hardware computing capability and resources, selecting the number of times of repeating the step 1 and the step 2 as 16 or 32 or 64 or 128, and processing the data by a convolutional neural network to obtain two feature sets, namely And/>Calculating the inner product according to equation (4):
（4）
The T is the transposed symbol.
2. A breaker life prediction method based on small sample learning as claimed in claim 1, wherein,
The second step, the training set comprises a support setAnd query set/>q；
The test set includes a support setAnd query set/>；
Support setAnd query set/>Q, query set/>The specific calculation steps of the required sample type number are as follows:
Step 1: the theoretical service life of the circuit breaker is the maximum opening and closing times ；
Step 2: setting minimum resolution according to the prediction accuracy required by the circuit breakerThe formula for calculating the number x of test sample types is as follows:
（2）。
3. A breaker life prediction method based on small sample learning as claimed in claim 2, wherein,
The training set comprises a plurality of labels and category data;
its support set includes tags and category data, and the query set includes only category data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110598719.XA CN113466681B (en) | 2021-05-31 | 2021-05-31 | Breaker service life prediction method based on small sample learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110598719.XA CN113466681B (en) | 2021-05-31 | 2021-05-31 | Breaker service life prediction method based on small sample learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113466681A CN113466681A (en) | 2021-10-01 |
| CN113466681B true CN113466681B (en) | 2024-05-10 |
Family
ID=77871825
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110598719.XA Active CN113466681B (en) | 2021-05-31 | 2021-05-31 | Breaker service life prediction method based on small sample learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113466681B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102022126547A1 (en) | 2022-10-12 | 2024-04-18 | Maschinenfabrik Reinhausen Gmbh | Device and method for monitoring an on-load tap changer |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103344909A (en) * | 2013-06-25 | 2013-10-09 | 国家电网公司 | Method for detecting service life of breaker |
| CN106680706A (en) * | 2017-02-27 | 2017-05-17 | 国网上海市电力公司 | Electric life prediction and health state online monitoring device of converter breaker |
| CN109948597A (en) * | 2019-04-26 | 2019-06-28 | 福州大学 | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method |
| CN110138303A (en) * | 2019-05-28 | 2019-08-16 | 华中科技大学 | A kind of AC Motor Control performance characteristic identification network model construction method and application |
| CN110174610A (en) * | 2019-05-17 | 2019-08-27 | 温州大学 | A method of obtaining A.C. contactor electric life based on convolutional neural networks |
| CN110398348A (en) * | 2019-06-18 | 2019-11-01 | 清华大学 | Memory, Mechanical Failure of HV Circuit Breaker diagnostic method and device |
| WO2020014893A1 (en) * | 2018-07-18 | 2020-01-23 | 深圳鲲云信息科技有限公司 | Method for implementing transposed convolution and related product |
| 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 |
| CN112084974A (en) * | 2020-09-14 | 2020-12-15 | 北京工商大学 | Multi-label rolling bearing fault diagnosis method based on meta-learning |
| CN112418013A (en) * | 2020-11-09 | 2021-02-26 | 贵州大学 | Complex working condition bearing fault diagnosis method based on meta-learning under small sample |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10012697B2 (en) * | 2015-08-13 | 2018-07-03 | Eaton Intelligent Power Limited | Failure prediction device and method for vacuum circuit breakers |
-
2021
- 2021-05-31 CN CN202110598719.XA patent/CN113466681B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103344909A (en) * | 2013-06-25 | 2013-10-09 | 国家电网公司 | Method for detecting service life of breaker |
| CN106680706A (en) * | 2017-02-27 | 2017-05-17 | 国网上海市电力公司 | Electric life prediction and health state online monitoring device of converter breaker |
| WO2020014893A1 (en) * | 2018-07-18 | 2020-01-23 | 深圳鲲云信息科技有限公司 | Method for implementing transposed convolution and related product |
| CN109948597A (en) * | 2019-04-26 | 2019-06-28 | 福州大学 | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method |
| CN110174610A (en) * | 2019-05-17 | 2019-08-27 | 温州大学 | A method of obtaining A.C. contactor electric life based on convolutional neural networks |
| CN110138303A (en) * | 2019-05-28 | 2019-08-16 | 华中科技大学 | A kind of AC Motor Control performance characteristic identification network model construction method and application |
| CN110398348A (en) * | 2019-06-18 | 2019-11-01 | 清华大学 | Memory, Mechanical Failure of HV Circuit Breaker diagnostic method and device |
| 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 |
| CN112084974A (en) * | 2020-09-14 | 2020-12-15 | 北京工商大学 | Multi-label rolling bearing fault diagnosis method based on meta-learning |
| CN112418013A (en) * | 2020-11-09 | 2021-02-26 | 贵州大学 | Complex working condition bearing fault diagnosis method based on meta-learning under small sample |
Non-Patent Citations (2)
| Title |
|---|
| 基于大数据技术的高压断路器寿命评估研究;赵洋洋;《中国优秀硕士学位论文全文数据库》(第2期);C042-1000 * |
| 高压断路器新型状态监测装置的研制;易慧;《中国优秀硕士学位论文全文数据库》(第3期);C042-182 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113466681A (en) | 2021-10-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107144428B (en) | A kind of rail traffic vehicles bearing residual life prediction technique based on fault diagnosis | |
| Hocaoğlu | Stochastic approach for daily solar radiation modeling | |
| CN107092744B (en) | Surface subsidence prediction method based on EMD-SVR | |
| CN103136539B (en) | Ground net corrosion speed grade Forecasting Methodology | |
| CN113157771A (en) | Data anomaly detection method and power grid data anomaly detection method | |
| CN108168924B (en) | Reciprocating compressor service life prediction method based on VMD (model-based decomposition) and MFSS (model-based decomposition) models | |
| WO2023103130A1 (en) | Quantum walk-based time-series multiscale analysis method | |
| CN110490366A (en) | Runoff forestry method based on variation mode decomposition and iteration decision tree | |
| CN113466681B (en) | Breaker service life prediction method based on small sample learning | |
| CN110263829A (en) | Medium-term and long-term Runoff Forecast method based on CEEMDAN and random forest | |
| CN115809716A (en) | Flight path prediction method and device based on multi-resolution analysis and wavelet packet reconstruction | |
| Jiao et al. | Short-term building energy consumption prediction strategy based on modal decomposition and reconstruction algorithm | |
| CN114397569A (en) | Circuit breaker fault arc detection method based on VMD parameter optimization and sample entropy | |
| CN112798888B (en) | Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train | |
| CN106528679A (en) | Time series analysis method based on multilinear autoregression model | |
| Huang et al. | A decomposition‐based multi‐time dimension long short‐term memory model for short‐term electric load forecasting | |
| CN109768549B (en) | Method for constructing thermal stability security domain of electric power system | |
| CN115510964B (en) | Computer calculation method for liquid chromatograph scientific instrument | |
| CN114564487B (en) | Meteorological raster data updating method combining forecast prediction | |
| CN116796189A (en) | Aerosol extinction coefficient profile prediction method based on deep learning technology | |
| Yun et al. | Forecasting of heart rate variability using wrist-worn heart rate monitor based on hidden Markov model | |
| CN115423175A (en) | Wind power prediction error decomposition method | |
| CN115577856A (en) | Method and system for predicting construction cost and controlling balance of power transformation project | |
| Zhu et al. | Prediction of air quality index based on wavelet transform combination model | |
| CN112348700B (en) | Line capacity prediction method combining SOM clustering and IFOU equation |
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 |