CN109116203A - Power equipment partial discharges fault diagnostic method based on convolutional neural networks - Google Patents

Power equipment partial discharges fault diagnostic method based on convolutional neural networks Download PDF

Info

Publication number
CN109116203A
CN109116203A CN201811288833.7A CN201811288833A CN109116203A CN 109116203 A CN109116203 A CN 109116203A CN 201811288833 A CN201811288833 A CN 201811288833A CN 109116203 A CN109116203 A CN 109116203A
Authority
CN
China
Prior art keywords
feature
acquisition device
high speed
signal
discharge pulse
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201811288833.7A
Other languages
Chinese (zh)
Inventor
毛恒
艾春
邓敏
田阳普
刘成宝
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Red Phase Ltd By Share Ltd
Original Assignee
Red Phase Ltd By Share Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Red Phase Ltd By Share Ltd filed Critical Red Phase Ltd By Share Ltd
Priority to CN201811288833.7A priority Critical patent/CN109116203A/en
Publication of CN109116203A publication Critical patent/CN109116203A/en
Pending legal-status Critical Current

Links

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01RMEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/12Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
    • G01R31/1227Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials

Abstract

The invention discloses a kind of power equipment partial discharges fault diagnostic method based on convolutional neural networks is related to measuring electric variable field, comprising the following steps: be acquired the Partial discharge signal of typical fault defect model, establish PRPD spectrum data library;Grayscale image constructing module obtains shelf depreciation grayscale image with PRPD;The identification feature of grayscale image is extracted with the characteristic extracting module that the convolutional neural networks for having residual error structure are established;Identification feature is sent into classifier and carries out recognition training;Using grayscale image constructing module, characteristic extracting module and classifier as the diagnostic module of detecting instrument.The PRPD data of shelf depreciation are changed into grayscale image by the present invention, with the powerful feature extracted in self-adaptive ability of the convolutional neural networks for having residual error structure, extract the identification feature of grayscale image, and feature is applied to classical taxonomy device, by deep learning method and conventional machines learning method effective integration, the feature identification degree of extraction is high, can promote the accuracy rate of fault diagnosis.

Description

Power equipment partial discharges fault diagnostic method based on convolutional neural networks
Technical field
The present invention relates to technology of sharing fields, more specifically refer to a kind of power equipment office based on convolutional neural networks Portion's discharge fault diagnostic method.
Background technique
Power equipment is possible to occur inside equipment in design, manufacture, transport, installation, longtime running and maintenance process It is various cause seriously endanger latency insulation defects, generate different types of shelf depreciation (Partial Discharge, PD).The insulation degradation mechanism of different shelf depreciation type reflections is different, also different to apparatus insulated impairment degree, so Safe and stable operation and field failure of the research of development power equipment Characteristics of Partial Discharge and its diagnostic method to power equipment The guidance of processing is of great significance.
Mainly there is time-based TRPD mode (Time Resolved for the identification of local discharge signal in electric power equipment Partial Discharge, TRPD) and based on phase PRPD mode (Phase Resolvd Partial Discharge, PRPD) two kinds.Compared to TRPD mode, PRPD pattern technology has maturation, and stability is good, the small advantage of data volume.Therefore PRPD mode has obtained relatively broad utilization, currently based on the feature extracting method of PRPD data mainly have statistical nature method, Pulse characteristics method, moment characteristics method, fractal characteristic method, wavelet analysis method etc..The shallow-layer that the above method mainly extracts PRPD data is special Sign expression has certain subjectivity, and information is lost seriously, and discrimination is caused to reduce.Shelf depreciation grayscale image is based on PRPD A kind of data processing method of mode, it is connected PRPD aggregation of data using gray level image vegetarian refreshments information, compared to original Beginning PRPD data, shelf depreciation grayscale image are concentrated with information, and the more intuitive advantage of feature is directed to shelf depreciation gray scale at present The feature extracting method of figure is mainly the textural characteristics for utilizing image analysis preprocess method to extract image, is such as based on gray scale symbiosis Matrix (Gray Level Co-occurrence Matrix, GLCM) and local binary patterns (Local Binary Pattern, LBP) image characteristic extracting method.But this method equally have rely on expertise, subjectivity is strong, discrimination is low The disadvantages of.In recent years, deep learning method is risen, the ability of extracted in self-adaptive data characteristics is all obtained in educational circles's industry Dramatically approve.Therefore, it is badly in need of convolutional neural networks (the Convolutional Neural in deep learning field Network, CNN) it is applied to GIS partial discharge feature extraction field, with deep learning method automatic mining shelf depreciation ash The characteristic of division contained in degree figure, the feature extraction for shelf depreciation PRPD data provide new method.
In addition, the capturing technology of existing local discharge signal is: complete signal being acquired and stored, then passes through software again Processing identifies from complete signal and extracts partial discharge pulse therein, then judges failure by analyzing partial discharge pulse Type.This technology has the disadvantage in that (1) data volume is big, causes equipment to need to store and transmit a large amount of data, influences to set Standby working efficiency;(2) frequency in shelf depreciation can not be acquired in the pulse of GHz.Therefore need to design a kind of shelf depreciation The feature of signal triggers capturing technology.
Summary of the invention
A kind of power equipment partial discharges fault diagnostic method based on convolutional neural networks provided by the invention, purpose It is to solve the above-mentioned problems in the prior art.
The technical solution adopted by the invention is as follows:
A kind of power equipment partial discharges fault diagnostic method based on convolutional neural networks, comprising the following steps:
(1), by high speed acquisition device, the feature triggering capturing technology typical events various to laboratory based on local discharge signal Partial discharge pulse's signal that barrier defect model generates carries out triggering collection and stores, and establishes PRPD spectrum data library;
(2), by grayscale image constructing module, the grayscale image of shelf depreciation type is constructed using PRPD spectrum data;
(3), by characteristic extracting module, the identification feature of grayscale image is extracted using the convolutional neural networks for having residual error structure;
(4), identification feature is sent into classifier and carries out recognition training, improve the accuracy of classifier identification failure;
(5), by grayscale image constructing module, the diagnostic module of trained characteristic extracting module and classifier as detecting instrument, Feature extraction and Analysis on Fault Diagnosis are carried out to the PRPD spectrum data of live shelf depreciation.
Further, the particular content of the step (1) is: high speed acquisition device includes sequentially connected shelf depreciation sensing Device, signal conditioning module, high speed acquisition device and computer;By the finger of typical fault defect model merging shelf depreciation simulator Positioning is set, and local discharge sensor is placed in shelf depreciation simulator and detects position, and guarantees shelf depreciation simulation dress It sets and the shell reliable ground of each detection device;Local Discharge Simulation device is calibrated using calibration square wave;Pass through calculating The sampling parameter of the control software set high speed acquisition device of machine, channel parameters and signal condition parameter;Adjustment signal conditioning module Shelf depreciation simulator is boosted to the firing potential of typical fault defect model, adopted at a high speed by gain to specified gear The acquisition of acquisition means commencing signal and preservation.
Further, the particular content of the step (2) is:
Using PRRD spectrum data, using partial discharge pulse signal amplitude q as the longitudinal axis, it is divided into M section, operating frequency phase φ is Horizontal axis is divided into N number of section, and q- φ plane is divided into M × N number of section;
Count partial discharge pulse's number in each section, construct q- φ-n map;
To partial discharge pulse's numberIt is normalized i.e.:;Wherein,After normalization Pulse number,For actual pulse number,For the maximum impulse number of the q- φ-n map;
Set the gray value of each point pixel:, makeMaxima and minima it is successively right The minimal gray grade and maximum gray scale of gray level image are answered, to construct the visualization grayscale image of shelf depreciation type.
Further, in the step (3), characteristic extracting module, input data is the shelf depreciation gray scale of size M × N Figure;Characteristic extracting module introduces residual error structure on the basis of classical convolutional neural networks, H (X)=F (X)+X can be obtained, wherein H (X) the mapping output of network is indicated, F (X) indicates that residual error portion, X indicate identical mapping;Characteristic extracting module is exported in network The last layer is using global average pond layer, using the feature point group of all output characteristic pattern mappings at feature vector as classifier Input feature vector.
Further, further include step (6): continuing to build up the PRPD map of various typical fault defect models by laboratory Data, while the PRPD spectrum data of collected portion's electric discharge when on-site test faulty equipment is constantly collected, not these data It is disconnected to accumulate into PDPR spectrum data library and diagnostic module is trained, is optimized.
The method that high speed acquisition device acquires partial discharge pulse's signal, specifically includes the following steps:
(S1), the sample rate and local discharge sensor type matching of high speed acquisition device are set, and adjust high speed acquisition device For low activation threshold value, partial discharge pulse's signal of continuous sampling different types of faults defect, while guaranteeing that sampling duration can cover Cover the discharge pulse signal of multiple frequency cycles;
(S2), collected partial discharge pulse's waveform is unfolded, the list of careful analysis different types of faults defect shelf depreciation The Time-distribution of multiple pulses on the temporal signatures and time series of a pulse, and quantify these two types of characteristic parameters;
(S3), the pulse temporal feature according to obtained in step (S2) and burst length distribution characteristics design high speed acquisition device To the triggering collection parameter of local discharge pulse signal;
(S4), according to the triggering collection parameter of partial discharge pulse's signal in step (S3), the triggering of configuration high-speed acquisition device Acquisition parameter is implemented to acquire to local discharge pulse information
(S5), according to the acquisition in step (S4) to local discharge pulse signal, partial discharge pulse's signal of acquisition is stored.
Further, the step (S1) comprising the following specific steps
(S1.1), the sample rate and local discharge sensor type matching of acquisition device are set, and adjustment high speed acquisition device is low Threshold values trigger condition;
(S1.2), it using partial discharge pulse's signal of high speed acquisition device continuous sampling different types of faults defect, protects simultaneously Card sampling duration can cover the discharge pulse signal of multiple frequency cycles, the return step (S1.1) if acquiring failure, if acquisition Successfully then follow the steps (S2);
Further, the step (S3) comprising the following specific steps
(S3.1), the single pulse waveforms of partial discharge pulse's signal of analytical procedure (S2) different types of faults defect collected Feature and multiple-pulse Time-distribution, and quantization characteristic parameter value;
(S3.2), the touching according to the design high speed acquisition device of characteristic ginseng value obtained by step (S3.1) to local discharge pulse signal Send out acquisition parameter;
Further, the step (S4) comprising the following specific steps
(S4.1), the triggering collection parameter of local discharge pulse signal is matched according to the high speed acquisition device that step (S3.2) is designed Set the triggering collection parameter of high speed acquisition device;
(S4.2), using high speed acquisition device, partial discharge pulse's signal is acquired, the return step (S4.1) if acquiring failure, (S5) is successfully thened follow the steps if capturing.
Further, further comprising the steps of:
(S6), corresponding upper computer software is developed, is realized to the control of high speed acquisition device and data flow interaction;
(S7), partial discharge pulse's sequence is showed and is reconstructed by upper computer software, further analyze the potential valence of data Value.
By the above-mentioned description of this invention it is found that being compared with existing technology, the present invention has the advantages that
The PRPD spectrum data of shelf depreciation is changed into grayscale image by the present invention, and design has the convolutional Neural net of residual error structure Network extracts the identification feature of grayscale image using its powerful feature extracted in self-adaptive ability, and feature is applied to classical taxonomy Device such as SVM, random forest, BP neural network etc. realize the effective integration of deep learning method and conventional machines learning method.It is real It tests and shows that the feature that this method is extracted has higher identification, can effectively promote the accuracy rate of PD Pattern Recognition.
In addition, the present invention in acquisition pulse signal, analyzes partial discharge pulse's letter of different typical fault defects first Number, analyze the various features parameter value of pulse;Recycle the touching of these characteristic ginseng values design high speed acquisition device signal acquisition Parameter is sent out, to capture the partial discharge pulse of useful (i.e. satisfaction joint trigger policy), only to reduce the data processing of equipment Amount improves working efficiency.Meanwhile the trigger parameter established by various features combining parameter values can guarantee high speed acquisition device Frequency is captured in the pulse of GHz, detection omission or distorted signals is avoided, improves the accuracy judged fault type.
Detailed description of the invention
Fig. 1 is the feature extracting method schematic diagram based on PRPD map.
Fig. 2 is the shelf depreciation grayscale image that tetra- kinds of typical fault defect PRPD maps of U generate.
Fig. 3 is characterized the structural schematic diagram of the convolutional neural networks for having residual error structure of extraction module.
Fig. 4 is residual error schematic network structure.
Fig. 5 is network performance contrast schematic diagram.
Fig. 6 is the feature visualization figure that the present invention extracts.Wherein 1,2,3,4 respectively corresponds and indicate that point discharge, particle are put Electricity, bubble-discharge, suspended discharge.
Fig. 7 is shelf depreciation grayscale image textural characteristics visualization figure.Wherein 1,2,3,4 respectively correspond indicate point discharge, Particle electric discharge, bubble-discharge, suspended discharge.
Fig. 8 is artificial PRPD statistical nature visualization figure.Wherein 1,2,3,4 respectively corresponds and indicate that point discharge, particle are put Electricity, bubble-discharge, suspended discharge.
Specific embodiment
Illustrate a specific embodiment of the invention with reference to the accompanying drawings.In order to fully understand the present invention, it is described below and is permitted More details, but to those skilled in the art, the present invention can also be realized without these details.
A kind of power equipment partial discharges fault diagnostic method based on convolutional neural networks, comprising the following steps:
(1), by high speed acquisition device, partial discharge pulse's signal that the various typical fault defect models in laboratory are generated into Row is acquired and is stored, and establishes PRPD spectrum data library.
Specifically, high speed acquisition device includes sequentially connected local discharge sensor, signal conditioning module, high speed acquisition Device and computer.By the designated position of typical fault defect model merging shelf depreciation simulator, by local discharge sensor It is placed in shelf depreciation simulator and detects position, and guarantee that the shell of shelf depreciation simulator and each detection device is reliable Ground connection;Local Discharge Simulation device is calibrated using calibration square wave;Pass through the control software set high speed acquisition of computer The sampling parameter of device, channel parameters and signal condition parameter;The gain of adjustment signal conditioning module is to specified gear, by shelf depreciation Simulator boosts to the firing potential of typical fault defect model, the acquisition of high speed acquisition device commencing signal and preservation.
(2), by grayscale image constructing module, the grayscale image of shelf depreciation type is constructed using PRPD spectrum data.
Specifically, the specific configuration process of grayscale image is:
(2.1) M section, power frequency phase are divided into using partial discharge pulse signal amplitude q as the longitudinal axis using PRRD spectrum data Position φ is horizontal axis, is divided into N number of section, q- φ plane is divided into M × N number of section;
(2.2) partial discharge pulse's number in each section is counted, construct q- φ-n map;
(2.3) to partial discharge pulse's numberIt is normalized i.e.:;Wherein,For normalization Pulse number afterwards,For actual pulse number,For the maximum impulse number of the q- φ-n map;
(2.4) gray value of each point pixel is set:, makeMaxima and minima It is corresponding in turn to the minimal gray grade and maximum gray scale of gray level image, to construct the visualization gray scale of shelf depreciation type Figure.
Four kinds of shelf depreciation types (point discharge, particle electric discharge, air gap are constructed such as Fig. 1 and Fig. 2, according to above step Electric discharge, suspended discharge) visualization grayscale image, Fig. 2 illustrate four kinds of typical fault defect PRPD spectrum datas generation part Discharge grayscale image.
(3), by characteristic extracting module, the identification for extracting grayscale image using the convolutional neural networks for having residual error structure is special Sign.
Referring to Fig. 3 and Fig. 4, features described above extraction module selects deep learning field widely applied convolutional Neural in recent years Network is realized comprising 4 residual error structures and 2 convolutional layers, 2 maximum value pond layers and 1 overall situation are averaged pond layer, The various pieces specific structure of features described above extraction module is introduced below.
A, input data: input data is size M × N shelf depreciation grayscale image.
B, residual error structure: features described above extraction module introduces residual error structure on the basis of classical convolutional neural networks, can H (X)=F (X)+X is obtained, wherein H (X) indicates the mapping output of network, and F (X) indicates that residual error portion, X indicate identical mapping, leads to Crossing introducing identical mapping X ensure that gradient passes back to low layer by identical mapping under conditions of no complexity for increasing network Network slows down the excessive bring gradient disappearance problem of the network number of plies, promotes the information flow between each level of network.
C, global average pond layer: for the characteristic dimension of further compression network output, features described above extraction module exists The last layer of network output has used global average pond layer, and the characteristic pattern that network exports not only is mapped as a feature Point, and the offset that convolutional layer parameter error causes estimation mean value is reduced, finally by the feature of all output characteristic pattern mappings Input feature vector of the point composition characteristic vector (64 dimension) as classifier.
(4), identification feature is sent into classifier and carries out recognition training, improve the accuracy of classifier identification failure.
Classifier, training this feature extraction module parameter and classifier parameters are followed by characteristic extracting module, make score Class device obtains higher recognition accuracy, saves the weight of each parameter in characteristic extracting module, then removes classifier.It is special at this time Sign extraction module the overall situation be averaged pond layer output for its extraction characteristic quantity.
(5), the diagnosis mould by grayscale image constructing module, trained characteristic extracting module and classifier as detecting instrument Block carries out feature extraction and Analysis on Fault Diagnosis to the PRPD data of live shelf depreciation.
(6), the PRPD spectrum data of various typical fault defect models is continued to build up by laboratory, while constantly being collected The PRPD spectrum data of collected portion's electric discharge, constantly accumulates these data into PDPR map number when on-site test faulty equipment It is trained, optimizes according to library and to diagnostic module.
In step (1), the method that high speed acquisition device acquires partial discharge pulse's signal, i.e., based on local discharge signal Feature triggers capturing technology, specifically includes the following steps:
(S1), the sample rate and local discharge sensor type matching of high speed acquisition device are set, and adjust high speed acquisition device For low activation threshold value, partial discharge pulse's signal of continuous sampling different types of faults defect, while guaranteeing that sampling duration can cover Cover the discharge pulse signal of multiple frequency cycles;
(S2), collected partial discharge pulse's waveform is unfolded, the list of careful analysis different types of faults defect shelf depreciation The Time-distribution of multiple pulses on the temporal signatures and time series of a pulse, and quantify these two types of characteristic parameters;
(S3), the pulse temporal feature according to obtained in step (S2) and burst length distribution characteristics design high speed acquisition device To the triggering collection parameter of local discharge pulse signal;
(S4), according to the triggering collection parameter of partial discharge pulse's signal in step (S3), the triggering of configuration high-speed acquisition device Acquisition parameter is implemented to acquire to local discharge pulse information
(S5), according to the acquisition in step (S4) to local discharge pulse signal, partial discharge pulse's signal of acquisition is stored.
(S6), corresponding upper computer software is developed, is realized to the control of high speed acquisition device and data flow interaction;
(S7), partial discharge pulse's sequence is showed and is reconstructed by upper computer software, further analyze the potential valence of data Value.
Specifically, above-mentioned steps (S1) comprising the following specific steps
(S1.1), the sample rate and local discharge sensor type matching of acquisition device are set, and adjustment high speed acquisition device is low Threshold values trigger condition;
(S1.2), it using partial discharge pulse's signal of high speed acquisition device continuous sampling different types of faults defect, protects simultaneously Card sampling duration can cover the discharge pulse signal of multiple frequency cycles, the return step (S1.1) if acquiring failure, if acquisition Successfully then follow the steps (S2);
Specifically, above-mentioned steps (S3) comprising the following specific steps
(S3.1), the single pulse waveforms of partial discharge pulse's signal of analytical procedure (S2) different types of faults defect collected Feature and multiple-pulse Time-distribution, and quantization characteristic parameter value;
(S3.2), the touching according to the design high speed acquisition device of characteristic ginseng value obtained by step (S3.1) to local discharge pulse signal Send out acquisition parameter;
Specifically, above-mentioned steps (S4) comprising the following specific steps
(S4.1), the triggering collection parameter of local discharge pulse signal is matched according to the high speed acquisition device that step (S3.2) is designed Set the triggering collection parameter of high speed acquisition device;
(S4.2), using high speed acquisition device, partial discharge pulse's signal is acquired, the return step (S4.1) if acquiring failure, (S5) is successfully thened follow the steps if capturing.
Experiment test and structural analysis
One, Data Processing in Experiment
The data acquisition platform for relying on independent research acquires four class shelf depreciation type (tips by shelf depreciation simulator Electric discharge, particle electric discharge, bubble-discharge, suspended discharge) Partial Discharge Data.To the data sample of acquisition introduce multiple types and The noise jamming of phase carries out data and increases expansion, effectively improves the diversity of sample data, by four classes finally obtained part The UHF data of electric discharge type are divided into training set and test set, specific to divide as shown in table 1
Point discharge Particle electric discharge Bubble-discharge Suspended discharge
Training set 570 570 570 570
Test set 500 500 500 500
1 shelf depreciation UHF data of table divide table
Two, experimental result and analysis
In 2.1 practical applications, the signal windowing number of different power equipment acquisitions can be different, in order to probe into the variation of phase windowing number Influence to diagnosis effect tests shelf depreciation type identification accuracy rate herein by the quantitative comparison for changing phase windowing.
Criticizing standardization (Batch Normalization, BN) is a kind of effective regularization method, and BN is used for neural network When, output standardization can be made to arrive the normal distribution of N (0,1), reduced to being standardized inside each lot data The change of intrinsic nerve member distribution, has and accelerates training speed, reduce the influence of weights initialisation, while improving generalization ability of network energy The advantages of power.
Probe into phase windowing number and data batch influence of the normalizing operation for network performance.Select SVM as classifier, It is 100,250,500 and 1000 that phase windowing number, which is arranged, while having carried out the standardized comparative experiments of no data batch, as a result as schemed Shown in 5.
As shown in figure 5, the feature that characteristic extracting module is extracted has not on classifier when out of phase windowing number is arranged Same performance.With the increase of windowing number, the recognition accuracy of classifier is gradually increased.In equally windowing number, by data Criticizing standardized classifier has higher recognition accuracy.
2.2 present invention extract the comparison of feature and PRPD data statistical characteristics and shelf depreciation grayscale image textural characteristics
In order to verify the validity for the feature that the present invention extracts, the present invention is extracted to the statistical nature drawn game of feature, PRPD data Portion's electric discharge grayscale image textural characteristics are tested in SVM, random forest, BP neural network respectively, to its shelf depreciation type Recognition result compare, discrimination of each category feature on each classifier such as table 2
Point discharge Particle electric discharge Bubble-discharge Suspended discharge Average Accuracy
SVM (PRPD statistical nature) 86.21% 42.74% 43.68% 46.23% 54.72%
SVM (shelf depreciation grayscale image textural characteristics) 69.36% 46.42% 77.62% 66.26% 64.91%
SVM (present invention extracts feature+BN+1000 window) 100% 93.63% 94.15% 99.59% 96.84%
Random forest (PRPD statistical nature) 98.38% 56.81% 66.59% 39.38% 65.29%
Random forest (shelf depreciation grayscale image textural characteristics) 67.75% 40.68% 89.70% 80.52% 69.66%
Random forest (present invention extracts feature+BN+1000 window) 79.74% 90.49% 78.79% 88.12% 84.29%
BP neural network (PRPD statistical nature) 91.80% 48.08% 67.25% 66.55% 68.42%
BP neural network (shelf depreciation grayscale image textural characteristics) 66.94% 35.32% 69.93% 58.25% 57.61%
BP neural network (present invention extracts feature+BN+1000 window) 100% 92.90% 91.95% 98.96% 96%
The comparison of 2 shelf depreciation type identification accuracy rate of table
It can be obtained by table 1, with the feature extracted of the present invention compared to other two category features in SVM, random forest, in BP neural network Obtain higher Average Accuracy.And the performance that the height of Average Accuracy can preferably evaluate a classifier is good Bad, which reflects the feature identification with higher that the present invention extracts from side.For more intuitive more different spies Three kinds of features are carried out t-SNE dimension reduction and visualization processing by the difference between sign, and t-SNE is a kind of data Nonlinear Dimension Reduction mode, It is handled by t-SNE, original high dimensional feature data are down to bidimensional and cluster visualization is carried out to data.
As shown in Fig. 6, Fig. 7 and Fig. 8, after handling through t-SNE, the feature that the present invention extracts has preferably cluster effect Fruit, this feature for demonstrating context of methods extraction again have higher discrimination compared to manual features.
To sum up, the PRPD spectrum data of shelf depreciation is changed into grayscale image by the present invention, and design has the volume of residual error structure Product neural network extracts the identification feature of grayscale image, and feature is applied to using its powerful feature extracted in self-adaptive ability Classical taxonomy device such as SVM, random forest, BP neural network etc., realization deep learning method and conventional machines learning method have Effect fusion.Experiment shows that the feature that this method is extracted has higher identification, can effectively promote PD Pattern Recognition Accuracy rate.
In addition, the present invention in acquisition pulse signal, analyzes partial discharge pulse's letter of different typical fault defects first Number, analyze the various features parameter value of pulse;Recycle the touching of these characteristic ginseng values design high speed acquisition device signal acquisition Parameter is sent out, to capture the partial discharge pulse of useful (i.e. satisfaction joint trigger policy), only to reduce the data processing of equipment Amount improves working efficiency.Meanwhile the trigger parameter established by various features combining parameter values can guarantee high speed acquisition device Frequency is captured in the pulse of GHz, detection omission or distorted signals is avoided, improves the accuracy judged fault type.
The above is only a specific embodiment of the present invention, but the design concept of the present invention is not limited to this, all to utilize this Design makes a non-material change to the present invention, and should all belong to behavior that violates the scope of protection of the present invention.

Claims (10)

1. the power equipment partial discharges fault diagnostic method based on convolutional neural networks, which comprises the following steps:
(1), by high speed acquisition device, the feature triggering capturing technology based on local discharge signal is to various typical fault defects Partial discharge pulse's signal that model generates carries out triggering collection and stores, and establishes PRPD spectrum data library;
(2), by grayscale image constructing module, the shelf depreciation gray scale of typical fault defect is constructed using PRPD spectrum data Figure;
(3), by characteristic extracting module, distinguishing for shelf depreciation grayscale image is extracted using the convolutional neural networks for having residual error structure Know feature;
(4), identification feature is sent into classifier and carries out recognition training, improve the accuracy of classifier identification failure;
(5), by shelf depreciation grayscale image constructing module, trained characteristic extracting module and classifier examining as detecting instrument Disconnected module carries out feature extraction and Analysis on Fault Diagnosis to the PRPD spectrum data of live shelf depreciation.
2. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as described in claim 1, feature Be, the particular content of the step (1) is: high speed acquisition device includes sequentially connected local discharge sensor, signal tune Manage module, high speed acquisition device and computer;It, will by the designated position of typical fault defect model merging shelf depreciation simulator Local discharge sensor, which is placed in shelf depreciation simulator, detects position, and guarantees shelf depreciation simulator and each detection The shell reliable ground of equipment;Local Discharge Simulation device is calibrated using calibration square wave;It is soft by the control of computer The sampling parameter of part setting high-speed collector, channel parameters and signal condition parameter;The gain of adjustment signal conditioning module is to specified Shelf depreciation simulator, is boosted to the firing potential of typical fault defect model by gear, and high speed acquisition device starts Signal acquisition and preservation.Wherein, high speed acquisition device is high speed acquisition oscillograph.
3. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as described in claim 1, feature It is, the particular content of the step (2) is:
Using PRRD spectrum data, using partial discharge pulse signal amplitude q as the longitudinal axis, it is divided into M section, operating frequency phase φ is Horizontal axis is divided into N number of section, and q- φ plane is divided into M × N number of section;
Count partial discharge pulse's number in each section, construct q- φ-n map;
To partial discharge pulse's numberIt is normalized i.e.:;Wherein,For the arteries and veins after normalization Jig frequency number,For actual pulse number,For the maximum impulse number of the q- φ-n map;
Set the gray value of each point pixel:, makeMaxima and minima it is successively right The minimal gray grade and maximum gray scale of gray level image are answered, to construct the visualization grayscale image of shelf depreciation type.
4. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as described in claim 1, feature Be: in the step (3), characteristic extracting module, input data is the shelf depreciation grayscale image of size M × N;Feature mentions Modulus block introduces residual error structure on the basis of classical convolutional neural networks, and H (X)=F (X)+X can be obtained, and wherein H (X) indicates net The mapping of network exports, and F (X) indicates that residual error portion, X indicate identical mapping;The last layer that characteristic extracting module is exported in network It is using the average pond layer of the overall situation, the feature point group of all output characteristic pattern mappings is special as the input of classifier at feature vector Sign.
5. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as described in claim 1, feature It is, further includes step (6): continues to build up the PRPD spectrum data of various typical fault defect models by laboratory, simultaneously The PRPD spectrum data of collected portion electric discharge when constantly collecting on-site test faulty equipment, these data constantly accumulate into PDPR spectrum data library is simultaneously trained diagnostic module, optimizes.
6. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as described in claim 1, feature It is: in the step (1), the method that high speed acquisition device acquires partial discharge pulse's signal, specifically includes the following steps:
(S1), the sample rate and local discharge sensor type matching of high speed acquisition device are set, and adjust high speed acquisition device For low activation threshold value, partial discharge pulse's signal of continuous sampling different types of faults defect, while guaranteeing that sampling duration can cover Cover the discharge pulse signal of multiple frequency cycles;
(S2), collected partial discharge pulse's waveform is unfolded, the list of careful analysis different types of faults defect shelf depreciation The Time-distribution of multiple pulses on the temporal signatures and time series of a pulse, and quantify these two types of characteristic parameters;
(S3), the pulse temporal feature according to obtained in step (S2) and burst length distribution characteristics design high speed acquisition device To the triggering collection parameter of local discharge pulse signal;
(S4), according to the triggering collection parameter of partial discharge pulse's signal in step (S3), the triggering of configuration high-speed acquisition device Acquisition parameter is implemented to acquire to local discharge pulse information
(S5), according to the acquisition in step (S4) to local discharge pulse signal, partial discharge pulse's signal of acquisition is stored.
7. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as claimed in claim 6, feature Be, the step (S1) comprising the following specific steps
(S1.1), the sample rate and local discharge sensor type matching of acquisition device are set, and adjustment high speed acquisition device is low Threshold values trigger condition;
(S1.2), it using partial discharge pulse's signal of high speed acquisition device continuous sampling different types of faults defect, protects simultaneously Card sampling duration can cover the discharge pulse signal of multiple frequency cycles, the return step (S1.1) if acquiring failure, if acquisition Successfully then follow the steps (S2).
8. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as claimed in claim 6, feature Be, the step (S3) comprising the following specific steps
(S3.1), the single pulse waveforms of partial discharge pulse's signal of analytical procedure (S2) different types of faults defect collected Feature and multiple-pulse Time-distribution, and quantization characteristic parameter value;
(S3.2), the touching according to the design high speed acquisition device of characteristic ginseng value obtained by step (S3.1) to local discharge pulse signal Send out acquisition parameter.
9. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as claimed in claim 6, feature Be, the step (S4) comprising the following specific steps
(S4.1), the triggering collection parameter of local discharge pulse signal is matched according to the high speed acquisition device that step (S3.2) is designed Set the triggering collection parameter of high speed acquisition device;
(S4.2), using high speed acquisition device, partial discharge pulse's signal is acquired, the return step (S4.1) if acquiring failure, (S5) is successfully thened follow the steps if capturing.
10. the power equipment partial discharges fault diagnostic method based on convolutional neural networks as claimed in claim 6, feature It is, further comprising the steps of:
(S6), corresponding upper computer software is developed, is realized to the control of high speed acquisition device and data flow interaction;
(S7), partial discharge pulse's sequence is showed and is reconstructed by upper computer software, further analyze the potential valence of data Value.
CN201811288833.7A 2018-10-31 2018-10-31 Power equipment partial discharges fault diagnostic method based on convolutional neural networks Pending CN109116203A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811288833.7A CN109116203A (en) 2018-10-31 2018-10-31 Power equipment partial discharges fault diagnostic method based on convolutional neural networks

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811288833.7A CN109116203A (en) 2018-10-31 2018-10-31 Power equipment partial discharges fault diagnostic method based on convolutional neural networks

Publications (1)

Publication Number Publication Date
CN109116203A true CN109116203A (en) 2019-01-01

Family

ID=64855694

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811288833.7A Pending CN109116203A (en) 2018-10-31 2018-10-31 Power equipment partial discharges fault diagnostic method based on convolutional neural networks

Country Status (1)

Country Link
CN (1) CN109116203A (en)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110208660A (en) * 2019-06-05 2019-09-06 国网江苏省电力有限公司电力科学研究院 A kind of training method and device for power equipment shelf depreciation defect diagonsis
CN110224401A (en) * 2019-06-19 2019-09-10 清华大学 In conjunction with the Power system transient stability prediction method of manual features and residual error network
CN110262463A (en) * 2019-07-09 2019-09-20 河海大学常州校区 A kind of rail traffic platform door fault diagnosis system based on deep learning
CN110412431A (en) * 2019-08-05 2019-11-05 国网湖南省电力有限公司 A kind of diagnostic method and diagnostic system of the shelf depreciation defect type of power equipment
CN110672988A (en) * 2019-08-29 2020-01-10 国网江西省电力有限公司电力科学研究院 Partial discharge mode identification method based on hierarchical diagnosis
CN110703057A (en) * 2019-11-04 2020-01-17 国网山东省电力公司电力科学研究院 Power equipment partial discharge diagnosis method based on data enhancement and neural network
CN110794243A (en) * 2019-11-14 2020-02-14 南方电网科学研究院有限责任公司 Fault diagnosis method, system and equipment for direct current system
CN110929697A (en) * 2019-12-17 2020-03-27 中国人民解放军海军航空大学 Neural network target identification method and system based on residual error structure
ES2780448A1 (en) * 2019-02-22 2020-08-25 Ormazabal Corp Technology Aie Partial discharge recognition method and system for electrical network diagnosis (Machine-translation by Google Translate, not legally binding)
CN111638047A (en) * 2020-05-25 2020-09-08 国网河北省电力有限公司电力科学研究院 GIS equipment mechanical fault diagnosis system and method
CN111856224A (en) * 2020-06-29 2020-10-30 国网河南省电力公司周口供电公司 Partial discharge multi-source fusion detection system and method for high-voltage switch cabinet
CN111999614A (en) * 2020-08-21 2020-11-27 国网河南省电力公司周口供电公司 Local discharge multi-source fusion detection early warning method for high-voltage switch cabinet
CN112036320A (en) * 2020-09-07 2020-12-04 华北电力大学(保定) Partial discharge state identification method and device, discharge simulator and identification equipment
CN112149549A (en) * 2020-09-18 2020-12-29 国网山东省电力公司泰安供电公司 GIS partial discharge type identification method based on depth residual error network
CN112763860A (en) * 2020-12-11 2021-05-07 贵州电网有限责任公司 Power transmission line fault and abnormal discharge monitoring system and method

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102628916A (en) * 2012-04-12 2012-08-08 江苏省电力试验研究院有限公司 System for positioning local discharge of transformer on line based on ultrahigh-frequency signal
CN103558532A (en) * 2013-11-20 2014-02-05 国家电网公司 Partial discharge on-line detection system of high-voltage crosslinked polyethylene power cable
CN104502820A (en) * 2014-12-25 2015-04-08 国家电网公司 Two-dimensional wavelet transformation based transformer partial discharge spectrum denoising method
CN106556781A (en) * 2016-11-10 2017-04-05 华乘电气科技(上海)股份有限公司 Shelf depreciation defect image diagnostic method and system based on deep learning
CN107167716A (en) * 2017-07-11 2017-09-15 国网福建省电力有限公司泉州供电公司 A kind of shelf depreciation default kind identification method and device
CN108108723A (en) * 2018-01-19 2018-06-01 深圳市恩钛控股有限公司 A kind of face feature extraction method based on deep learning
CN108334948A (en) * 2018-02-09 2018-07-27 武汉理工大学 A kind of mechanical bearing fault diagnosis technology based on wide residual error network learning model
CN108573225A (en) * 2018-03-30 2018-09-25 国网天津市电力公司电力科学研究院 A kind of local discharge signal mode identification method and system
CN108627743A (en) * 2018-05-15 2018-10-09 国网江苏省电力有限公司电力科学研究院 A kind of lossless method for catching of shelf depreciation nanosecond narrow pulse sequence

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102628916A (en) * 2012-04-12 2012-08-08 江苏省电力试验研究院有限公司 System for positioning local discharge of transformer on line based on ultrahigh-frequency signal
CN103558532A (en) * 2013-11-20 2014-02-05 国家电网公司 Partial discharge on-line detection system of high-voltage crosslinked polyethylene power cable
CN104502820A (en) * 2014-12-25 2015-04-08 国家电网公司 Two-dimensional wavelet transformation based transformer partial discharge spectrum denoising method
CN106556781A (en) * 2016-11-10 2017-04-05 华乘电气科技(上海)股份有限公司 Shelf depreciation defect image diagnostic method and system based on deep learning
CN107167716A (en) * 2017-07-11 2017-09-15 国网福建省电力有限公司泉州供电公司 A kind of shelf depreciation default kind identification method and device
CN108108723A (en) * 2018-01-19 2018-06-01 深圳市恩钛控股有限公司 A kind of face feature extraction method based on deep learning
CN108334948A (en) * 2018-02-09 2018-07-27 武汉理工大学 A kind of mechanical bearing fault diagnosis technology based on wide residual error network learning model
CN108573225A (en) * 2018-03-30 2018-09-25 国网天津市电力公司电力科学研究院 A kind of local discharge signal mode identification method and system
CN108627743A (en) * 2018-05-15 2018-10-09 国网江苏省电力有限公司电力科学研究院 A kind of lossless method for catching of shelf depreciation nanosecond narrow pulse sequence

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
HUI SONG 等: "GIS Partial Discharge Pattern Recognition via Deep Convolutional Neural Network under Complex Data Sources", 《IEEE TRANSACTIONS ON DIELECTRICS AND ELECTRICAL INSULATION》 *
丁清山: "基于小波的GIS局部放电类型识别算法及其实现", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 *
刘兵 等: "基于卷积神经网络的变压器局部放电模式识别", 《高压电器》 *
汪可 等: "基于灰度图像分解的局部放电特征提取与优化", 《电机与控制学报》 *
王云浩 等: "基于稀疏表示算法的气体绝缘组合电器(GIS)局部放电识别", 《科学技术与工程》 *

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
ES2780448A1 (en) * 2019-02-22 2020-08-25 Ormazabal Corp Technology Aie Partial discharge recognition method and system for electrical network diagnosis (Machine-translation by Google Translate, not legally binding)
EP3699614A1 (en) * 2019-02-22 2020-08-26 Ormazabal Corporate Technology, A.I.E. Method and system of partial discharge recognition for diagnosing electrical networks
CN110208660A (en) * 2019-06-05 2019-09-06 国网江苏省电力有限公司电力科学研究院 A kind of training method and device for power equipment shelf depreciation defect diagonsis
CN110208660B (en) * 2019-06-05 2021-07-27 国网江苏省电力有限公司电力科学研究院 Training method and device for diagnosing partial discharge defects of power equipment
CN110224401B (en) * 2019-06-19 2020-09-01 清华大学 Power system transient stability prediction method combining artificial features and residual error network
CN110224401A (en) * 2019-06-19 2019-09-10 清华大学 In conjunction with the Power system transient stability prediction method of manual features and residual error network
CN110262463A (en) * 2019-07-09 2019-09-20 河海大学常州校区 A kind of rail traffic platform door fault diagnosis system based on deep learning
CN110262463B (en) * 2019-07-09 2021-12-10 河海大学常州校区 Rail transit platform door fault diagnosis system based on deep learning
CN110412431A (en) * 2019-08-05 2019-11-05 国网湖南省电力有限公司 A kind of diagnostic method and diagnostic system of the shelf depreciation defect type of power equipment
CN110672988A (en) * 2019-08-29 2020-01-10 国网江西省电力有限公司电力科学研究院 Partial discharge mode identification method based on hierarchical diagnosis
CN110703057A (en) * 2019-11-04 2020-01-17 国网山东省电力公司电力科学研究院 Power equipment partial discharge diagnosis method based on data enhancement and neural network
CN110703057B (en) * 2019-11-04 2022-02-25 国网山东省电力公司电力科学研究院 Power equipment partial discharge diagnosis method based on data enhancement and neural network
CN110794243A (en) * 2019-11-14 2020-02-14 南方电网科学研究院有限责任公司 Fault diagnosis method, system and equipment for direct current system
CN110929697B (en) * 2019-12-17 2021-04-13 中国人民解放军海军航空大学 Neural network target identification method and system based on residual error structure
CN110929697A (en) * 2019-12-17 2020-03-27 中国人民解放军海军航空大学 Neural network target identification method and system based on residual error structure
CN111638047A (en) * 2020-05-25 2020-09-08 国网河北省电力有限公司电力科学研究院 GIS equipment mechanical fault diagnosis system and method
CN111856224A (en) * 2020-06-29 2020-10-30 国网河南省电力公司周口供电公司 Partial discharge multi-source fusion detection system and method for high-voltage switch cabinet
CN111999614A (en) * 2020-08-21 2020-11-27 国网河南省电力公司周口供电公司 Local discharge multi-source fusion detection early warning method for high-voltage switch cabinet
CN112036320A (en) * 2020-09-07 2020-12-04 华北电力大学(保定) Partial discharge state identification method and device, discharge simulator and identification equipment
CN112149549A (en) * 2020-09-18 2020-12-29 国网山东省电力公司泰安供电公司 GIS partial discharge type identification method based on depth residual error network
CN112763860A (en) * 2020-12-11 2021-05-07 贵州电网有限责任公司 Power transmission line fault and abnormal discharge monitoring system and method
CN112763860B (en) * 2020-12-11 2022-05-24 贵州电网有限责任公司 Power transmission line fault and abnormal discharge monitoring system and method

Similar Documents

Publication Publication Date Title
CN109116203A (en) Power equipment partial discharges fault diagnostic method based on convolutional neural networks
CN108573225A (en) A kind of local discharge signal mode identification method and system
CN109100627A (en) A kind of power equipment partial discharges fault diagnostic method based on end-to-end mode
CN107340456B (en) Power distribution network operating condition intelligent identification Method based on multiple features analysis
CN108509976A (en) The identification device and method of animal
CN109871746A (en) A kind of mode identification method based on the diagnosis of GIS insulation defect
CN108710889A (en) A kind of scarce cylinder method for diagnosing faults of automobile engine
CN105701470A (en) Analog circuit fault characteristic extraction method based on optimal wavelet packet decomposition
CN110161388A (en) A kind of the fault type recognition method and its system of high-tension apparatus
CN103558519A (en) GIS partial discharge ultrasonic signal identification method
CN109858352A (en) A kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement
CN102279358A (en) MCSKPCA based neural network fault diagnosis method for analog circuits
CN107037327A (en) Partial discharges fault judges feature extracting method and decision method
CN110298085A (en) Analog-circuit fault diagnosis method based on XGBoost and random forests algorithm
CN110108992B (en) Cable partial discharge fault identification method and system based on improved random forest algorithm
CN106443380B (en) A kind of distribution cable local discharge signal recognition methods and device
CN109284672A (en) A kind of Mechanical Failure of HV Circuit Breaker diagnostic method based on PSO-Kmeans algorithm
CN108919067A (en) A kind of recognition methods for GIS partial discharge mode
CN108921110A (en) Radar signal classification method of the novel convolutional neural networks in conjunction with Wigner-Ville distribution
CN111025100A (en) Transformer ultrahigh frequency partial discharge signal mode identification method and device
CN111289251A (en) Rolling bearing fine-grained fault identification method
CN110261746B (en) Cable defect detection method based on periodic attenuation characteristics of oscillating wave voltage
Puspitasari et al. Partial discharge waveform identification using image with convolutional neural network
CN109472239A (en) A kind of frequency hopping radio set individual discrimination method
CN108537266A (en) A kind of cloth textured fault sorting technique of depth convolutional network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20190101