CN109164362B - Method and system for identifying partial discharge defect fault of direct current cable - Google Patents
Method and system for identifying partial discharge defect fault of direct current cable Download PDFInfo
- Publication number
- CN109164362B CN109164362B CN201811113943.XA CN201811113943A CN109164362B CN 109164362 B CN109164362 B CN 109164362B CN 201811113943 A CN201811113943 A CN 201811113943A CN 109164362 B CN109164362 B CN 109164362B
- Authority
- CN
- China
- Prior art keywords
- partial discharge
- pulse waveform
- waveform signal
- discharge pulse
- deep belief
- 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
Images
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
- G01R31/12—Testing 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
- G01R31/1227—Testing 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
- G01R31/1263—Testing 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 of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation
- G01R31/1272—Testing 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 of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation of cable, line or wire insulation, e.g. using partial discharge measurements
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Testing Relating To Insulation (AREA)
Abstract
The invention discloses a method for identifying partial discharge defect faults of a direct current cable, which comprises the following steps: (1) collecting partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable; (2) extracting effective information of the partial discharge pulse waveform signal to obtain a training sample; (3) constructing a deep belief network based on a limited Boltzmann machine, and performing unsupervised training on the deep belief network by adopting a training sample to obtain network parameters; (4) carrying out supervised training on the deep belief network so as to optimize network parameters; (5) and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network so as to obtain a recognition result from the output of the deep belief network. In addition, the invention also discloses a system for identifying the partial discharge defect fault of the direct current cable, which comprises the following steps: the device comprises a signal acquisition module, a preprocessing module and a signal processing module. The method and the system for identifying the partial discharge defect fault of the direct current cable are high in accuracy.
Description
Technical Field
The present invention relates to an identification method and system, and more particularly, to an identification method and system for dc cable faults.
Background
With the rapid development of flexible high-voltage direct-current transmission, Cross-linked polyethylene (XLPE) cables are increasingly applied by virtue of excellent insulating properties. With the continuous operation of more and more high-voltage direct-current transmission projects, the improvement of an online monitoring and fault early warning system is urgently needed to ensure the reliability of power supply. Since Partial Discharge (PD) is listed as an important index for determining the state of the power equipment by the international electrotechnical commission, the mode identification and fault diagnosis research of the Partial discharge of the dc cable is still in the beginning stage.
At present, the research on the partial discharge of the cable is mostly directed to an alternating current XLPE cable, and a uniform detection method and an evaluation standard are not formed for a direct current cable, so that the mode identification of the partial discharge signal of the direct current cable has a larger research space.
Disclosure of Invention
One of the purposes of the present invention is to provide a method for identifying a dc cable partial discharge defect fault, which is based on a pulse current signal generated when a dc cable is subjected to partial discharge, then preprocesses the pulse current signal and extracts effective information, then trains by constructing a neural network, and identifies the signal to be identified by the trained neural network, thereby finally realizing fault diagnosis of the dc cable partial discharge defect. Compared with the prior art, the identification method has higher accuracy in fault classification.
Based on the above purpose, the present invention provides a method for identifying a dc cable partial discharge defect fault, which comprises the steps of:
(1) collecting partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable;
(2) extracting effective information of the partial discharge pulse waveform signal to obtain a training sample;
(3) constructing a deep belief network based on a limited Boltzmann machine, and performing unsupervised training on the deep belief network by adopting a training sample to obtain network parameters;
(4) carrying out supervised training on the deep belief network so as to optimize network parameters;
(5) and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network so as to obtain a recognition result from the output of the deep belief network.
In the method for identifying the partial discharge defect fault of the direct current cable, when the direct current cable is subjected to partial discharge, the partial discharge pulse waveform signal is collected, and then the collected partial discharge pulse waveform signal is preprocessed to extract effective information of the partial discharge pulse waveform signal. Constructing a depth belief network based on a Restricted Boltzmann Machine (RBM) for unsupervised training to obtain network parameters, then carrying out supervised training on the depth belief network, optimizing the network parameters to finally obtain a trained depth belief network, and inputting a partial discharge pulse waveform signal to be recognized into the trained depth belief network to obtain a recognition result from the output of the local discharge pulse waveform signal.
The method for identifying the partial discharge defect fault of the direct current cable has high identification accuracy.
Further, in the identification method of the present invention, the insulation defect discharge model includes at least an air gap discharge model, a corona discharge model, a scratch discharge model, and a creeping discharge model.
Further, in the identification method according to the present invention, in the step (2), the Canny algorithm is adopted to extract the valid information of the partial discharge pulse waveform signal, wherein the valid information includes the segment in which the partial discharge pulse waveform signal changes.
Further, in the identification method according to the present invention, in the step (2), extracting valid information of the partial discharge pulse waveform signal by using a Canny algorithm includes the steps of:
smoothing the waveform signal f (x) of the local discharge pulse by adopting a one-dimensional Gaussian function to obtain a Gaussian-filtered waveform signal
Obtaining
The derivative g (x) of (a), performing non-maximum suppression, and only keeping the maximum point of the derivative;
performing double-threshold detection: setting a low threshold δlAnd a high threshold value deltahIf the derivative g (x)i) Less than deltalThen mark xiNon-edge points, if greater than deltahMark point xiThe points are strong edge points, and the rest points are marked as weak edge points;
and (3) inhibiting isolated weak edge points: setting a neighborhood epsilon, and if no strong edge point exists in the weak edge point neighborhood epsilon, taking the weak edge point as a non-edge point; if the weak edge point exists, selecting the 1 st weak edge point as the starting point of the segment of the partial discharge pulse waveform signal, otherwise, selecting the 1 st strong edge point as the starting point of the segment of the partial discharge pulse waveform signal.
In some embodiments, the filter window length in gaussian filtering is 20, and the standard deviation of the gaussian distribution is 1.5.
Further, in the identification method of the present invention, in step (3), the deep belief network is trained layer by layer using a contrastive divergence algorithm to obtain the network parameters, so that the data reconstructed by the deep belief network and the data of the training sample are kept consistent as much as possible.
Further, in the identification method of the present invention, in step (4), an ADAM algorithm is used to perform supervised training on the deep belief network.
It should be noted that the supervised training of the deep belief network may adopt a gradient descent method and a conjugate gradient descent method, and preferably adopts Adaptive Moment Estimation (ADAM) to select a more appropriate learning rate, so as to easily converge to a local optimum.
Accordingly, another object of the present invention is to provide a system for identifying partial discharge defects of a dc cable, which can quickly and accurately identify partial discharge defects occurring in the dc cable.
Based on the above object, the present invention further provides a system for identifying a partial discharge fault of a dc cable, which includes:
the signal acquisition module is used for acquiring partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable;
the preprocessing module extracts effective information of the partial discharge pulse waveform signal to obtain a training sample;
the signal processing module is used for carrying out unsupervised training on the constructed deep belief network based on the limited Boltzmann machine by adopting the training samples to obtain network parameters; carrying out supervised training on the deep belief network to obtain optimized network parameters;
and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network, and obtaining a recognition result from the output of the deep belief network.
Further, in the identification system of the present invention, the preprocessing module uses a Canny algorithm to extract valid information of the partial discharge pulse waveform signal, where the valid information includes a segment in which the partial discharge pulse waveform signal changes, and the extracting of the valid information of the partial discharge pulse waveform signal using the Canny algorithm includes the steps of:
smoothing the waveform signal f (x) of the local discharge pulse by adopting a one-dimensional Gaussian function to obtain a Gaussian-filtered waveform signal
Obtaining
The derivative g (x) of (a), performing non-maximum suppression, and only keeping the maximum point of the derivative;
performing double-threshold detection: setting a low threshold δlAnd a high threshold value deltahIf the derivative g (x)i) Less than deltalThen mark xiNon-edge points, if greater than deltahMark point xiThe points are strong edge points, and the rest points are marked as weak edge points;
and (3) inhibiting isolated weak edge points: setting a neighborhood epsilon, and if no strong edge point exists in the weak edge point neighborhood epsilon, taking the weak edge point as a non-edge point; if the weak edge point exists, selecting the 1 st weak edge point as the starting point of the segment of the partial discharge pulse waveform signal, otherwise, selecting the 1 st strong edge point as the starting point of the segment of the partial discharge pulse waveform signal.
Further, in the recognition system of the present invention, the signal processing module uses a contrast divergence algorithm to unsupervised train the deep belief network to obtain the network parameters.
Further, in the recognition system of the present invention, the signal processing module performs supervised training on the deep belief network by using an ADAM algorithm to obtain the optimized network parameters.
The method and the system for identifying the partial discharge defect fault of the direct current cable have the following advantages and beneficial effects:
the method and the system for identifying the partial discharge defect fault of the direct current cable can identify the fault aiming at the partial discharge of the direct current cable, collect partial discharge pulse waveform signals, then preprocess the collected partial discharge pulse waveform signals and extract effective information of the partial discharge pulse waveform signals. The method comprises the steps of constructing a deep belief network based on a limited Boltzmann machine to carry out unsupervised training to obtain network parameters, then carrying out supervised training on the deep belief network, optimizing the network parameters to finally obtain a trained deep belief network, and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network to obtain a recognition result from the output of the partial discharge pulse waveform signal.
The method and the system for identifying the partial discharge defect fault of the direct current cable have high identification accuracy.
Drawings
Fig. 1 is a structural framework diagram of a dc partial discharge fault identification system according to an embodiment of the present invention.
Fig. 2 is a schematic flow chart of a method for identifying a dc partial discharge defect fault according to an embodiment of the present invention.
Fig. 3 shows a partial discharge pulse waveform signal when a partial discharge defect fault occurs in a dc current.
Fig. 4 shows a partial discharge pulse waveform signal under a corona discharge insulation defect fault model after pretreatment.
Fig. 5 shows a partial discharge pulse waveform signal under an air gap discharge insulation defect fault model after pretreatment.
Fig. 6 shows a partial discharge pulse waveform signal under a scratch discharge insulation defect fault model after pretreatment.
Fig. 7 shows a partial discharge pulse waveform signal under a surface creepage discharge insulation defect fault model after pretreatment.
Fig. 8 schematically shows a deep belief network structure in the identification method of the partial discharge defect fault of the direct current cable according to the invention.
Fig. 9 shows a confusion matrix of the insulation defect fault identification results using the identification of comparative example 1.
Fig. 10 shows a confusion matrix of the insulation defect fault identification results using the identification of comparative example 2.
Fig. 11 shows a confusion matrix of the insulation defect fault identification results identified using embodiment 1.
Detailed Description
The method and system for identifying partial discharge fault of dc cable according to the present invention will be further described with reference to the following specific embodiments and the accompanying drawings, but the description should not be construed as an undue limitation on the technical solution of the present invention.
As shown in fig. 1, in the present embodiment, the system for identifying a partial discharge fault in a dc cable includes: the device comprises a signal acquisition module, a preprocessing module and a signal processing module.
The signal acquisition module acquires partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable.
And the preprocessing module extracts effective information of the partial discharge pulse waveform signal to obtain a training sample. In this embodiment, the preprocessing module extracts effective information of the partial discharge pulse waveform signal by using a Canny algorithm, where the effective information includes a segment in which the partial discharge pulse waveform signal changes, and the extracting the effective information of the partial discharge pulse waveform signal by using the Canny algorithm includes:
smoothing the waveform signal f (x) of the local discharge pulse by adopting a one-dimensional Gaussian function to obtain a Gaussian-filtered waveform signal
For example, in the smoothing process, assuming that the center point of the filter window moves to x ═ μ, a one-dimensional gaussian function is used to assign weights to the center point and other points in the window, as shown in the following equation:
in the formula, ωxσ is the standard deviation of the Gaussian distribution as the weight of point x. Obtaining filtered values by weighted averaging of points within a window
Moving the filtering window point by point to obtain the waveform signal after Gaussian filtering
In the scheme, the filter window length is set to be 20, and sigma is 1.5.
Obtaining
The derivative g (x) of (a), non-maxima suppression is performed, only the maxima points of the derivative are retained.
Performing double-threshold detection: setting a low threshold δlAnd a high threshold value deltahIf the derivative g (x)i) Less than deltalThen mark xiNon-edge points, if greater than deltahMark point xiStrong edge points, the remaining points are marked as weak edge points.
And (3) inhibiting isolated weak edge points: setting a neighborhood epsilon, and if no strong edge point exists in the weak edge point neighborhood epsilon, taking the weak edge point as a non-edge point; if the weak edge point exists, selecting the 1 st weak edge point as the starting point of the segment of the partial discharge pulse waveform signal, otherwise, selecting the 1 st strong edge point as the starting point of the segment of the partial discharge pulse waveform signal. In view of a large acquisition amount when acquiring the partial discharge pulse waveform signal, it is preferable that the waveform length of the partial discharge pulse waveform signal be intercepted as 600 points.
The signal processing module is used for carrying out unsupervised training on the constructed deep belief network based on the limited Boltzmann machine by adopting the training samples to obtain network parameters; and carrying out supervised training on the deep belief network to obtain optimized network parameters. In the embodiment, the signal processing module unsupervised trains the deep belief network by using a contrast divergence algorithm, and supervised training is performed on the deep belief network by using an ADAM algorithm to obtain optimized network parameters.
And inputting the partial discharge pulse waveform signal to be recognized into the trained deep belief network, and obtaining a recognition result from the output of the partial discharge pulse waveform signal.
In order to verify the identification effect of the identification system for the partial discharge defect fault of the direct current cable, the identification method shown in fig. 2 is adopted for identification. As shown in fig. 2, the method for identifying a partial discharge defect fault of a dc cable includes the steps of:
(1) collecting partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable;
(2) extracting effective information of the partial discharge pulse waveform signal to obtain a training sample;
(3) constructing a deep belief network based on RBM, and carrying out unsupervised training on the deep belief network by adopting a training sample to obtain network parameters;
(4) carrying out supervised training on the deep belief network so as to optimize network parameters;
(5) and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network so as to obtain a recognition result from the output of the deep belief network.
In order to obtain partial discharge pulse waveform signals of different insulation defect discharge models, four different direct current cable defect fault models are constructed in a simulation mode, and the construction process is as follows:
corona discharge insulation defect fault model: when a corona discharge insulation defect fault model of an XLPE insulation tip is manufactured, a metal needle with the length of 3cm is inserted into the insulation position of the XLPE and contacts with a wire core, and therefore the corona discharge defect fault is simulated.
Air gap discharge insulation defect fault model: when an air gap discharge insulation defect fault model is manufactured, a plurality of micropores are punched on the surface of XLPE and sealed by epoxy resin, so that air gap discharge defect faults caused by residual air bubbles in insulation are simulated.
Scratch discharge insulation defect fault model: when a scratch discharge insulation defect fault model is manufactured, a scratch with the width of 2mm, the length of 10mm and the depth of 1mm is scribed on XLPE insulation, so that a scratch discharge defect fault is simulated.
Creepage discharge insulation defect fault model along surface: when the creepage discharge insulation defect fault model is manufactured, the residue with the width of 3mm and the length of 10mm is left when the outer semi-conducting layer is stripped at one end, so that creepage discharge fault faults are simulated.
It should be noted that the insulation defect discharge model is only an illustrative example, and the method for identifying a partial discharge defect fault of a dc cable is not limited to identifying the insulation defect discharge model of the above type, and may also identify insulation defect faults of other dc cables known to those skilled in the art.
The defect fault models of the four different direct current cables are connected with the recognition system, the signal acquisition module is used for acquiring partial discharge pulse waveform signals, in the embodiment, the high-frequency current sensor is used for acquiring partial discharge pulse waveform signals, and the acquired partial discharge model pulse waveform signals refer to fig. 3. As shown in fig. 3, when the partial discharge occurs, the voltage fluctuates sharply, and thus, a start point of the fluctuation is a start point of the partial discharge signal. In some embodiments, the collected partial discharge pulse waveform signal has more interference information influence, so that the collected partial discharge pulse waveform signal is not easily obtained as a starting point, and therefore, the local discharge pulse waveform signal is preprocessed by adopting a Canny algorithm, so that effective information of the local discharge pulse waveform signal is extracted. It should be noted that in fig. 3, the ordinate represents the actual amplitude U of the point, and the abscissa represents the actual nth data point divided by 102The latter value, e.g. x being the abscissa of a certain point, is the x 10 th value2A data point.
Fig. 4 to fig. 7 respectively show effective information of partial discharge pulse waveform signals obtained by different insulation defect fault models by adopting the Canny algorithm. Wherein, fig. 4 shows the partial discharge pulse waveform signal under the fault model of the corona discharge insulation defect after pretreatment. Fig. 5 shows a partial discharge pulse waveform signal under an air gap discharge insulation defect fault model after pretreatment. Fig. 6 shows a partial discharge pulse waveform signal under a scratch discharge insulation defect fault model after pretreatment. Fig. 7 shows a partial discharge pulse waveform signal under a surface creepage discharge insulation defect fault model after pretreatment.
It should be noted that the ordinate in fig. 4 to 7 represents the actual amplitude U of the point divided by the maximum amplitude U of the waveformmThe resulting value, the abscissa representing the actual nth data point divided by 102The latter value, e.g. x being the abscissa of a certain point, is the x 10 th value2A data point.
The mechanism of the deep belief network based on RBM is shown in fig. 8. As shown in FIG. 8, RBM is a undirected probabilistic graphical model consisting of a visual layer and a hidden layer, where there is no connection between cells in the visual layer or the hidden layer. For RBM comprising a visual layer v and an implicit layer h, its energy function Eθ(v, h) can be expressed by the following formula:
in the formula, θ is an RBM model parameter, and θ ═ w, a, b }; a isiAnd bjAre respectively a display element viAnd hidden element hjBias of (3); w is aijIs a display unit viAnd hidden element hjA connection weight between; n isvAnd nhAre respectively a display element viAnd hidden element hjThe number of (2).
For the RBM-based deep belief network, the RBM-based deep belief network is formed by stacking a plurality of RBMs, and the hidden layer of the RBM at the lower layer is used as the input layer of the RBM at the upper layer, namely the RBMkRBM representing the kth iteration with hidden layer hkThe input layer is hk-1And RBMkInput layer h ofk-1Which is also the hidden layer of the RBM for the (k-1) th iteration.
The RBM is the RBM with the given significant element and hidden element unit number, the deep belief network based on the RBM is trained by adopting the training sample, so that the parameter theta of the RBM model is determined, and the training aim is to ensure that the data reconstructed by the RBM under the control of the parameter theta of the RBM model is consistent with the given training sample data as far as possible. And (4) carrying out fast and unsupervised training on the RBM layer by layer through a contrast divergence algorithm.
During training, firstly, the RBM model parameter theta is initialized randomly, and the training sample is used as a display element v before reconstructionoAnd calculating the hidden element h before reconstruction according to the following formula0：
In the above formula, Pθ(hj 0＝1|v0) Representing a given display element v0The jth hidden element hj 0Probability of setting as 1, sigmoid is activation function, sigmoid is (1+ exp (-x))-1To map x into the (0,1) interval.
The reconstructed display element v according to1：
In the above formula, Pθ(vi 1＝1|h0) Representing a given hidden element h0The ith display element vj 1Probability of being set to 1.
Then the reconstructed display element v1Recalculating to obtain reconstructed hidden element h1. And obtaining each parameter of the RBM-based deep belief network by analogy, wherein an updating formula is shown as the following formula:
in the formula, epsilon is the learning rate of a contrast divergence method; < > is a mathematical expectation. Δ wijIs a weight wijAmount of change of (a), Δ aiIs offset by aijAmount of change of (a), Δ bjIs offset by bijAmount of change of vi 0Is the ith display element before reconstruction， hj 0For the j-th hidden element before reconstruction, vi 1For the ith reconstructed rendering, hj 1Is the reconstructed jth hidden element.
In order to select a proper learning rate and easily converge to local optimum, an ADAM algorithm is adopted to carry out supervised training on a deep belief network to obtain optimized parameters.
During optimization, the parameter theta of the network is obtained in the k-th iterationkOn the basis of (1), obtaining a gradient by the following formula
Updating biased first moment estimate mk+1And biased second moment estimate vk+1：
In the formula, beta1And beta2Respectively, are hyper-parameters.
Further, the first moment deviation is obtained by the following formula
And second moment deviation
Finally, the parameter theta of the k +1 th iteration after updating is obtainedk+1：
Where α is the step size and τ is the stability constant.
In the present embodiment, the step α is 0.001 and the stability constant τ is 10-8Hyperparameter beta1And beta20.9 and 0.999 respectively.
In order to verify the identification effect of the scheme, the deep belief network based on the RBM trained in the scheme is taken as an example 1, and a comparative example 1 adopting a support vector machine and a comparative example 2 adopting a back propagation neural network are respectively drawn to obtain confusion matrices shown in fig. 9 to 11, wherein fig. 9 shows the confusion matrices of the insulation defect fault identification result identified by the comparative example 1. Fig. 10 shows a confusion matrix of the insulation defect fault identification results using the identification of comparative example 2. Fig. 11 shows a confusion matrix of the insulation defect fault identification results identified using embodiment 1.
In fig. 9 to 11, I represents a corona discharge insulation defect fault model, II represents a creepage discharge insulation defect fault model, III represents an air gap discharge insulation defect fault model, and IV represents a scratch discharge insulation defect fault model.
As can be seen from fig. 9 to 11, comparative examples 1 to 2 and example 1 both have good identification effects on defects I and II, and the recall rate exceeds 97.50%. However, for the defects of III and IV, the recall ratios of comparative example 1 and comparative example 2 were greatly reduced, with comparative example 1 having recall ratios of 92.03% and 92.73% for the two defects, respectively, and comparative example 2 having recall ratios of 90.32% and 91.82% for the two defects, respectively. Example 1 of the present case is more effective than comparative examples 1 and 2, and the recall rates at these two defects are 95.16% and 95.39%, respectively. Therefore, the identification effect of the embodiment 1 is best, and the defect identification method has high identification capability on the defects I and II and good identification effect on the defects III and IV.
Compared with the traditional deep belief network, the ADAM has higher efficiency of supervising and optimizing the network weight of the RBM-based deep belief network, so that the extracted features have stronger distinguishing capability on partial discharge defects, and the recognition effect is further improved. The identification capability of the defects III and IV in the comparative examples 1 and 2 is poor, because the air gaps and the insulation scratches are partially filled with air in the defect forming process of the defects III and IV, and certain similarity exists in the mechanism, so that the air gaps and the insulation scratches are mutually mistakenly identified as main types, so that the comparative examples 1 and 2 need to artificially extract the characteristics, otherwise, the difference of the partial discharge pulse waveforms of the defects III and IV is not enough to be sufficiently expressed. In the embodiment 1 of the scheme, the partial discharge waveform characteristics are mapped to the top layer of the deep belief network through unsupervised pre-training of the RBM. The characteristics comprise time-frequency and distribution characteristics and other deep characteristics which are difficult to observe, so that the characteristics can describe original data more accurately, and therefore the scheme has a more ideal identification effect.
In addition, in order to further verify the identification effect of the present application, effective information of the partial discharge pulse waveform signal obtained after the acquisition preprocessing is used as 6400 samples, 4000 of the 6400 samples are used as test samples, and the remaining samples are respectively used as training sets in scales of 400, 800, 1200, 1600, 2000 and 2400 to train the RBM-based deep belief network of the present application and comparative example 3 adopting the conventional deep belief network. The extracted feature quantity samples are also trained for comparative example 1 and comparative example 2 according to the same proportion, and the recognition results of different methods are listed in table 1.
Table 1.
As can be seen from table 1, the average recognition accuracy of example 1 and comparative examples 1 to 3 is above 90% at the set training set scale, and the recognition accuracy of comparative example 1, comparative example 2, comparative example 3 and example 1 is higher as the training set increases. Under the condition of small samples (for example, the sample size of the training set is 400 and 800), the identification effect of the embodiment 1 is equivalent to that of the comparative examples 1 to 3, but with the increase of the training set size, the partial discharge pulse waveform features extracted by the deep belief network are more comprehensive, and the average identification accuracy of the embodiment 1 is obviously better than that of the comparative examples 1 to 3. Meanwhile, the convergence speed of the improved deep belief network based on the ADAM algorithm and supervised fine tuning is increased in the embodiment 1 of the scheme, so that the method is faster, and the algorithm identification accuracy is higher under the same training scale.
It should be noted that the prior art in the protection scope of the present invention is not limited to the examples given in the present application, and all the prior art which is not inconsistent with the technical scheme of the present invention, including but not limited to the prior patent documents, the prior publications and the like, can be included in the protection scope of the present invention.
It should be noted that the combination of the features in the present application is not limited to the combination described in the claims or the combination described in the embodiments, and all the features described in the present application may be freely combined or combined in any manner unless contradictory to each other.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.
Claims (7)
1. A method for identifying partial discharge defect faults of a direct current cable is characterized by comprising the following steps:
(1) collecting partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable;
(2) extracting effective information of the partial discharge pulse waveform signal by adopting a Canny algorithm to obtain a training sample, wherein the effective information comprises a segment of the partial discharge pulse waveform signal which changes; the method for extracting the effective information of the partial discharge pulse waveform signal by adopting the Canny algorithm comprises the following steps:
smoothing the waveform signal f (x) of the local discharge pulse by adopting a one-dimensional Gaussian function to obtain a Gaussian-filtered waveform signal
Obtaining
The derivative g (x) of (a), performing non-maximum suppression, and only keeping the maximum point of the derivative;
performing double-threshold detection: setting a low threshold δlAnd a high threshold value deltahIf the derivative g (x) is less than δlThen mark xiNon-edge points, if greater than deltahMark point xiThe points are strong edge points, and the rest points are marked as weak edge points;
and (3) inhibiting isolated weak edge points: setting a neighborhood epsilon, and if no strong edge point exists in the weak edge point neighborhood epsilon, taking the weak edge point as a non-edge point; if the weak edge point exists, selecting the 1 st weak edge point as the starting point of the segment of the partial discharge pulse waveform signal, otherwise, selecting the 1 st strong edge point as the starting point of the segment of the partial discharge pulse waveform signal;
(3) constructing a deep belief network based on a limited Boltzmann machine, and performing unsupervised training on the deep belief network by adopting a training sample to obtain network parameters;
(4) carrying out supervised training on the deep belief network so as to optimize network parameters;
(5) and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network so as to obtain a recognition result from the output of the deep belief network.
2. The identification method according to claim 1, wherein the insulation defect discharge model includes at least an air gap discharge model, a corona discharge model, a scratch discharge model, and a creeping discharge model.
3. The recognition method of claim 1, wherein in step (3), the deep belief network is trained layer by layer using a contrastive divergence algorithm to obtain the network parameters.
4. The identification method according to claim 1, wherein in step (4), the deep belief network is supervised trained using an ADAM algorithm.
5. A system for identifying partial discharge fault faults in a dc cable, comprising:
the signal acquisition module is used for acquiring partial discharge pulse waveform signals of a plurality of insulation defect discharge models of the direct current cable;
the preprocessing module extracts effective information of the partial discharge pulse waveform signal by adopting a Canny algorithm to obtain a training sample; the effective information comprises a segment of which the partial discharge pulse waveform signal changes, and the extraction of the effective information of the partial discharge pulse waveform signal by adopting a Canny algorithm comprises the following steps:
smoothing the waveform signal f (x) of the local discharge pulse by adopting a one-dimensional Gaussian function to obtain a Gaussian-filtered waveform signal
Obtaining
The derivative g (x) of (a), performing non-maximum suppression, and only keeping the maximum point of the derivative;
performing double-threshold detection: setting a low threshold δlAnd a high threshold value deltahIf the derivative g (x) is less than δlThen mark xiNon-edge points, if greater than deltahMark point xiThe points are strong edge points, and the rest points are marked as weak edge points;
and (3) inhibiting isolated weak edge points: setting a neighborhood epsilon, and if no strong edge point exists in the weak edge point neighborhood epsilon, taking the weak edge point as a non-edge point; if the weak edge point exists, selecting the 1 st weak edge point as the starting point of the segment of the partial discharge pulse waveform signal, otherwise, selecting the 1 st strong edge point as the starting point of the segment of the partial discharge pulse waveform signal;
the signal processing module is used for carrying out unsupervised training on the constructed deep belief network based on the limited Boltzmann machine by adopting the training samples to obtain network parameters; carrying out supervised training on the deep belief network to obtain optimized network parameters;
and inputting a partial discharge pulse waveform signal to be recognized into the trained deep belief network, and obtaining a recognition result from the output of the deep belief network.
6. The recognition system of claim 5, wherein the signal processing module unsupervised trains the deep belief network to obtain the network parameters using a contrast divergence algorithm.
7. The identification system of claim 5, wherein the signal processing module employs an ADAM algorithm to supervised training of the deep belief network to obtain the optimized network parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811113943.XA CN109164362B (en) | 2018-09-25 | 2018-09-25 | Method and system for identifying partial discharge defect fault of direct current cable |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811113943.XA CN109164362B (en) | 2018-09-25 | 2018-09-25 | Method and system for identifying partial discharge defect fault of direct current cable |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109164362A CN109164362A (en) | 2019-01-08 |
| CN109164362B true CN109164362B (en) | 2021-01-08 |
Family
ID=64880196
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811113943.XA Active CN109164362B (en) | 2018-09-25 | 2018-09-25 | Method and system for identifying partial discharge defect fault of direct current cable |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109164362B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109799423B (en) * | 2019-01-09 | 2021-05-07 | 西安科技大学 | Cable fault on-line diagnosis method |
| CN112307851A (en) * | 2019-08-02 | 2021-02-02 | 上海交通大学烟台信息技术研究院 | Method and system for identifying bird nest on electric power iron tower |
| CN112305381A (en) * | 2020-09-21 | 2021-02-02 | 国网山东省电力公司临沂供电公司 | Method and system for monitoring and positioning online partial discharge of distribution cable |
| CN112363099B (en) * | 2020-10-30 | 2023-05-09 | 天津大学 | TMR current sensor temperature drift and geomagnetic field correction device and method |
| CN113376478B (en) * | 2021-06-22 | 2023-06-16 | 清华大学 | Power transmission line short-circuit fault positioning method based on edge detection |
| CN113848435B (en) * | 2021-08-31 | 2023-12-15 | 四川大学 | Direct-current partial discharge signal classification and identification method based on frequency spectrum energy distribution probability |
| CN114034990A (en) * | 2021-10-22 | 2022-02-11 | 中核核电运行管理有限公司 | High-voltage electrical equipment online diagnosis system and method |
| CN114252737B (en) * | 2021-12-23 | 2022-12-02 | 武汉大学 | State identification and evaluation method and system for power cable partial discharge development process |
| CN116736051A (en) * | 2023-05-22 | 2023-09-12 | 内蒙古工业大学 | GIS insulation defect intelligent diagnosis method, device, equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103954897A (en) * | 2014-05-20 | 2014-07-30 | 电子科技大学 | Intelligent power grid high voltage insulation damage monitoring system and method based on ultraviolet imagery |
| CN104700421A (en) * | 2015-03-27 | 2015-06-10 | 中国科学院光电技术研究所 | Adaptive threshold edge detection algorithm based on canny |
| CN107255779A (en) * | 2017-06-15 | 2017-10-17 | 国网河南省电力公司济源供电公司 | A kind of high-tension switch cabinet Analysis of Partial Discharge method |
| CN107449994A (en) * | 2017-07-04 | 2017-12-08 | 国网江苏省电力公司电力科学研究院 | Partial discharge method for diagnosing faults based on CNN DBN networks |
| CN107831409A (en) * | 2017-09-22 | 2018-03-23 | 国网山东省电力公司电力科学研究院 | The method and method for detecting abnormality of superfrequency partial discharge detection TuPu method parameter extraction |
-
2018
- 2018-09-25 CN CN201811113943.XA patent/CN109164362B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103954897A (en) * | 2014-05-20 | 2014-07-30 | 电子科技大学 | Intelligent power grid high voltage insulation damage monitoring system and method based on ultraviolet imagery |
| CN104700421A (en) * | 2015-03-27 | 2015-06-10 | 中国科学院光电技术研究所 | Adaptive threshold edge detection algorithm based on canny |
| CN107255779A (en) * | 2017-06-15 | 2017-10-17 | 国网河南省电力公司济源供电公司 | A kind of high-tension switch cabinet Analysis of Partial Discharge method |
| CN107449994A (en) * | 2017-07-04 | 2017-12-08 | 国网江苏省电力公司电力科学研究院 | Partial discharge method for diagnosing faults based on CNN DBN networks |
| CN107831409A (en) * | 2017-09-22 | 2018-03-23 | 国网山东省电力公司电力科学研究院 | The method and method for detecting abnormality of superfrequency partial discharge detection TuPu method parameter extraction |
Non-Patent Citations (4)
| Title |
|---|
| A Computational Approach to Edge Detection;JOHN CANNY;《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》;19861130;第8卷(第6期);第I-X节 * |
| ADAM : A METHOD FOR STOCHASTIC OPTIMIZATION;Diederik P. Kingma 等;《International Conference on Learning Representations2015》;20151231;第1-8节 * |
| 基于振荡波局部放电检测的电缆缺陷类型识别研究;龙红星;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20180715;正文第三-四章 * |
| 龙红星.基于振荡波局部放电检测的电缆缺陷类型识别研究.《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》.2018, * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109164362A (en) | 2019-01-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109164362B (en) | Method and system for identifying partial discharge defect fault of direct current cable | |
| CN108732465B (en) | Power distribution network fault positioning method based on wavelet transformation and CNN | |
| CN109685138B (en) | XLPE power cable partial discharge type identification method | |
| CN109212392B (en) | Method and system for identifying partial discharge defect fault of direct current cable and Caffe convolutional neural network | |
| CN110161388B (en) | Fault type identification method and system of high-voltage equipment | |
| CN108957251B (en) | Cable joint partial discharge mode identification method | |
| CN109102508B (en) | Method for identifying insulation defects of alternating current cable based on partial discharge image characteristics | |
| Sharkawy et al. | SVM classification of contaminating particles in liquid dielectrics using higher order statistics of electrical and acoustic PD measurements | |
| CN109085468A (en) | A kind of recognition methods of cable local discharge insulation defect | |
| CN110118928B (en) | Breaker fault diagnosis method based on error inverse propagation algorithm | |
| CN108693448B (en) | Partial discharge mode recognition system applied to power equipment | |
| CN106503612A (en) | Direct current cables local discharge signal mode identification method and system based on compressive sensing theory | |
| CN106338674B (en) | Based on the direct current cables splice insulation fault diagnosis method and system for improving ECOC classifier | |
| CN112613423B (en) | Epileptic brain electrical signal identification method based on machine learning | |
| CN116243115A (en) | High-voltage cable mode identification method and device based on time sequence topology data analysis | |
| Chang et al. | Application of fuzzy entropy to improve feature selection for defect recognition using support vector machine in high voltage cable joints | |
| CN112305379A (en) | Mode identification method and system for GIS insulation defect | |
| CN112147474A (en) | XLPE power cable typical defect partial discharge type identification system and method | |
| CN112557833A (en) | Cable partial discharge mode identification method based on depth sample enhancement | |
| CN115311584A (en) | Unmanned aerial vehicle high-voltage power grid video inspection floating hanging method based on deep learning | |
| CN114118149A (en) | Induction motor fault diagnosis system based on finite element simulation and symmetric feature migration | |
| CN109975634A (en) | A kind of fault diagnostic method for transformer winding based on atom sparse decomposition | |
| Rostaminia et al. | An efficient partial discharge pattern recognition method using texture analysis for transformer defect models | |
| CN113608073A (en) | Cable partial discharge pulse separation method under variable frequency resonance | |
| Sharkawy et al. | Particle identification in terms of acoustic partial discharge measurements in transformer oils |
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 |