CN117828333A - Cable partial discharge feature extraction method based on signal mixing enhancement and CNN - Google Patents
Cable partial discharge feature extraction method based on signal mixing enhancement and CNN Download PDFInfo
- Publication number
- CN117828333A CN117828333A CN202410011871.7A CN202410011871A CN117828333A CN 117828333 A CN117828333 A CN 117828333A CN 202410011871 A CN202410011871 A CN 202410011871A CN 117828333 A CN117828333 A CN 117828333A
- Authority
- CN
- China
- Prior art keywords
- enhancement
- partial discharge
- image
- layer
- signal
- 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
Links
- 238000000605 extraction Methods 0.000 title claims abstract description 23
- 238000002156 mixing Methods 0.000 title claims abstract description 12
- 238000010586 diagram Methods 0.000 claims abstract description 63
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 61
- 238000000034 method Methods 0.000 claims abstract description 39
- 238000012937 correction Methods 0.000 claims abstract description 16
- 230000003044 adaptive effect Effects 0.000 claims abstract description 14
- 230000008569 process Effects 0.000 claims abstract description 12
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 230000004927 fusion Effects 0.000 claims description 33
- 230000006870 function Effects 0.000 claims description 32
- 238000011176 pooling Methods 0.000 claims description 31
- 238000000354 decomposition reaction Methods 0.000 claims description 30
- 238000004422 calculation algorithm Methods 0.000 claims description 27
- 238000012545 processing Methods 0.000 claims description 23
- 230000009466 transformation Effects 0.000 claims description 20
- 238000004364 calculation method Methods 0.000 claims description 11
- 230000000694 effects Effects 0.000 claims description 10
- 238000005457 optimization Methods 0.000 claims description 10
- 238000013519 translation Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 7
- 238000013507 mapping Methods 0.000 claims description 7
- 239000000284 extract Substances 0.000 claims description 6
- 230000004807 localization Effects 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000005192 partition Methods 0.000 claims description 6
- 230000009467 reduction Effects 0.000 claims description 5
- 238000001228 spectrum Methods 0.000 claims description 5
- 238000012549 training Methods 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 4
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 3
- 125000003275 alpha amino acid group Chemical group 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 150000001875 compounds Chemical class 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 3
- 238000003384 imaging method Methods 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 2
- 238000009413 insulation Methods 0.000 description 7
- 230000008901 benefit Effects 0.000 description 4
- 239000011810 insulating material Substances 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 230000005684 electric field Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000006866 deterioration Effects 0.000 description 2
- 239000012774 insulation material Substances 0.000 description 2
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000008602 contraction Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000012535 impurity Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000004091 panning Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
Landscapes
- Image Processing (AREA)
Abstract
The invention relates to a cable partial discharge characteristic extraction method based on signal mixing enhancement and CNN, and belongs to the technical field of power cables. The method comprises the following steps: s1: acquiring an originally acquired cable partial discharge signal diagram; s2: partial discharge signal preprocessing based on wavelet transform denoising and non-lower shear wave transform; s3: hybrid partial discharge signal enhancement based on adaptive parameter gamma correction and BPDFHE; s4: feature extraction based on convolutional neural networks. The invention can be used for identifying partial discharge characteristic information in the online operation process of the power cable, effectively removing noise and optimizing data, and can be used for establishing a more efficient and quicker auxiliary decision model for cable fault classification and providing effective data.
Description
Technical Field
The invention belongs to the technical field of power cables, and relates to a cable partial discharge feature extraction method based on signal mixing enhancement and CNN.
Background
Partial discharge means that a discharge is generated in the insulating material under the effect of an electric field, but such a discharge does not extend through the entire insulating material, but only in a partial region. Any factor that may cause maldistribution of the electric field of the insulation system may become a cause of partial discharge. Such as insulation structure design, have drawbacks; impurities are mixed into the insulating material to cause insufficient purity; problems in the manufacturing process create bubbles, cracks, and the like. These causes create air gaps in the insulating material where partial discharges are created when the electric field strength is high to some extent. If measures are not taken in time, long-time partial discharge can gradually develop into dendritic discharge, so that surface flashover is caused, insulation degradation is caused and is continuously expanded, further, the service life and insulation performance of the insulation material are reduced, and even the insulation material is possibly broken down, so that large-scale electric power accidents are caused. Therefore, partial discharge is a sign of insulation deterioration and a cause of insulation deterioration. In order to avoid such power accidents, detection of partial discharge of power equipment is very important in the case of enabling normal power transmission and transformation.
The existing cable fault partial discharge characteristic recognition methods are quite many, but generally comprise a wavelet transformation method, a statistical parameter method, a fractal method and a pulse waveform method.
1. Wavelet transformation
In 1974, based on the research of fourier transform, the french engineer Morlet proposed a method for processing signals by using wavelet transform for the first time, and through the research in recent years, the method for processing signals by using wavelet transform is mature. Conventional fourier analysis methods are suitable for processing stationary signals, and use thereof for analyzing non-stationary high frequency signals on electrical lines have limitations. The wavelet transformation can simultaneously characterize the characteristic information of the signal in the time domain and the frequency domain, and is particularly suitable for processing the non-stationary abrupt change signal of the high-frequency signal of the cable line.
2. Fractal method
In 1990, jacquin proposed a block fractal coding algorithm based on a local iteration function, so that complexity of fractal coding of a common image is greatly reduced, automatic segmentation and automatic coding of the image are realized, a gate for fractal coding practicality is opened, and various improved algorithms proposed by a learner are all algorithms based on Jacquin, so that the block fractal coding algorithm is also called a basic fractal coding algorithm. During encoding, firstly dividing an image into a definition domain block and a value domain block, performing spatial domain contraction and equidistant transformation on the definition domain block to generate a codebook, and then searching and matching each sub-block in the definition domain in the codebook to complete fractal encoding, wherein the flow is shown in figure 1. And during decoding, finding out an optimal matching block corresponding to each value range block according to the fractal code, iteratively recovering each value range block through compression transformation, and splicing all the value range blocks to form a final decoded image.
In summary, in order to solve the problem of accuracy of identifying the partial discharge characteristics of the cable fault, a number of methods have emerged, and some achievements in the fields of digital processing technology and artificial intelligence have been widely used. Thus, to date, partial discharge fault signature has remained the hotspot in partial discharge on-line monitoring research.
A discharge waveform is generated during the partial discharge, which is characterized by a high frequency, but very weak signal amplitude, typically only a few microvolts. Due to this weak nature of the signal amplitude, the disturbance signal is easily submerged once it exists outside. Therefore, in the case of partial discharge signal feature recognition, it is necessary to take into consideration the influence of various causes, and it is particularly important to take various effective feature recognition measures.
1. Firstly, because the collected cable fault signals may contain unnecessary noise, a certain denoising filtering algorithm is needed to perform preprocessing.
2. Secondly, the characteristic information of the partial discharge signal diagram is weak, the accuracy of defect identification is improved by carrying out enhancement processing on the cable partial discharge signal diagram, and an effective signal diagram enhancement method is provided for the cable partial discharge signal diagram.
3. And (3) carrying out effective feature extraction on the enhanced signal and analyzing main components of the enhanced signal.
Disclosure of Invention
In view of the above, the present invention aims to provide a cable partial discharge feature extraction method based on signal mixing enhancement and CNN.
In order to achieve the above purpose, the present invention provides the following technical solutions:
a cable partial discharge feature extraction method based on signal mixing enhancement and CNN, the method comprising the following steps:
s1: acquiring an originally acquired cable partial discharge signal diagram;
s2: partial discharge signal preprocessing based on wavelet transform denoising and non-lower shear wave transform;
s3: hybrid partial discharge signal enhancement based on adaptive parameter gamma correction and BPDFHE;
s4: feature extraction based on convolutional neural networks.
Optionally, the S2 specifically is:
s21: wavelet denoising of cable original signal
For a one-dimensional signal x (n), the discrete wavelet transform W is expressed as:
wherein a and b represent scale and translation parameters, respectively, ψ a,b (n) is a wavelet function; the scale parameter a controls the width of the wavelet function, while the translation parameter b controls the position of the wavelet function;
s22: filtering the partial discharge signal:
(1) wavelet divisionSolution: selecting Daubechies wavelet basis function, decomposing cable original partial discharge signal into wavelet coefficients c of multiple scales through wavelet transformation j,k Where j represents a scale and k represents a translation; c j,k ＝<x(n),ψ j,k (n)>，<·,·>Representing an inner product operation;
(2) and (3) threshold processing: for wavelet coefficient c j,k Soft thresholding is performed, formulated as:
c′ j,k ＝sign(c j,k )·(|c j,k |-threshold) + (2)
wherein sign (·) represents a signed function, (·) + Representing a positive part;
(3) wavelet reconstruction: the processed wavelet coefficient c 'is transformed by inverse wavelet' j,k Reconstructing a denoised signal x' (n);
s23: non-downsampled shear wave conversion
The non-downsampled shear wave transform (non-subsample shearlet transform, NSST) consists of two parts, multiscale decomposition and direction localization;
the multi-scale decomposition is accomplished by using a pyramid filter bank that is not downsampled, and the filters used in each stage of decomposition are all from the filters used in the previous stage in a matrixThe up-sampled filter is carried out, so that the down-sampling operation is avoided; after the image is decomposed by a first-stage non-downsampling pyramid filter (NonSubsampled Pyramid, NSP), a low-frequency subband image and a high-frequency subband image can be obtained, NSP decomposition from the second stage is iteratively performed on the upper-stage low-frequency component, and finally a low-frequency subband image and k high-frequency subband images are obtained, wherein the sizes of the subband images are the same as the original image;
the localization of the direction is achieved by shifting the window function in the pseudo-polar grid system by using a standard Shear Filter (SF), mapping the SF in the pseudo-polar grid system to a Cartesian coordinate system and performing an inverse Fourier transform, and then passing through twoThe dimensional convolution obtains an improved shear filter, and ensures translational invariance of non-downsampled shear wave transformation; carrying out L-level direction decomposition on the sub-band image under the k-level scale to obtain 2 L A subband image;
after NSST processing is carried out on the cable partial discharge signal diagram, a low-frequency sub-band diagram (Low frequency directional subband image, LFDSI) and k high-frequency direction sub-band diagrams (High frequency directional subband image, HFDS) are obtained for extracting features of each sub-image by a subsequent method.
Optionally, the S3 specifically is:
s31: gamma corrected low frequency signal enhancement of adaptive parameters
(1) Adaptive gamma
Adopting gamma transformation of self-adaptive parameters to enhance the preliminary noise-reduced image to obtain n preprocessed images; let the input image be in uint8 format, the mathematical form of the gamma transformation of the adaptive parameters be:
γ＝log 2 (m-F median ) (4)
in the formula (3) and the formula (4), I in Brightness for the input image; i out Representing the brightness of the output image; c is a proportionality coefficient, 1 is taken, F represents I in Is a normalized luminance of (2); f (F) median To take I in A median value; gamma is a gamma correction index, the value of which is determined by a target brightness average value m, and m takes the following value:
in the formula (5), the amino acid sequence of the compound,is I in Is the average value of (2); the combination of the formula (3), the formula (4) and the formula (5) can ensure that the brightness of the input n Zhang Chuci noise reduction chart sequence is corrected to be in a uniform range, thereby effectively overcoming the brightnessInterference caused by the degree fluctuation on imaging;
s32: BPDFHE high frequency signal enhancement
Implementing image enhancement on a plurality of high-frequency direction sub-band diagrams in the cable partial discharge signal diagram decomposed by NSST, wherein an algorithm is BPDFHE; the method comprises the following specific steps:
s321: calculating a fuzzy histogram;
s322: partitioning the histogram for convenient classification;
s323: equalizing the dynamic histogram to improve performance;
s324: the brightness of the normalized signal diagram is convenient to calculate;
s325: finally, a high-frequency subband signal diagram after BPDFHE enhancement is obtained;
s323 is divided into partition mapping and equalization operation;
s3231: partition mapping to dynamic range
Δp i ＝h i -l i (6)
In the formulas (6) to (9): l (L) i And h i The lowest and highest intensity values of the ith sub-graph are used for defining a section; input histogram sub-graph and intensity difference Δp for dynamic range of output sub-graph i And adjusting parameter r i Determining, starting with S ai Stop is S ti ；
S3232: equalizing each histogram subgraph
Equalizing each histogram sub-graph operates similarly to global equalization, and the equalized ith sub-graph is shown in formula (10)
Wherein: y (j) is the intensity value of the original signal diagram at the j-th level;
s33: weighting each sub-band map
Based on NSST reconstruction, weighting the sub-band diagram obtained after the enhancement in the previous step in a Cartesian coordinate system and a pseudo-polar coordinate system; NSST decomposition is carried out on the signal diagram to obtain a plurality of low-frequency and high-frequency sub-band diagrams, the enhancement effect applicable to each of gamma correction and BPDFHE can be more fully exerted, and NSST reconstruction is carried out by further combining the algorithm idea of weighted fusion; firstly, respectively using gamma correction and BPDFHE enhancement algorithm to carry out image enhancement on a low-frequency and high-frequency direction sub-band diagram after NSST decomposition; then, carrying out weighted average fusion on the enhanced subgraphs, and finally reconstructing a new enhanced partial discharge signal diagram;
let LFDSI after Gamma enhancement be alpha I 1 Each HFDSI after BPDFHE enhancement is (2-alpha) I h I=1, 2, …,8; wherein, alpha is a weighting coefficient, which is critical in the processing, the value of alpha is 0 to 2, different enhancement reconstruction images are calculated through repeated adjustment of the weighting coefficient and the NSST decomposition reconstruction, and the definition, the gray average value, the edge strength and the contrast of the enhancement signal image under each weighting coefficient are compared, so that the weighting coefficient is finally determined.
Optionally, the S4 specifically is:
s41: building convolutional neural network
The deep neural network of the two-dimensional plane data consists of a convolution layer, a pooling layer and a full connection layer;
for a convolution layer, the feature diagram output by the previous layer is subjected to convolution calculation by a convolution kernel, and the feature diagram of the layer is obtained after a function is activated, wherein the specific process is as follows:
wherein: c (C) i Represents a characteristic diagram of the ith layer of CNN,representing convolution calculations, W i Is the weight matrix of the convolution kernel of the ith layer, b i For the bias vector of the i-th layer, σ represents the activation function; during network training, the linear rectification function (Rectified Linear Unit, reLU) reduces the calculated amount by debugging the activity of neurons in the network, improves the expression capacity of the network, has
The calculation process of the pooling layer is expressed as:
C i ＝p(C i-1 ) (13)
wherein p represents a pooling operation;
s42: basic principle of cross-layer feature fusion and CNN optimization
Designing a Cross-layer feature fusion and optimization CNN model (Cross-layer Feature Fusion and Optimization CNN, CFFO-CNN); improving a network structure by introducing cross-layer connection, selecting a plurality of pooled layer output characteristics capable of effectively expressing signal characteristics, and completing fusion of deep and shallow layer characteristics so that the network can autonomously extract essential characteristics of signals from shallow and deep places;
firstly, inputting a preprocessed image into a convolutional neural network to perform convolutional transformation, wherein the network comprises n convolutional layers (conv) and n pooling layers (pool); then, selecting k (k=1, …, n) pooling layer output feature graphs to perform feature fusion, wherein the fused features are expressed as:
F＝[X pool1 X pool2 …,X poolk ] (14)
wherein: x is X pool1 ,X poot2 …,X poolk Outputting characteristic graphs for the 1 st, 2 nd and … th and k th pooling layers; shielding part of neurons of the full-connection layer by adopting a Dropout method;
finally, inputting the fused feature images into a full-connection layer for classification;
s43: gamma-BPDFHE enhancement and CFFO-CNN based feature extraction
S431: outputting the Gamma-BPDFHE enhanced time-frequency spectrogram of the partial discharge signal into an image format, carrying out gray level processing, and compressing the image input size with the same proportion to 64 multiplied by 64:
s432: constructing a basic structure of CNN, wherein the network comprises 2 convolution layers and corresponding pooling layers, the convolution kernel size of the network is set to be 5 multiplied by 5, the pooling kernel size is 2 multiplied by 2, the batch size is 8, the iteration number is 200, and an RLU activation function, a maximum pooling mode and a random gradient descent (Stochastic Gradient Descent with Momentum, SGDM) algorithm based on momentum are adopted;
s433: introducing cross-layer connection, and determining a feature fusion layer with highest recognition accuracy by comparing recognition effects of different feature fusion layer structure networks;
s434: and optimizing the super parameters of the cross-layer feature fusion convolutional neural network by using an Adam algorithm to obtain CFFO-CNN, and extracting the intrinsic features of the spectrum image when the local discharge enhanced signal is extracted by using the CFFO-CNN.
The invention has the beneficial effects that: the method can be used for identifying partial discharge characteristic information in the online operation process of the power cable, effectively removing noise and optimizing data, and can be used for establishing a more efficient and quicker auxiliary decision-making model for cable fault classification and providing effective data.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the instrumentalities and combinations particularly pointed out in the specification.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in the following preferred detail with reference to the accompanying drawings, in which:
FIG. 1 is a flow chart of the present invention;
FIG. 2 is an exploded view of NSST;
FIG. 3 is a typical structure of a CNN network;
FIG. 4 is a flow chart of CNN feature extraction;
fig. 5 is a schematic diagram of the present invention.
Detailed Description
Other advantages and effects of the present invention will become apparent to those skilled in the art from the following disclosure, which describes the embodiments of the present invention with reference to specific examples. The invention may be practiced or carried out in other embodiments that depart from the specific details, and the details of the present description may be modified or varied from the spirit and scope of the present invention. It should be noted that the illustrations provided in the following embodiments merely illustrate the basic idea of the present invention by way of illustration, and the following embodiments and features in the embodiments may be combined with each other without conflict.
Wherein the drawings are for illustrative purposes only and are shown in schematic, non-physical, and not intended to limit the invention; for the purpose of better illustrating embodiments of the invention, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the size of the actual product; it will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The same or similar reference numbers in the drawings of embodiments of the invention correspond to the same or similar components; in the description of the present invention, it should be understood that, if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc., that indicate an azimuth or a positional relationship based on the azimuth or the positional relationship shown in the drawings, it is only for convenience of describing the present invention and simplifying the description, but not for indicating or suggesting that the referred device or element must have a specific azimuth, be constructed and operated in a specific azimuth, so that the terms describing the positional relationship in the drawings are merely for exemplary illustration and should not be construed as limiting the present invention, and that the specific meaning of the above terms may be understood by those of ordinary skill in the art according to the specific circumstances.
As shown in fig. 1, the present invention includes the following parts:
s1: acquiring an originally acquired cable partial discharge signal diagram;
s2: partial discharge signal preprocessing based on wavelet transform denoising and non-lower shear wave transform;
s3: hybrid partial discharge signal enhancement based on adaptive parameter gamma correction and BPDFHE;
s4: feature extraction based on convolutional neural networks;
1. partial discharge signal preprocessing based on wavelet transform denoising and non-lower shear wave transform
1. Wavelet denoising of cable original signal
The core idea of wavelet denoising is to thresholde the high frequency subbands of the signal in the domain of the wavelet transform, setting the noise figure to zero or reducing to a small value. In this way, high frequency noise components in the signal can be eliminated, preserving the dominant signal characteristics. For a one-dimensional signal x (n), its discrete wavelet transform W can be expressed as:
wherein a and b represent scale and translation parameters, respectively, ψ a,b (n) is a wavelet function. The scale parameter a controls the width of the wavelet function and the panning parameter b controls the position of the wavelet function.
The invention filters the partial discharge signal by the following steps:
(1) wavelet decomposition: selecting Daubechies wavelet basis function, decomposing cable original partial discharge signal into wavelet coefficients c of multiple scales through wavelet transformation j,k Where j represents the scale and k represents the translation. c j,k ＝<x(n),ψ j,k (n)>(here<·,·>Representing an inner product operation. )
(2) And (3) threshold processing: for wavelet coefficient c j,k Soft thresholding is performed, formulated as:
c′ j,k ＝sign(c j,k )·(|c j,k |-threshold) + (2)
wherein sign (·) represents a signed function, (·) + Representing the positive part.
(3) Wavelet reconstruction: the processed wavelet coefficient c 'is transformed by inverse wavelet' j,k And reconstructing the denoised signal x' (n).
3. Non-downsampled shear wave conversion
The non-downsampled shear wave transform (non-subsample shearlet transform, NSST) consists of two parts, multiscale decomposition and direction localization.
The multi-scale decomposition process is mainly accomplished by using a pyramid filter bank that is not downsampled, which is an orthogonal transform that uses only upsampling. Since the filters used in each stage of decomposition are all from the filters used in the previous stage according to a matrixThe upsampled filter is performed, thus avoiding a downsampling operation. After the image is decomposed by a first-stage non-downsampling pyramid filter (NonSubsampled Pyramid, NSP), a low-frequency subband image and a high-frequency subband image can be obtained, NSP decomposition from the second stage is performed iteratively on the upper-stage low-frequency component, and the images are repeatedly repeated, so that a low-frequency subband image and k (k-stage NSP decomposition) high-frequency subband images can be finally obtained, and the sizes of the subband images are the same as that of the original image.
The directional localization of the shear wave transformation is achieved by shifting the window function in the pseudo-polarization grid system using standard Shear Filters (SF), but downsampling is inevitably taken during this operation and thus the shift invariance is lost. For this purpose, SF in the pseudo-polar grid system is mapped to a cartesian coordinate system and subjected to inverse fourier transform, and then an improved shear filter is obtained through two-dimensional convolution, so that translational invariance of non-downsampled shear wave transform is ensured. If the sub-band image under the k-th scale is subjected to L-level direction decomposition, 2 can be obtained L Subband images. The process of NSST multi-scale multi-directional decomposition is shown in FIG. 2. FIG. 2 shows NSST being subjected to three-level scale divisionThe solution process, wherein the direction progression L of the decomposition on the first level, second level and third level scales is 3, 3 and 2, respectively.
As can be seen from fig. 2, after NSST processing is performed on the cable partial discharge signal diagram, a low-frequency subband diagram (Low frequency directional subband image, LFDSI) and k high-frequency subband diagrams (High frequency directional subband image, HFDS) are obtained for the subsequent method to extract features from each sub-image.
2. Hybrid partial discharge signal enhancement based on adaptive parameter gamma correction and BPDFHE
LFDSI and HFDS obtained by decomposing the original cable partial discharge signal diagram are adopted to process low-frequency components by adopting the gamma correction of self-adaptive parameters, and high-frequency components are processed by adopting a brightness-maintaining dynamic fuzzy histogram equalization algorithm (brightness preserving dynamic fuzzy histogram equalization, BPDFHE) so as to achieve the mixing enhancement of signals.
1. Gamma corrected low frequency signal enhancement of adaptive parameters
(1) Adaptive gamma
Conventional gamma correction results in a fixed degree of brightness variation for each pixel of an image due to the fixed gamma value setting; if the unified gamma value is set to correct the input image sequence, the brightness fluctuation of the corrected image sequence is larger, and the noise reduction failure of the subsequent algorithm is possibly caused. In order to solve the problem, the invention adopts the gamma transformation of the self-adaptive parameters to enhance the image with preliminary noise reduction to obtain n preprocessed images. Assuming that the input image is in uint8 format, the mathematical form of the gamma transformation of the adaptive parameters is:
γ＝log 2 (m-F median ) (4)
in the formula (3) and the formula (4), I in Brightness for the input image; i out Representing the brightness of the output image; c is a proportionality coefficient, usually 1, F is I in Is a normalized luminance of (2); f (F) median To take I in A median value; gamma is a gamma correction index, the value of which is determined by a target brightness average value m, and m takes the following value:
in the formula (5), the amino acid sequence of the compound,is I in Is a mean value of (c). The combination of the formula (3), the formula (4) and the formula (5) can ensure that the brightness of the input n Zhang Chuci noise reduction chart sequence is corrected to be in a uniform range, and effectively overcomes the interference of brightness fluctuation on imaging.
BPDFHE high frequency signal enhancement
The BPDFHE algorithm is mainly used in the field of image enhancement which needs to be compatible with contrast enhancement and relatively low operation complexity. The invention realizes image enhancement of a plurality of high-frequency direction subband images in the cable partial discharge signal image decomposed by the section NSST, and the main algorithm is BPDFHE. The specific steps are as follows: (1) computing a fuzzy histogram; (2) partitioning the histogram for classification; (3) equalizing the dynamic histogram to improve performance; (4) normalizing the brightness of the signal diagram to facilitate calculation; (5) And finally obtaining the high-frequency subband signal diagram after BPDFHE enhancement. Step (3) is a key step and can be divided into 2 stages, namely partition mapping and equalization operation.
(1) Partition mapping to dynamic range
Δp i ＝h i -l i (6)
In the formulas (6) to (9): l (L) i And h i The lowest and highest intensity values of the ith sub-graph are used for defining a section; input histogram sub-graph and intensity difference Δp for dynamic range of output sub-graph i And adjusting parameter r i Determining, starting with S ai Stop is S ti 。
(2) Equalizing each histogram subgraph
Equalizing each histogram sub-graph operates similarly to global equalization, with the equalized ith sub-graph, e.g. of formula (10)
Wherein: y (j) is the intensity value of the original signal diagram at the j-th level.
3. Weighting each sub-band map
Based on NSST reconstruction, the sub-band diagram obtained after the enhancement in the previous step is weighted and reconstructed in a Cartesian coordinate system and a pseudo-polar coordinate system. The NSST decomposition is carried out on the signal diagram to obtain a plurality of low-frequency and high-frequency sub-band diagrams, the enhancement effect applicable to each of the gamma correction and the BPDFHE can be fully exerted, and the NSST reconstruction is carried out by further combining the algorithm idea of weighted fusion. Firstly, respectively using gamma correction and BPDFHE enhancement algorithm to carry out image enhancement on the low-frequency and high-frequency direction sub-band diagrams after NSST decomposition. And then carrying out weighted average fusion on the enhanced subgraphs, and finally reconstructing a new enhanced partial discharge signal diagram.
Let LFDSI after Gamma enhancement be alpha I 1 Each HFDSI after BPDFHE enhancement is (2-alpha) I h I=1, 2, …,8. Wherein, alpha is a weighting coefficient, which is critical in the processing, the value of alpha is 0 to 2, different enhancement reconstruction modes can be calculated through repeated adjustment of the weighting coefficient and the NSST decomposition reconstruction, and indexes such as definition, gray average value, edge intensity, contrast and the like of the enhancement signal diagram under each weighting coefficient are compared, so that the weighting coefficient is finally determined.
3. Feature extraction based on convolutional neural network
1. Principle of convolutional neural network
CNN is a deep neural network for processing two-dimensional plane data, and is generally composed of a convolution layer, a pooling layer, a full connection layer, and the like, and a typical structure is shown in fig. 3.
The convolution layer mainly extracts the characteristics of the input data, and because the convolution layer has weight sharing property, learning parameters are effectively reduced, and the overfitting phenomenon can be reduced. For a convolution layer, the feature diagram output by the previous layer is subjected to convolution calculation by a convolution kernel, and the feature diagram of the layer is obtained after a function is activated, and the specific process can be expressed as follows:
wherein: c (C) i Represents a characteristic diagram of the ith layer of CNN,representing convolution calculations, W i Is the weight matrix of the convolution kernel of the ith layer, b i Sigma represents the activation function for the bias vector of the i-th layer. In network training, the conventional activation function Sigmoid has the problems of gradient disappearance, long time consumption and the like, compared with the conventional linear rectification function (Rectified Linear Unit, reLU) which is relatively commonly used at present, effectively reduces the calculated amount, improves the expression capacity of the network by debugging the liveness of neurons in the network, and has the following characteristics of
Since the resolution of the image is generally high, a large burden is imposed on the computer operation if all the features extracted by all the convolution layers are input into the CNN model. Therefore, the aggregation statistics of partial features of the image can effectively reduce the calculation cost and prevent the overfitting phenomenon, which is the pooling operation. Because continuous pooling has translational invariance, the training result is not affected. The calculation process of the pooling layer can be expressed as:
C i ＝p(C i-1 ) (13)
where p represents the pooling operation, average pooling and maximum pooling are common methods. After the rolling and pooling operations, the network has higher distortion tolerance capability when the input samples are identified.
2. Basic principle of cross-layer feature fusion and CNN optimization
Currently, conventional CNNs typically extract semantic information contained in deep features from top to bottom, while ignoring image details contained in shallow features. Therefore, the feature extraction capability of the traditional CNN still has room for improvement, and on the basis, the invention designs a Cross-layer feature fusion and optimization CNN model (Cross-layer Feature Fusion and Optimization CNN, CFFO-CNN). By introducing cross-layer connection to improve the network structure and selecting a plurality of pooled layer output characteristics capable of effectively expressing signal characteristics, fusion of deep and shallow layer characteristics is completed, so that the network can autonomously extract essential characteristics of signals from shallow and deep places.
Firstly, inputting a preprocessed image into a convolutional neural network to perform convolutional transformation, wherein the network comprises n convolutional layers (conv) and n pooling layers (pool); then, k (k=1, …, n) pooling layer output feature graphs are selected for feature fusion, and the fused features can be expressed as:
F＝[X pool1 X pool2 …,X poolk ] (14)
wherein: x is X pool1 ,X poot2 …,X poolk The feature map is output for the 1,2, …, k pooled layers. In addition, the dimension surge phenomenon generated after feature fusion can increase the calculated amount and the risk of overfitting, and in order to solve the problem, a Dropout method is adopted to shield part of neurons of the full-connection layer. And finally, inputting the fused feature images into a full-connection layer for classification.
3. Feature extraction method based on Gamma-BPDFHE enhancement and CFFO-CNN
The invention provides a feature extraction method based on Gamma-BPDFHE enhancement and CFFO-CNN, which can realize the direct extraction of time spectrum image features by using the network model and avoid complex artificial feature design engineering and feature loss phenomenon. The feature extraction steps based on Gamma-BPDFHE enhancement and CFFO-CNN are as follows:
step 1: outputting the Gamma-BPDFHE enhanced time-frequency spectrogram of the partial discharge signal into an image format, carrying out gray level processing, and compressing the image input size with the same proportion to 64 multiplied by 64:
step 2: the basic structure of CNN is built, the network contains 2 convolution layers and corresponding pooling layers, the convolution kernel size of the network is set to be 5 multiplied by 5, the pooling kernel size is 2 multiplied by 2, the batch size is 8, the iteration number is 200, and an RLU activation function, a maximum pooling mode and a random gradient descent (Stochastic Gradient Descent with Momentum, SGDM) algorithm based on momentum are adopted.
Step 3: introducing cross-layer connection, and determining an optimal feature fusion layer by comparing the identification effects of different feature fusion layer structure networks;
step 4: and optimizing the super parameters of the cross-layer feature fusion convolutional neural network by using an Adam algorithm, so as to obtain CFFO-CNN, and extracting the intrinsic features of the spectrum image when the local discharge enhanced signal is extracted by using the CFFO-CNN.
A flow chart based on Gamma-BPDFHE images and CFFO-CNN is shown in fig. 4.
Examples
As shown in fig. 5, the cable partial discharge characteristic extraction method based on signal mixing enhancement and CNN comprises the following steps:
(1) And processing the direct current cable partial discharge signal diagrams with different typical insulation defects by utilizing non-downsampled shear wave conversion (non-subsample shearlet transform, NSST), carrying out signal diagram enhancement processing on the low-frequency subband diagram and each high-frequency direction subband diagram after NSST decomposition by utilizing a Gamma-BPDFHE algorithm, and then obtaining the cable partial discharge signal diagram after enhancement processing by utilizing an addition weight construction algorithm on each enhanced subband diagram.
(2) On the basis of a discharge signal time-frequency spectrogram, a feature extraction model based on cross-layer feature fusion optimization convolution network is designed. Firstly, gray-scale and compression processing are carried out on a Gamma-BPDFHE time spectrum image: secondly, determining a basic structure of a convolutional network through a comparison experiment, and introducing cross-layer connection and an Adam algorithm to improve the network structure, so as to build a cross-layer feature fusion and optimized convolutional neural network model;
(3) And finally, training the network to learn and extract the internal characteristics of the Gamma-BPDFHE image, and further identifying the discharge characteristics of the test sample.
Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the present invention, which is intended to be covered by the claims of the present invention.
Claims (4)
1. A cable partial discharge characteristic extraction method based on signal mixing enhancement and CNN is characterized in that: the method comprises the following steps:
s1: acquiring an originally acquired cable partial discharge signal diagram;
s2: partial discharge signal preprocessing based on wavelet transform denoising and non-lower shear wave transform;
s3: hybrid partial discharge signal enhancement based on adaptive parameter gamma correction and BPDFHE;
s4: feature extraction based on convolutional neural networks.
2. The method for extracting the cable partial discharge characteristics based on signal mixing enhancement and CNN according to claim 1, wherein the method comprises the following steps: the step S2 is specifically as follows:
s21: wavelet denoising of cable original signal
For a one-dimensional signal x (n), the discrete wavelet transform W is expressed as:
wherein a and b represent scale and translation parameters, respectively, ψ a,b (n) is a wavelet function; the scale parameter a controls the width of the wavelet function, while the translation parameter b controls the position of the wavelet function;
s22: filtering the partial discharge signal:
(1) wavelet decomposition: selecting Daubechies wavelet basis function, decomposing cable original partial discharge signal into wavelet coefficients c of multiple scales through wavelet transformation j,k Where j represents a scale and k represents a translation; c j,k ＝<x(n),ψ j,k (n)>，<·,·>Representing an inner product operation;
(2) and (3) threshold processing: for wavelet coefficient c j,k Soft thresholding is performed, formulated as:
c′ j,k ＝sign(c j,k )·(|c j,k |-threshold) + (2)
wherein sign (·) represents a signed function, (·) + Representing a positive part;
(3) wavelet reconstruction: the processed wavelet coefficient c 'is transformed by inverse wavelet' j,k Reconstructing a denoised signal x' (n);
s23: non-downsampled shear wave conversion
The non-downsampled shear wave transform (non-subsample shearlet transform, NSST) consists of two parts, multiscale decomposition and direction localization;
the multi-scale decomposition is accomplished by using a pyramid filter bank that is not downsampled, and the filters used in each stage of decomposition are all from the filters used in the previous stage in a matrixThe up-sampled filter is carried out, so that the down-sampling operation is avoided; after the image is decomposed by a first-stage non-downsampling pyramid filter (NonSubsampled Pyramid, NSP), a low-frequency subband image and a high-frequency subband image can be obtained, NSP decomposition from the second stage is iteratively performed on the upper-stage low-frequency component, and finally a low-frequency subband image and k high-frequency subband images are obtained, wherein the sizes of the subband images are the same as the original image;
Directionthe localization adopts a standard Shear Filter (SF), is realized by translating a window function in a pseudo-polarization grid system, maps the SF in the pseudo-polarization grid system to a Cartesian coordinate system and carries out inverse Fourier transform, and then the improved shear Filter is obtained by two-dimensional convolution, so that the translation invariance of non-downsampled shear wave transformation is ensured; carrying out L-level direction decomposition on the sub-band image under the k-level scale to obtain 2 L A subband image;
after NSST processing is carried out on the cable partial discharge signal diagram, a low-frequency sub-band diagram (Low frequency directional subband image, LFDSI) and k high-frequency direction sub-band diagrams (High frequency directional subband image, HFDS) are obtained for extracting features of each sub-image by a subsequent method.
3. The method for extracting the cable partial discharge characteristics based on signal mixing enhancement and CNN according to claim 2, wherein the method comprises the following steps: the step S3 is specifically as follows:
s31: gamma corrected low frequency signal enhancement of adaptive parameters
(1) Adaptive gamma
Adopting gamma transformation of self-adaptive parameters to enhance the preliminary noise-reduced image to obtain n preprocessed images; let the input image be in uint8 format, the mathematical form of the gamma transformation of the adaptive parameters be:
γ＝log 2 (m-F median ) (4)
in the formula (3) and the formula (4), I in Brightness for the input image; i out Representing the brightness of the output image; c is a proportionality coefficient, 1 is taken, F represents I in Is a normalized luminance of (2); f (F) median To take I in A median value; gamma is a gamma correction index, the value of which is determined by a target brightness average value m, and m takes the following value:
in the formula (5), the amino acid sequence of the compound,is I in Is the average value of (2); the combination of the formula (3), the formula (4) and the formula (5) can ensure that the brightness of the input n Zhang Chuci noise reduction chart sequence is corrected to be in a uniform range, and the interference of brightness fluctuation on imaging is effectively overcome;
s32: BPDFHE high frequency signal enhancement
Implementing image enhancement on a plurality of high-frequency direction sub-band diagrams in the cable partial discharge signal diagram decomposed by NSST, wherein an algorithm is BPDFHE; the method comprises the following specific steps:
s321: calculating a fuzzy histogram;
s322: partitioning the histogram for convenient classification;
s323: equalizing the dynamic histogram to improve performance;
s324: the brightness of the normalized signal diagram is convenient to calculate;
s325: finally, a high-frequency subband signal diagram after BPDFHE enhancement is obtained;
s323 is divided into partition mapping and equalization operation;
s3231: partition mapping to dynamic range
Δp i ＝h i -l i (6)
In the formulas (6) to (9): l (L) i And h i Respectively isThe lowest and highest intensity values of the ith sub-graph are used to define the interval; input histogram sub-graph and intensity difference Δp for dynamic range of output sub-graph i And adjusting parameter r i Determining, starting with S ai Stop is S ti ；
S3232: equalizing each histogram subgraph
Equalizing each histogram sub-graph operates similarly to global equalization, and the equalized ith sub-graph is shown in formula (10)
Wherein: y (j) is the intensity value of the original signal diagram at the j-th level;
s33: weighting each sub-band map
Based on NSST reconstruction, weighting the sub-band diagram obtained after the enhancement in the previous step in a Cartesian coordinate system and a pseudo-polar coordinate system; NSST decomposition is carried out on the signal diagram to obtain a plurality of low-frequency and high-frequency sub-band diagrams, the enhancement effect applicable to each of gamma correction and BPDFHE can be more fully exerted, and NSST reconstruction is carried out by further combining the algorithm idea of weighted fusion; firstly, respectively using gamma correction and BPDFHE enhancement algorithm to carry out image enhancement on a low-frequency and high-frequency direction sub-band diagram after NSST decomposition; then, carrying out weighted average fusion on the enhanced subgraphs, and finally reconstructing a new enhanced partial discharge signal diagram;
let LFDSI after Gamma enhancement be alpha I 1 Each HFDSI after BPDFHE enhancement is (2-alpha) I h I=1, 2, …,8; wherein, alpha is a weighting coefficient, which is critical in the processing, the value of alpha is 0 to 2, different enhancement reconstruction images are calculated through repeated adjustment of the weighting coefficient and the NSST decomposition reconstruction, and the definition, the gray average value, the edge strength and the contrast of the enhancement signal image under each weighting coefficient are compared, so that the weighting coefficient is finally determined.
4. A method for extracting cable partial discharge characteristics based on signal mixing enhancement and CNN according to claim 3, wherein: the step S4 specifically comprises the following steps:
s41: building convolutional neural network
The deep neural network of the two-dimensional plane data consists of a convolution layer, a pooling layer and a full connection layer;
for a convolution layer, the feature diagram output by the previous layer is subjected to convolution calculation by a convolution kernel, and the feature diagram of the layer is obtained after a function is activated, wherein the specific process is as follows:
wherein: c (C) i Represents a characteristic diagram of the ith layer of CNN,representing convolution calculations, W i Is the weight matrix of the convolution kernel of the ith layer, b i For the bias vector of the i-th layer, σ represents the activation function; during network training, the linear rectification function (Rectified Linear Unit, reLU) reduces the calculated amount by debugging the activity of neurons in the network, improves the expression capacity of the network, has
The calculation process of the pooling layer is expressed as:
C i ＝p(C i-1 ) (13)
wherein p represents a pooling operation;
s42: basic principle of cross-layer feature fusion and CNN optimization
Designing a Cross-layer feature fusion and optimization CNN model (Cross-layer Feature Fusion and Optimization CNN, CFFO-CNN); improving a network structure by introducing cross-layer connection, selecting a plurality of pooled layer output characteristics capable of effectively expressing signal characteristics, and completing fusion of deep and shallow layer characteristics so that the network can autonomously extract essential characteristics of signals from shallow and deep places;
firstly, inputting a preprocessed image into a convolutional neural network to perform convolutional transformation, wherein the network comprises n convolutional layers (conv) and n pooling layers (pool); then, selecting k (k=1, …, n) pooling layer output feature graphs to perform feature fusion, wherein the fused features are expressed as:
F＝[X pool1 X pool2 …,X poolk ] (14)
wherein: x is X pool1 ,X poot2 …,X poolk Outputting characteristic graphs for the 1 st, 2 nd and … th and k th pooling layers; shielding part of neurons of the full-connection layer by adopting a Dropout method;
finally, inputting the fused feature images into a full-connection layer for classification;
s43: gamma-BPDFHE enhancement and CFFO-CNN based feature extraction
S431: outputting the Gamma-BPDFHE enhanced time-frequency spectrogram of the partial discharge signal into an image format, carrying out gray level processing, and compressing the image input size with the same proportion to 64 multiplied by 64:
s432: constructing a basic structure of CNN, wherein the network comprises 2 convolution layers and corresponding pooling layers, the convolution kernel size of the network is set to be 5 multiplied by 5, the pooling kernel size is 2 multiplied by 2, the batch size is 8, the iteration number is 200, and an RLU activation function, a maximum pooling mode and a random gradient descent (Stochastic Gradient Descent with Momentum, SGDM) algorithm based on momentum are adopted;
s433: introducing cross-layer connection, and determining a feature fusion layer with highest recognition accuracy by comparing recognition effects of different feature fusion layer structure networks;
s434: and optimizing the super parameters of the cross-layer feature fusion convolutional neural network by using an Adam algorithm to obtain CFFO-CNN, and extracting the intrinsic features of the spectrum image when the local discharge enhanced signal is extracted by using the CFFO-CNN.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410011871.7A CN117828333A (en) | 2024-01-04 | 2024-01-04 | Cable partial discharge feature extraction method based on signal mixing enhancement and CNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410011871.7A CN117828333A (en) | 2024-01-04 | 2024-01-04 | Cable partial discharge feature extraction method based on signal mixing enhancement and CNN |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117828333A true CN117828333A (en) | 2024-04-05 |
Family
ID=90507580
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410011871.7A Pending CN117828333A (en) | 2024-01-04 | 2024-01-04 | Cable partial discharge feature extraction method based on signal mixing enhancement and CNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117828333A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118130984A (en) * | 2024-05-10 | 2024-06-04 | 山东博通节能科技有限公司 | Cable partial discharge fault real-time monitoring method based on data driving |
-
2024
- 2024-01-04 CN CN202410011871.7A patent/CN117828333A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118130984A (en) * | 2024-05-10 | 2024-06-04 | 山东博通节能科技有限公司 | Cable partial discharge fault real-time monitoring method based on data driving |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109841226B (en) | Single-channel real-time noise reduction method based on convolution recurrent neural network | |
| Xu et al. | Image inpainting by patch propagation using patch sparsity | |
| CN109035142B (en) | Satellite image super-resolution method combining countermeasure network with aerial image prior | |
| CN109389585B (en) | Brain tissue extraction method based on full convolution neural network | |
| CN112233026A (en) | SAR image denoising method based on multi-scale residual attention network | |
| CN117828333A (en) | Cable partial discharge feature extraction method based on signal mixing enhancement and CNN | |
| CN110490816B (en) | Underwater heterogeneous information data noise reduction method | |
| CN111583285A (en) | Liver image semantic segmentation method based on edge attention strategy | |
| CN113344933A (en) | Glandular cell segmentation method based on multi-level feature fusion network | |
| CN112785539A (en) | Multi-focus image fusion method based on image adaptive decomposition and parameter adaptive | |
| CN116246138A (en) | Infrared-visible light image target level fusion method based on full convolution neural network | |
| CN117593187A (en) | Remote sensing image super-resolution reconstruction method based on meta-learning and transducer | |
| CN116681623A (en) | SAR image target detection method based on multistage Laplacian pyramid denoising | |
| CN117079005A (en) | Optical cable fault monitoring method, system, device and readable storage medium | |
| CN116630964A (en) | Food image segmentation method based on discrete wavelet attention network | |
| CN116309221A (en) | Method for constructing multispectral image fusion model | |
| CN111047537A (en) | System for recovering details in image denoising | |
| CN117196963A (en) | Point cloud denoising method based on noise reduction self-encoder | |
| CN115565671A (en) | Atrial fibrillation auxiliary analysis method based on cross-model mutual teaching semi-supervision | |
| CN114049357B (en) | Mammary gland ultrasonic segmentation method based on feature set association degree | |
| CN115937302A (en) | Hyperspectral image sub-pixel positioning method combined with edge preservation | |
| CN115205308A (en) | Fundus image blood vessel segmentation method based on linear filtering and deep learning | |
| Annavarapu et al. | An adaptive watershed segmentation based medical image denoising using deep convolutional neural networks | |
| CN114331894A (en) | Face image restoration method based on potential feature reconstruction and mask perception | |
| CN113205471A (en) | Log-Gabor transformation and direction region entropy-based guided filtering remote sensing image fusion method |
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 |