CN115620107A - Transformer substation bird-involved fault related bird species identification method based on deep learning - Google Patents
Transformer substation bird-involved fault related bird species identification method based on deep learning Download PDFInfo
- Publication number
- CN115620107A CN115620107A CN202211386862.3A CN202211386862A CN115620107A CN 115620107 A CN115620107 A CN 115620107A CN 202211386862 A CN202211386862 A CN 202211386862A CN 115620107 A CN115620107 A CN 115620107A
- Authority
- CN
- China
- Prior art keywords
- bird
- module
- convnext
- bird species
- image
- 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
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 238000001514 detection method Methods 0.000 claims abstract description 77
- 238000012549 training Methods 0.000 claims abstract description 36
- 238000012360 testing method Methods 0.000 claims abstract description 21
- 238000012795 verification Methods 0.000 claims abstract description 9
- 238000000605 extraction Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 38
- 238000010586 diagram Methods 0.000 claims description 37
- 230000004927 fusion Effects 0.000 claims description 18
- 230000003044 adaptive effect Effects 0.000 claims description 12
- 238000005070 sampling Methods 0.000 claims description 9
- 238000002372 labelling Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 4
- 239000000654 additive Substances 0.000 claims description 3
- 230000000996 additive effect Effects 0.000 claims description 3
- 230000002708 enhancing effect Effects 0.000 claims 1
- 230000002265 prevention Effects 0.000 abstract description 4
- 241000894007 species Species 0.000 description 82
- 238000012935 Averaging Methods 0.000 description 5
- 230000004913 activation Effects 0.000 description 5
- 238000010606 normalization Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 238000009499 grossing Methods 0.000 description 3
- 238000007689 inspection Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 230000001629 suppression Effects 0.000 description 3
- 241000271566 Aves Species 0.000 description 2
- 241001137251 Corvidae Species 0.000 description 2
- 206010037180 Psychiatric symptoms Diseases 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000013508 migration Methods 0.000 description 2
- 230000005012 migration Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 240000001689 Cyanthillium cinereum Species 0.000 description 1
- 241001464763 Dichagyris flammatra Species 0.000 description 1
- 244000148064 Enicostema verticillatum Species 0.000 description 1
- 241000245165 Rhododendron ponticum Species 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 235000005770 birds nest Nutrition 0.000 description 1
- 244000240635 birds nest Species 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 235000005765 wild carrot Nutrition 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method for identifying bird species related to bird-involved faults of a transformer substation based on deep learning, which comprises the steps of firstly, aiming at a transformer substation polling image and a public bird species image set, expanding bird species images by adopting motion blur and defocusing blur, and constructing a transformer substation bird-involved fault related bird species image data set; then, making a label for each image in the data set to obtain a real rectangular frame containing bird species categories and coordinate information, and dividing the frame into a test set, a training set and a verification set; secondly, improving a YOLOv5 target detection model structure, replacing an original main feature extraction network with ConvNeXt-T, and extracting richer bird image features; and finally, training the improved YOLOv5 model by combining a plurality of training skills, and detecting the divided test set images by using the trained model. The invention can provide reference for differential prevention and control of bird-involved faults of the transformer substation.
Description
Technical Field
The invention belongs to the technical field of transformer substation monitoring image data processing, and particularly relates to a transformer substation bird-related fault related bird species identification method based on deep learning.
Background
The transformer substation is a core place for bearing electric energy conversion and distribution in a power grid, and safe and stable operation of the transformer substation has great significance to the power grid. With the continuous development of power grids, power transmission lines, transformer substations and the like are widely spread, and part of the transformer substations are in remote areas with relatively good ecological environments, so that a proper production environment is brought to birds. However, birds nest, fly, defecate and other activities in the open-type transformer substation, easily cause insulation flashover, disconnecting link switching failure, switch tripping, even equipment burning loss and other faults, and seriously threaten the safe and stable operation of the transformer equipment. However, due to the characteristics of randomness and burstiness, the problems of manpower and material resource consumption, low efficiency and the like are caused because the system cannot be found in time depending on manual routing inspection. In recent years, equipment such as a robot and an unmanned aerial vehicle are used for carrying an intelligent detection model to patrol a transformer substation, bird species images related to bird-related faults of the transformer substation are collected, and intelligent detection is achieved. According to the bird species category that detects out, carry out differentiation prevention and cure, very big improvement patrols and examines efficiency.
At present, with the further development of machine learning, the deep learning intelligent detection algorithm is applied to various fields. Compared with the traditional image recognition method, the method has higher detection precision and speed. The convolutional neural network can automatically extract the image features, so that the complex process of manually extracting the image features is avoided, and the efficiency is greatly improved. However, intelligent identification of bird species related to bird-related faults in a transformer substation is less, most of bird species are concentrated in the field of bird ecology, and an intelligent bird identification method is urgently needed to assist inspection personnel.
Disclosure of Invention
In view of this, the invention aims to provide a method for identifying bird species related to bird-involved faults of a transformer substation based on deep learning, which can realize accurate intelligent identification of bird species related to bird-involved faults of the transformer substation.
In order to achieve the purpose, the invention provides the following technical scheme: a transformer substation bird-related fault related bird species identification method based on deep learning comprises the following steps:
s1, constructing a bird species image data set related to bird-related faults of a transformer substation;
s2, constructing an improved Yolov5 bird species target detection model: constructing four parts, namely an input end, a backbone section network ConvNeXt-T for generating characteristics, a neck section network FPN + PAN for carrying out characteristic fusion and a head section network for carrying out prediction; updating parameters through back propagation of a loss function, wherein the loss function consists of detection frame loss, confidence coefficient loss and classification loss, and finally obtaining an improved Yolov5 bird target detection model;
s3, training an improved Yolov5 bird species target detection model;
and S4, detecting the images of the test set by using the trained improved Yolov5 bird species target detection model.
Preferably, in step S1, the substation patrol image and the bird species image set are expanded by using motion blur and defocus blur, a substation bird-related fault-related bird species image data set including multiple bird species is constructed, a tool Labeling is applied to label each image in the substation bird-related fault-related bird species image data set, the obtained bird species and coordinate information are prepared into a standard Pascal VOC data set format, and the standard Pascal VOC data set format is divided into a test set, a training set and a verification set.
More preferably, in step S1, the bird species image is motion blurred by a motion blur function (PSF):
in the formulaf(x, y) Andg(x, y) Respectively representing the input bird species image and the motion-blurred bird species image,h(x, y) For a motion blur function (PSF) that convolves the input bird species image,n(x, y) For an additive noise function, the formula for the motion blur function is expressed as:
in the formulaLAnd
is the length and direction of motion blur; defocus blur is obtained by substituting the motion blur function with the defocus functionh r (x, y) The result of the convolution:
in the formularIs the defocus blur radius.
Further preferably, the input end performs mosaic data enhancement, adaptive anchor frame calculation and adaptive picture scaling technical processing on the input image to obtain an appropriate anchor frame and an input image with a uniform size.
It is further preferred that the backbone segment network of the original YOLOv5 is replaced by ConvNeXt-T, which in turn consists of 1 separate 4 × 4 convolutional layer, a normative layer (LayerNorm), a ConvNeXt _1 module, a ConvNeXt _2 module, a ConvNeXt _3 module, a downsampling module i, a ConvNeXt _4 module, a ConvNeXt _5 module, a ConvNeXt _6 module, a downsampling module ii, a ConvNeXt _7 module, a ConvNeXt _8 module, a ConvNeXt _9 module, a ConvNeXt _10 module, a ConvNeXt _11 module, a ConvNeXt _12 module, a ConvNeXt _13 module, a ConvNeXt _14 module, a ConvNeXt _15 module, a downsampling module, a ConvNeXt _16 module, a ConvNeXt _17, a normative _ iii module, a linvre _16 module, a global averaging layer (layernom _ 18), and a global averaging layer.
Further preferably, the first features F are extracted through a ConvNeXt _3 module, a ConvNeXt _6 module, a ConvNeXt _15 module and a ConvNeXt _18 module in the backbone segment network ConvNeXt-T respectively 1 A second feature F 2 The third feature F 3 And fourth feature F 4 And the convolution characteristic extraction of the image from shallow to deep is realized, and then the image is input to a neck segment network FPN + PAN for fusion.
Further preferably, the first feature F extracted from the neck segment network FPN + PAN pair 1 Second characteristic F 2 A third feature F 3 Fourth feature F 4 Carrying out characteristic fusion from deep to shallow and from shallow to deep, specifically: fourth feature F 4 Obtaining a first feature map P through a CSPN2 module and a CBL convolution block 1 First characteristic diagram P 1 After upsampling and the second feature F 2 Performing fusion to obtain a second feature map P 2 (ii) a Second characteristic diagram P 2 Obtaining a third characteristic diagram P through a CSPN2 module and a CBL convolution block 3 Third characteristic diagram P 3 After up-sampling and the first characteristic F 1 Performing fusion and obtaining a fourth characteristic diagram P through a CSPN2 module 4 (ii) a Fourth feature map P 4 Obtaining a fifth feature map P through the CBL convolution block 5 Fifth characteristic diagram P 5 And the third characteristic diagram P 3 Merging and obtaining a sixth feature map P through a CSPN2 module 6 (ii) a Sixth characteristic diagramP 6 Obtaining a seventh feature map P by CBL convolution block 7 Seventh feature diagram P 7 And the third feature F 3 Performing fusion and obtaining an eighth characteristic diagram P through a CSPN2 module 8 。
More preferably, convNeXt _iThe module is used for carrying out the following steps,i 
1, 2, 3, \ 823018, which consists of 17 × 7 layer-by-layer convolutional layer, a normative layer, 1 convolutional layer, 1 GELU activation function, 1 convolutional layer, a scaling and regularization layer in sequence, and adds input and output by a residual connection.
Further preferably, the CSPN2 module has 2 branches, 1 branch is 2 CBL convolution blocks and 1 convolution layer, the other 1 branch is 1 convolution layer, and the output of the two branches is subjected to feature fusion and then sequentially passes through the batch normalization layer, the leak RuLU activation function and the CBL convolution block to obtain the final output.
Preferably, three detection frames with different prediction sizes exist in the head segment network, and the detection frames respectively correspond to large, medium and small detection targets; the obtained fourth feature map P with strengthened fusion 4 The sixth characteristic diagram P 6 And an eighth feature map P 8 After convolution, the data are input into detection boxes with large, medium and small prediction sizes.
Loss function loss by detection box in head segment networkLossRect, loss of confidenceLossConf and classification lossLossA cls component, updating the weight and the deviation by a loss function through back propagation, and finally obtaining an optimal detection frame, a most accurate confidence score and a classification result at the same time; the detection frame Loss adopts a CIOU _ Loss function, confidence coefficient Loss and classification Loss as cross entropy Loss functions.
Firstly, pre-training an improved Yolov5 bird species target detection model by using a COCO public data set to obtain pre-training weights; and then loading pre-training weights, and retraining the improved Yolov5 bird species target detection model by using the divided training set and verification set, wherein the training skills comprise: the method comprises the following steps of multi-stage transfer learning, mixed precision training fp16, mosaic and mixed data enhancement, label smoothing, SGDM model parameter optimization and Step length (Step) learning rate reduction, wherein the training skills can improve the speed, generalization capability and robustness of model training; secondly, evaluating the trained improved YOLOv5 bird species target detection model through a verification set during training, and optimizing parameters of the improved YOLOv5 bird species target detection model until the training is finished; finally, obtaining the optimal weight of the improved Yolov5 bird target detection model;
further preferably, a multi-stage migration learning method is adopted to train the improved Yolov5 bird species target detection model, a first stage freezes a main stem to train, and a second stage finely tunes all layer parameters; and (4) optimizing parameters by adopting an SGDM model.
Preferably, in step S4, the optimal weight is loaded, the test set image is scaled to a uniform size by using an adaptive picture scaling technique, and then input to the improved YOLOv5 bird species target detection model for detection, and a flexible non-maximum value is used for suppressing and selecting a positioning accurate frame, so as to realize detection of the bird species target, wherein the detected image includes the species and specific position information of the bird species and the confidence of the prediction result, and the number of the bird species is statistically output.
Further preferably, in step S4, the adaptive image scaling technique (lettbex) scales the test set image to 640 × 640, inputs the scaled test set image into the improved yollov 5 bird species target detection model for detection, and selects a positioning accurate frame through flexible non-maximum suppression: prediction candidate boxY r The higher the coincidence degree with the highest-score prediction candidate frame M, the higher the prediction candidate frame MY r The lower the score of (d); prediction candidate boxY r The degree of coincidence with the highest-score prediction candidate frame M is less than the threshold valueN t The prediction candidate block is retainedY r 。
Compared with the prior art, the invention has the beneficial effects that: the bird species image is expanded through motion blur and defocusing blur, the improved YOLOv5 bird species target detection model is trained on the basis of the improved YOLOv5 bird species target detection model and in combination with various training skills, the trained improved YOLOv5 bird species target detection model is used for detecting the divided test set image, and accurate intelligent identification of bird species related to bird-related faults of the transformer substation can be achieved. The technical scheme provided by the invention has better detection precision and detection speed, is more favorable for practical application, and can be used for carrying out differential prevention and control on bird-involved faults of the transformer substation for transformer substation operation and maintenance personnel.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
FIG. 2 is a schematic diagram of an improved Yolov5 bird species target detection model.
Fig. 3 is a schematic diagram of the ConvNeXt _1 module.
Fig. 4 is a schematic diagram of a CSPN2 module.
Detailed Description
The present invention is further described in the following examples, which should not be construed as limiting the scope of the invention, but rather as providing the following examples which are set forth to illustrate and not limit the scope of the invention.
A method for identifying bird species related to bird-involved faults in a transformer substation based on deep learning is disclosed, wherein a flow chart of the method is shown in figure 1, and comprises the following steps:
s1, constructing a bird species image data set related to bird-involved faults of a transformer substation: aiming at a transformer substation inspection image and a bird species image set, extending bird species images by adopting motion blur and defocusing blur, constructing a transformer substation bird-related fault related bird species image data set containing 10 bird species, labeling each image in the transformer substation bird-related fault related bird species image data set by using a tool Labeling, and making the obtained bird species and coordinate information into a standard Pascal VOC data set format and dividing the standard bird species and coordinate information into a test set, a training set and a verification set;
in the embodiment, 10 birds such as pied magpie, ash pied magpie, rhododendron, vernonia nucifera, white head 40526, swallow, red falcon and black collar 26891, silkete 26891, bird and big mouth crow are selected as research objects from the bird species image, each bird species comprises 200 bird species images, and 2000 bird species image samples are counted. The bird species image is augmented with motion blur and defocus blur: motion blurring the bird species image by a motion blur function:
in the formulaf(x, y) Andg(x, y) Respectively representing the input bird species image and the motion-blurred bird species image,h(x, y) For the motion blur function of convolving the input bird species image,n(x, y) For an additive noise function, the formula for the motion blur function is expressed as:
in the formulaLAnd
is the length and direction of motion blur; defocus blur is obtained by substituting the motion blur function with the defocus functionh r (x, y) The result of the convolution:
in the formularIs the defocus blur radius. The existing bird species image is expanded through motion blur and defocusing blur, a real environment is simulated, the diversity of the bird species image is enhanced, and overfitting is prevented. And expanding 100 bird species images for each bird, adding the previously collected bird species images, and constructing a bird species image data set related to bird-related faults in the transformer substation, wherein each bird comprises 300 bird species images, and 3000 bird species image samples are calculated. Labeling each image in the data set by using a tool Labeling, wherein the obtained bird species and coordinate information are prepared into a standard Pascal VOC data set format according to the following steps of 9:1 into training and test sets, 10% of the test sets beingAnd the number of the pictures in the verification set is 2430, 300 and 270 respectively.
S2, constructing an improved Yolov5 bird species target detection model: constructing four parts of an input end, a backbone segment network ConvNeXt-T for generating characteristics, a neck segment network FPN + PAN for carrying out characteristic fusion and a head segment network for carrying out prediction; updating parameters through back propagation of a loss function, wherein the loss function consists of detection frame loss, confidence coefficient loss and classification loss, and finally obtaining an improved Yolov5 bird target detection model; as shown in fig. 2, the improved YOLOv5 bird species target detection model constructed in this embodiment is composed of four parts, a first part is an input end, a second part is a backbone segment network ConvNeXt-T, a third part is a neck segment network FPN + PAN, and a fourth part is a head segment network:
the input end carries out mosaic data enhancement, adaptive Anchor frame (Anchor) calculation and adaptive picture scaling technology (Letterbox) processing on an input image to obtain an appropriate Anchor frame (Anchor) and an input image with uniform size, wherein the input size is 640 multiplied by 3, the mosaic data enhancement can enrich a data set and greatly improve the training speed of a network at the same time, the memory requirement of a model is reduced, the adaptive Anchor frame (Anchor) calculation can avoid artificially defining Anchor frames with different sizes and aspect ratios, the quality of the Anchor frame (Anchor) is improved, the adaptive picture scaling technology (Letterbox) can reduce a large amount of information redundancy, and the reasoning speed of an algorithm is improved;
replacing the backbone segment network of the original YOLOv5 with ConvNeXt-T, the backbone segment network ConvNeXt-T being composed of 1 separate 4 × 4 convolutional layer, a normative layer (LayerNorm), a ConvNeXt _1 module, a ConvNeXt _2 module, a ConvNeXt _3 module, a downsampling module i, a ConvNeXt _4 module, a ConvNeXt _5 module, a ConvNeXt _6 module, a downsampling module ii, a ConvNeXt _7 module, a ConvNeXt _8 module, a ConvNeXt _9 module, a ConvNeXt _10 module, a ConvNeXt _11 module, a convnexxt _12 module, a ConvNeXt _13 module, a convnextt _14 module, a convnextt _15 module, a downsampling module, a convnextt _16 module, a ConvNeXt _17 module, a normative layer (layerjiii) module, a global averaging layer (layerjt) and a global averaging layer (global averaging layer) in sequenceTo avoid pictures that are too large to be ambiguous, the global average pooling layer, the normalization layer (LayerNorm), and the linearization layer (Linear) are not plotted in fig. 2). As shown in fig. 3, the ConvNeXt _1 module is composed of 17 × 7 Layer-by-Layer convolutional Layer, a normative Layer (LayerNorm), 1 × 1 convolutional Layer, 1 GELU activation function, 1 × 1 convolutional Layer, scaling (Layer Scale), and a regularization Layer (Drop Path) in sequence, and adds the input and output by a residual connection; the down-sampling module (a down-sampling module I, a down-sampling module II and a down-sampling module III) consists of a standard layer (LayerNorm) and 1 convolution layer of 2 multiplied by 2. ConvNeXt _2, convNeXt _3, \8230, convNeXt _18, with the same structure as ConvNeXt _ 1; extracting first characteristics F through a ConvNeXt _3 module, a ConvNeXt _6 module, a ConvNeXt _15 module and a ConvNeXt _18 module in the backbone segment network ConvNeXt-T respectively 1 And a second feature F of 80X 192 2 And a third feature F of 40X 384 3 And a fourth feature F of 20X 768 4 The method comprises the steps of extracting the convolution characteristics of an image from shallow to deep, inputting the image into a neck segment network FPN + PAN for fusion, wherein the selected lightweight skeleton segment network ConvNeXt-T also reduces the parameter quantity of a model while keeping the characteristic quality, and only the convolution characteristic extraction is realized but not the classification function is realized;
first feature F extracted from neck segment network FPN + PAN pair 1 Second characteristic F 2 A third feature F 3 Fourth feature F 4 Carrying out characteristic fusion from deep to shallow and from shallow to deep, specifically: fourth feature F 4 Obtaining a first characteristic diagram P through a CSPN2 module and a CBL convolution block 1 First characteristic diagram P 1 After up-sampling and the second characteristic F 2 Fusing to obtain a second characteristic diagram P 2 (ii) a Second characteristic diagram P 2 Obtaining a third characteristic diagram P through a CSPN2 module and a CBL convolution block 3 Third characteristic diagram P 3 After up-sampling and the first characteristic F 1 Merging and obtaining a fourth characteristic diagram P through a CSPN2 module 4 (ii) a Fourth feature map P 4 Obtaining a fifth feature map P through the CBL convolution block 5 Fifth characteristic diagram P 5 And the third characteristic diagram P 3 Go on to meltCombining the six characteristic maps P obtained by the CSPN2 module 6 (ii) a Sixth feature map P 6 Obtaining a seventh feature map P by CBL convolution block 7 Seventh feature diagram P 7 And the third feature F 3 Merging and obtaining an eighth feature map P through a CSPN2 module 8 (ii) a As shown in fig. 4, the CSPN2 module has 2 branches, 1 branch is 2 CBL convolution blocks, 1 convolution layer, and the other 1 branch is 1 convolution layer, and the output of the two branches is subjected to feature fusion and then sequentially passes through a batch normalization layer (BN), a leak RuLU activation function and a CBL convolution block to obtain a final output, wherein the CBL convolution block is composed of a normal convolution, a batch normalization layer (BN) and a leak RuLU activation function, and a neck segment network FPN + PAN shortens an information path between features of a lower layer and a top layer, so that information of the lower layer is more easily transmitted;
three detection frames with different prediction sizes of 20 multiplied by 45, 40 multiplied by 45 and 80 multiplied by 45 exist in the head segment network, and the detection frames respectively correspond to a large detection target, a medium detection target and a small detection target; the obtained fourth feature map P with strengthened fusion 4 The sixth characteristic diagram P 6 And an eighth feature map P 8 After convolution, inputting the data into detection frames with large, medium and small prediction sizes; loss function in head segment network is lost by detection box (LossRect), loss of confidence (LossConf) and classification loss (LossCls), updating the weight and deviation by a loss function through back propagation, and finally obtaining an optimal detection frame, a most accurate confidence score and a classification result at the same time; the detection frame Loss adopts a CIOU _ Loss function, confidence coefficient Loss and classification Loss as cross entropy Loss functions, and the improved YOLOv5 bird target detection model Loss is shown as a formula (4):
s3, training an improved Yolov5 bird species target detection model: firstly, pre-training an improved Yolov5 bird species target detection model by using a COCO public data set to obtain pre-training weights; then loading pre-training weights, and retraining the improved Yolov5 bird species target detection model by using the divided training set and verification set, wherein the training skills comprise: the method comprises the following steps of multi-stage transfer learning, mixed precision training fp16, mosaic and mixed data enhancement, label smoothing, SGDM model parameter optimization and Step length (Step) learning rate reduction, wherein the training skills can improve the speed, generalization capability and robustness of model training; secondly, evaluating the trained improved YOLOv5 bird species target detection model through a verification set during training, and optimizing parameters of the improved YOLOv5 bird species target detection model until the training is finished; finally, obtaining the optimal weight of the improved Yolov5 bird target detection model;
in the embodiment, a multistage migration learning is adopted to train an improved yollov 5 bird species target detection model, a first stage freezes a trunk to train, the number of training rounds is 50 rounds, the batch size (batch size) is 8, a second stage finely tunes all layer parameters, and the number of training rounds is 50 rounds, and the batch size (batch size) is 4; the mosaic data enhancement is set to be used at a 50% probability in each iteration, the mixed data enhancement is set to be used after the mosaic data enhancement at a 50% probability, and the value of the label smoothing is set to be 0.01; adopting SGDM model parameter optimization, setting momentum value to be 0.9, and using Step length (Step) learning rate reduction strategy, the maximum learning rate is 1 multiplied by 10 -2 The minimum learning rate is 1 × 10 -4 (ii) a The training model was evaluated through a validation set during training, set to evaluate once every 10 rounds, optimizing the model parameters.
S4, detecting the images of the test set by using the trained improved Yolov5 bird species target detection model: the method comprises the steps of loading optimal weight, utilizing a self-adaptive picture scaling technology (Letterbox) to scale a test set image to a uniform size, inputting the test set image to an improved YOLOv5 bird species target detection model for detection, selecting a positioning accurate frame through flexible non-maximum suppression (Soft-NMS), realizing the detection of a bird species target, wherein the detected image comprises bird species, specific position information and confidence coefficient of a prediction result, and carrying out statistical output on bird species quantity;
in this embodiment, the adaptive image scaling technique (lettbex) is used to scale the test set image to 640 × 640, and then the scaled test set image is input into the improved YOLOv5 bird species target detection modelDetecting, and selecting a positioning accurate frame through flexible non-maximum suppression (Soft-NMS): prediction candidate boxY r The higher the coincidence degree with the highest-score prediction candidate frame M is, the prediction candidate frame isY r The lower the score of (a); prediction candidate boxY r The coincidence degree with the highest-score prediction candidate frame M is less than the threshold valueN t Then the prediction candidate box is retainedY r
In the formulaS i Is a set of prediction candidate box scores,Mthe candidate box is predicted for the highest score,Y r in order to predict the candidate block(s),N t for threshold, the threshold is set to 0.3 and IOU is the cross-over ratio.
In this embodiment, the trained improved YOLOv5 bird species target detection model is used to detect the test set images, so as to realize the bird classification identification related to the bird-related fault of the transformer substation. The accuracy and detection speed of the model are evaluated by mean average precision (mAP) and Frame Per Second (FPS) to evaluate the classification performance. The result shows that the average precision mean value of the detection of 300 test sets containing 10 types of bird species can reach 97.47 percent, the FPS is 23, and the detection speed is related to the experimental environment. The effectiveness of the method for identifying the bird species related to the bird-involved fault of the transformer substation based on deep learning is verified. The bird-involved fault differential prevention and control improvement reference can be developed for transformer substation operation and maintenance personnel.
The foregoing description is of the preferred embodiment of the present invention only, and is not intended to limit the invention in any way, so that those skilled in the art, having the benefit of this disclosure, may modify and/or adapt the same to equivalent embodiments without departing from the scope of the present invention. However, any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.
Claims (10)
1. A method for identifying bird species related to bird-related faults of a transformer substation based on deep learning is characterized by comprising the following steps:
s1, constructing a bird species image data set related to bird-involved faults of a transformer substation;
s2, constructing an improved Yolov5 bird species target detection model: constructing four parts, namely an input end, a backbone section network ConvNeXt-T for generating characteristics, a neck section network FPN + PAN for carrying out characteristic fusion and a head section network for carrying out prediction; updating parameters through back propagation of a loss function, wherein the loss function consists of detection frame loss, confidence coefficient loss and classification loss, and finally obtaining an improved Yolov5 bird target detection model;
s3, training an improved Yolov5 bird species target detection model;
and S4, detecting the test set image by using the trained improved YOLOv5 bird species target detection model.
2. The method for identifying the bird species related to the bird-related fault of the transformer substation based on the deep learning of claim 1, wherein in the step S1, for a transformer substation patrol image and a bird species image set, the bird species image is expanded by adopting motion blur and defocus blur, a transformer substation bird-related fault related bird species image data set containing a plurality of bird species is constructed, a tool Labeling is used for Labeling each image in the transformer substation bird-related fault related bird species image data set, and the obtained bird species and coordinate information are manufactured into a standard Pascal VOC data set format and are divided into a test set, a training set and a verification set.
3. The method for identifying the bird species related to the bird-involved fault in the transformer substation based on the deep learning as claimed in claim 2, wherein in the step S1, the bird species image is subjected to motion blurring through a motion blurring function:
in the formulaf(x, y) Andg(x, y) Respectively representing the input bird species image and the motion-blurred bird species image,h(x, y) For the motion blur function of convolving the input bird species image,n(x, y) For an additive noise function, the formula for the motion blur function is expressed as:
in the formulaLAnd
is the length and direction of motion blur; defocus blur is obtained by substituting the motion blur function with the defocus functionh r (x, y) The result of the convolution:
in the formularIs the defocus blur radius.
4. The deep learning-based identification method for bird-involved fault-related bird species of transformer substation according to claim 1, wherein the input end performs mosaic data enhancement, adaptive anchor frame calculation and adaptive picture scaling on the input image to obtain an appropriate anchor frame and input image with uniform size.
5. The deep learning based substation bird-related fault related bird species identification method of claim 1, wherein the backbone segment network of the original YOLOv5 is replaced by ConvNeXt-T, which in turn consists of 1 individual 4 x 4 convolutional layer, a normative layer, a ConvNeXt _1 module, a ConvNeXt _2 module, a ConvNeXt _3 module, a downsampling module i, a ConvNeXt _4 module, a ConvNeXt _5 module, a ConvNeXt _6 module, a downsampling module ii, a ConvNeXt _7 module, a ConvNeXt _8 module, a ConvNeXt _9 module, a ConvNeXt _10 module, a convnexxt _11 module, a convnexxt _12 module, a convnexxt _13 module, a convnexxt _14 module, a vvnext _15 module, a ConvNeXt _ iii, a ConvNeXt _16 module, a normative layer, a global average layer, a ConvNeXt _17 module, and a global average layer.
6. The deep learning-based substation bird-related fault related bird species identification method according to claim 5, wherein first features F are extracted through a ConvNeXt _3 module, a ConvNeXt _6 module, a ConvNeXt _15 module and a ConvNeXt _18 module in a backbone segment network ConvNeXt-T respectively 1 Second characteristic F 2 A third feature F 3 And a fourth feature F 4 And the convolution characteristic extraction of the image from shallow to deep is realized, and then the image is input to a neck segment network FPN + PAN for fusion.
7. The deep learning-based bird species identification method related to bird-involved fault in transformer substation according to claim 6, characterized in that the first feature F extracted from the neck segment network FPN + PAN pair 1 A second feature F 2 The third feature F 3 Fourth feature F 4 Carrying out characteristic fusion from deep to shallow and from shallow to deep, specifically: fourth feature F 4 Obtaining a first characteristic diagram P through a CSPN2 module and a CBL convolution block 1 First characteristic diagram P 1 After up-sampling and the second characteristic F 2 Fusing to obtain a second characteristic diagram P 2 (ii) a Second characteristic diagram P 2 Obtaining a third characteristic diagram P through a CSPN2 module and a CBL convolution block 3 Third characteristic diagram P 3 After up-sampling and the first characteristic F 1 Merging and obtaining a fourth characteristic diagram P through a CSPN2 module 4 (ii) a Fourth characteristic diagram P 4 Obtaining a fifth feature map P through the CBL convolution block 5 Fifth characteristic diagram P 5 And the third characteristic diagram P 3 Performing fusion and obtaining a sixth characteristic diagram P through a CSPN2 module 6 (ii) a Sixth feature map P 6 Obtaining a seventh feature map P by CBL convolution block 7 Seventh feature diagram P 7 And the third feature F 3 Merging and obtaining an eighth feature map P through a CSPN2 module 8 。
8. The deep learning-based method for identifying bird related faults of transformer substations according to claim 7, wherein three detection frames with different prediction sizes are present in the head section network, and are respectively corresponding to prediction of large, medium and small detection targets; the obtained fourth feature map P for enhancing fusion 4 The sixth characteristic diagram P 6 And an eighth feature map P 8 After convolution, the data are input into detection boxes with large, medium and small prediction sizes.
9. The deep learning-based identification method for bird-involved fault-related bird species of transformer substation according to claim 8, wherein in the head segment network, the loss function updates the weight and the deviation through back propagation, and finally, the optimal detection frame and the most accurate confidence score and classification result are obtained simultaneously; the detection frame Loss adopts a CIOU _ Loss function, confidence coefficient Loss and classification Loss as cross entropy Loss functions.
10. The method for identifying bird species related to bird-involved fault in the transformer substation based on deep learning of claim 9, wherein in step S4, optimal weights are loaded, the image of the test set is scaled to a uniform size by using an adaptive picture scaling technique and then input to an improved YOLOv5 bird species target detection model for detection, and an accurate positioning frame is selected by suppressing a flexible non-maximum value to realize detection of bird species targets, the detected image includes bird species, specific position information and confidence of prediction results, and statistical output is performed on the number of birds.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211386862.3A CN115620107A (en) | 2022-11-07 | 2022-11-07 | Transformer substation bird-involved fault related bird species identification method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211386862.3A CN115620107A (en) | 2022-11-07 | 2022-11-07 | Transformer substation bird-involved fault related bird species identification method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115620107A true CN115620107A (en) | 2023-01-17 |
Family
ID=84879155
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211386862.3A Pending CN115620107A (en) | 2022-11-07 | 2022-11-07 | Transformer substation bird-involved fault related bird species identification method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115620107A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116229380A (en) * | 2023-05-10 | 2023-06-06 | 国网江西省电力有限公司电力科学研究院 | Method for identifying bird species related to bird-related faults of transformer substation |
| CN116912890A (en) * | 2023-09-14 | 2023-10-20 | 国网江苏省电力有限公司常州供电分公司 | Method and device for detecting birds in transformer substation |
| CN117095411A (en) * | 2023-10-16 | 2023-11-21 | 青岛文达通科技股份有限公司 | Detection method and system based on image fault recognition |
| CN117874173A (en) * | 2024-03-11 | 2024-04-12 | 腾讯科技（深圳）有限公司 | Training method and related device of vector model |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111104906A (en) * | 2019-12-19 | 2020-05-05 | 南京工程学院 | Transmission tower bird nest fault detection method based on YOLO |
| CN112215822A (en) * | 2020-10-13 | 2021-01-12 | 北京中电兴发科技有限公司 | Face image quality evaluation method based on lightweight regression network |
| CN113486865A (en) * | 2021-09-03 | 2021-10-08 | 国网江西省电力有限公司电力科学研究院 | Power transmission line suspended foreign object target detection method based on deep learning |
| CN114842208A (en) * | 2022-06-08 | 2022-08-02 | 南昌大学 | Power grid harmful bird species target detection method based on deep learning |
| CN115082401A (en) * | 2022-06-22 | 2022-09-20 | 桂林电子科技大学 | SMT production line chip mounter fault prediction method based on improved YOLOX and PNN |
-
2022
- 2022-11-07 CN CN202211386862.3A patent/CN115620107A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111104906A (en) * | 2019-12-19 | 2020-05-05 | 南京工程学院 | Transmission tower bird nest fault detection method based on YOLO |
| CN112215822A (en) * | 2020-10-13 | 2021-01-12 | 北京中电兴发科技有限公司 | Face image quality evaluation method based on lightweight regression network |
| CN113486865A (en) * | 2021-09-03 | 2021-10-08 | 国网江西省电力有限公司电力科学研究院 | Power transmission line suspended foreign object target detection method based on deep learning |
| CN114842208A (en) * | 2022-06-08 | 2022-08-02 | 南昌大学 | Power grid harmful bird species target detection method based on deep learning |
| CN115082401A (en) * | 2022-06-22 | 2022-09-20 | 桂林电子科技大学 | SMT production line chip mounter fault prediction method based on improved YOLOX and PNN |
Non-Patent Citations (4)
| Title |
|---|
| TIANYU WANG等: "An improved YOLOv5-based bird detection algorithm for transmission line" * |
| 孙卫红等: "基于MSRCR与注意力机制的群体蚕茧智能识别算法" * |
| 李悦言等: "一种改进的轻量化目标检测算法" * |
| 陈金广等: "改进YOLOv5网络的轻量级服装目标检测方法" * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116229380A (en) * | 2023-05-10 | 2023-06-06 | 国网江西省电力有限公司电力科学研究院 | Method for identifying bird species related to bird-related faults of transformer substation |
| CN116229380B (en) * | 2023-05-10 | 2023-09-12 | 国网江西省电力有限公司电力科学研究院 | Method for identifying bird species related to bird-related faults of transformer substation |
| CN116912890A (en) * | 2023-09-14 | 2023-10-20 | 国网江苏省电力有限公司常州供电分公司 | Method and device for detecting birds in transformer substation |
| CN116912890B (en) * | 2023-09-14 | 2023-11-24 | 国网江苏省电力有限公司常州供电分公司 | Method and device for detecting birds in transformer substation |
| CN117095411A (en) * | 2023-10-16 | 2023-11-21 | 青岛文达通科技股份有限公司 | Detection method and system based on image fault recognition |
| CN117095411B (en) * | 2023-10-16 | 2024-01-23 | 青岛文达通科技股份有限公司 | Detection method and system based on image fault recognition |
| CN117874173A (en) * | 2024-03-11 | 2024-04-12 | 腾讯科技（深圳）有限公司 | Training method and related device of vector model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115620107A (en) | Transformer substation bird-involved fault related bird species identification method based on deep learning | |
| CN113486865B (en) | Power transmission line suspended foreign object target detection method based on deep learning | |
| CN114627360B (en) | Substation equipment defect identification method based on cascade detection model | |
| CN110968866B (en) | Defense method for resisting attack for deep reinforcement learning model | |
| CN109446925A (en) | A kind of electric device maintenance algorithm based on convolutional neural networks | |
| CN115167442A (en) | Power transmission line inspection path planning method and system | |
| CN115861853A (en) | Transmission line bird nest detection method in complex environment based on improved yolox algorithm | |
| CN114943689A (en) | Method for detecting components of steel cold-rolling annealing furnace based on semi-supervised learning | |
| CN115019302A (en) | Improved YOLOX target detection model construction method and application thereof | |
| CN115410087A (en) | Transmission line foreign matter detection method based on improved YOLOv4 | |
| CN112861646B (en) | Cascade detection method for oil unloading worker safety helmet in complex environment small target recognition scene | |
| Yan et al. | Learning probabilistic multi-modal actor models for vision-based robotic grasping | |
| CN113011308A (en) | Pedestrian detection method introducing attention mechanism | |
| CN115690730A (en) | High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation | |
| CN114862768A (en) | Improved YOLOv5-LITE lightweight-based power distribution assembly defect identification method | |
| CN114677357A (en) | Model, method and equipment for detecting self-explosion defect of aerial photographing insulator and storage medium | |
| CN117974985A (en) | High-voltage line foreign matter detection method and system based on improvement YOLOv s | |
| CN114359167A (en) | Insulator defect detection method based on lightweight YOLOv4 in complex scene | |
| CN112837281A (en) | Pin defect identification method, device and equipment based on cascade convolutional neural network | |
| CN116152699B (en) | Real-time moving target detection method for hydropower plant video monitoring system | |
| CN110110651B (en) | Method for identifying behaviors in video based on space-time importance and 3D CNN | |
| CN116385950A (en) | Electric power line hidden danger target detection method under small sample condition | |
| CN112199980A (en) | Overhead line robot obstacle identification method | |
| CN110991366A (en) | Shipping monitoring event identification method and system based on three-dimensional residual error network | |
| CN112233071A (en) | Multi-granularity hidden danger detection method and system based on power transmission network picture in complex environment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20230117 |