CN113139464B - Power grid fault detection method - Google Patents
Power grid fault detection method Download PDFInfo
- Publication number
- CN113139464B CN113139464B CN202110442697.8A CN202110442697A CN113139464B CN 113139464 B CN113139464 B CN 113139464B CN 202110442697 A CN202110442697 A CN 202110442697A CN 113139464 B CN113139464 B CN 113139464B
- Authority
- CN
- China
- Prior art keywords
- label
- power grid
- convolutional neural
- neural network
- network model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/182—Network patterns, e.g. roads or rivers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a power grid fault detection method, which comprises the following steps: the method comprises the steps of obtaining a power grid image by means of cruise shooting of an unmanned aerial vehicle, and dividing the obtained image into a training set and a test set; inputting the images of the training set into a deep learning convolutional neural network model to obtain a network classification label; inputting the obtained network classification label into a label alignment layer to obtain an aligned network classification label with the same format as the manually marked label; calculating AccLoss loss function values of the aligned network classification labels and the manually marked labels, performing back propagation by using the obtained AccLoss loss function values, and continuously adjusting parameters of the convolutional neural network model to finish training of the convolutional neural network model; and inputting the images of the test set into the trained convolutional neural network model for testing to obtain a power grid fault classification result. The method disclosed by the invention can obtain high accuracy in the power grid fault detection.
Description
Technical Field
The invention relates to the technical field of power grid detection, in particular to a power grid fault detection method.
Background
With the development of economy and society, the role played by the power grid is more and more important, so that the maintenance of the power supply reliability of the power grid is a primary task of power operation development. However, with the improvement of living standard of people, the scale of the power grid is continuously enlarged, the cost of manual power grid detection is gradually increased and difficult to implement, and how to realize automatic power grid fault detection by means of scientific and technological strength is of great importance. And the image obtained in the unmanned aerial vehicle cruising process can be used as the media information for judging the fault, namely, the power grid fault detection problem is defined as an image classification problem (a two-classification problem of a fault class and a non-fault class). This will greatly improve the efficiency of grid fault detection as well as the labor cost.
Image classification is to solve the problem whether pictures belong to a certain class, and is a basic research direction in the fields of artificial intelligence and computer vision. Generally, an object classification algorithm describes the whole image globally by a manual feature or feature learning method, and then uses a classifier to determine whether the image belongs to a certain class of objects. Most feature extraction processes are designed manually, but there is still a large "semantic gap" between these features and the high-level subject matter of the image. The deep learning completely learns the hierarchical structural characteristics of the image from the training data by using the set network structure, and can extract the abstract characteristics closer to the high-level semantics of the image, so the expression of the image recognition is far superior to that of the traditional method. Deep learning was proposed by Hinton et al in 2006. The convolutional neural network has great superiority in feature representation. Therefore, how to reduce the labor cost of power grid inspection and improve the fault detection accuracy rate through a deep learning mode is very important.
The existing image classification method is based on the improvement on a model structure, and loss functions used in training are all cross entropy loss functions. The cross entropy loss function is a concept in information theory and is originally used for estimating the average coding length. That is, given two probability distributions p and q, the cross-entropy formula for p by q is:
the larger the difference between the two probability distributions p and q, the larger the loss function. Where p is a known distribution, i.e., a label of the image; and q is a prediction result of the deep learning model. The cross entropy loss function is used for measuring the loss of classification accuracy from the perspective of probability theory, and can achieve better effect, but the cross entropy loss function can also achieve better effectIt is always an error function
(i.e., the deep learning model predicts the correct total number of samples) it is difficult to give the network a correct direction of optimization. The error function has judgment operation, so that the training and back propagation cannot be directly performed.
Disclosure of Invention
In order to solve the technical problems, the invention provides a power grid fault detection method, which utilizes certain calculation equivalent error functions capable of being reversely propagated and directly uses a calculation formula of accuracy rate to carry out reverse propagation, so that the network training convergence is faster and can converge to a better solution, the generalization capability is enhanced, and the method is applied to the fault detection of a power grid, so that the power grid fault detection obtains higher accuracy rate.
In order to achieve the purpose, the technical scheme of the invention is as follows:
a power grid fault detection method comprises the following steps:
(1) Acquiring a power grid image by using unmanned aerial vehicle cruise shooting, and dividing the acquired image into a training set and a test set;
(2) Inputting the images of the training set into a deep learning convolutional neural network model to obtain a network classification label;
(3) Inputting the obtained network classification label into a label alignment layer to obtain an aligned network classification label with the same format as the manually marked label; the label alignment layer performs maximum pooling layer operation, inverse maximum pooling layer operation and normalized maximum position operation on input data;
(4) Calculating AccLoss loss function values of the aligned network classification labels and the manually marked labels, performing back propagation by using the obtained AccLoss loss function values, and continuously adjusting parameters of the convolutional neural network model to finish training of the convolutional neural network model; the calculation formula of the AccLoss loss function value is as follows:
wherein x is the aligned network classification label, label is the manually labeled label, i is the ith sample, j is the sample category,
the aligned network class labels of class j for the ith sample,
for the artificially labeled label of the jth class of the ith sample, the ReLU is a neural network activation function, has a derivable property, changes the numbers less than 0 into 0, and does not change the numbers more than 0;
(5) And inputting the images of the test set into the trained convolutional neural network model for testing to obtain a power grid fault classification result.
In the above scheme, in step (3), the network classification label x 1 ＝[c 1 ，c 2 ]Obtaining the maximum score c of a certain class of images after passing through the maximum pooling layer max Then the maximum score c is given max Inputting the image into an inverse maximum pooling layer to obtain a network classification label x containing the index of the class to which the image belongs 2 ＝[c max ，0]Or x 2 ＝[0，c max ]Network classification label x containing an index to which an image belongs 2 Obtaining the aligned network classification label x = [1,0] in the same format as the manually marked label after the operation of normalizing the position of the maximum value]Or x = [0,1 =]。
In the above scheme, the convolutional neural network model deeply learned in step (2) includes a resnet, vgg, and Alexnet network model.
Through the technical scheme, the power grid fault detection method provided by the invention has the following beneficial effects:
(1) The method uses a series of conductible operations (maximum pooling layer operation, inverse maximum pooling layer operation and normalized maximum position operation) to directly calculate the cost function, the gradient obtained by directly calculating the classification accuracy rate is the optimal optimization direction, and the network training reaches a globally appropriate solution by giving the optimal optimization direction to the network.
(2) According to the invention, accurate optimization direction is provided for the model through accurate loss calculation, so that the effect of the deep learning model on an image recognition task is improved, and the classification accuracy and generalization capability are improved by applying the deep learning model in power grid fault detection.
(3) The invention constructs a data set for power grid fault detection, applies the Accloss function value to the power grid fault detection, and performs experiments in various convolutional neural networks, and the result shows that the accuracy of the power grid fault detection is greatly improved by the Accloss function provided by the invention.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below.
Fig. 1 is a flowchart of a method for detecting a power grid fault according to an embodiment of the present invention;
FIG. 2 is a basic architecture diagram of a network model training process according to an embodiment of the present invention;
FIG. 3 is a diagram of a label alignment layer structure according to an embodiment of the present invention;
fig. 4 is a flowchart of determining and calculating a loss function according to an embodiment of the present invention.
Detailed Description
The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
The invention provides a power grid fault detection method, which comprises the following steps as shown in figure 1:
(1) The method comprises the steps of obtaining a power grid image by means of cruise shooting of an unmanned aerial vehicle, and randomly dividing the obtained image into a training set and a testing set;
(2) Inputting the images of the training set into a deep learning convolutional neural network model (classification backbone network) to obtain a network classification label x 1 ＝[c 1 ，c 2 ](ii) a Deep learning convolutional neural network modelIncluding resnet, vgg, alexnet network models. The training process of the network model is shown in fig. 2.
(3) Inputting the obtained network classification label into a label alignment layer to obtain an aligned network classification label in the same format as a label labeled manually; as shown in fig. 3, the label alignment layer performs a max pooling layer operation, an inverse max pooling layer operation, and a normalized maximum position operation on input data;
network classification label x 1 ＝[c 1 ，c 2 ]Obtaining the maximum score c of a certain class of images after passing through the maximum pooling layer max Then the maximum score c is given max Inputting the image into the inverse maximum pooling layer to obtain a network classification label x containing the index of the image belonging to the class 2 ＝[c max ，0]Or x 2 ＝[0，c max ]These two operations correspond to indexing with the largest score taken while retaining the score.
The manually labeled label is a One-hot vector (One-hot example: label belonging to the first class is denoted label = [1,0 ]), and the result of label-to-output alignment is achieved by subjecting the output of the CNN to max pooling layer operation and anti-max pooling layer operation to the same One-hot form as label.
It is then necessary to continue to normalize the label to output in-format alignment. The error function of the accuracy is to use the decision operation of if (x = = label) to achieve this, which may lead to a collapse of the gradient. Therefore, the invention designs a conductive step to equivalently judge operation, and the operation is named as the position of the normalized maximum value.
It is calculated as follows:
(calculated in bits) epsilon is a very small number such as 10 -8 . If x =0, then 0 remains after the above calculation; if x = other (other numbers), 1 is obtained after the above calculation.
Network classification label x containing index to which image belongs 2 ＝[c max ，0]Or x 2 ＝[0，c max ]Through normalizationObtaining the aligned network classification label x = [1,0] with the same format as the label marked manually after the large value position operation]Or x = [0,1 =]。
(4) Calculating Accloss function values of the aligned network classification labels and the manually marked labels, performing back propagation by using the obtained Accloss function values, and continuously adjusting parameters of the convolutional neural network model to finish training of the convolutional neural network model;
after one-hot forms of network output and label are obtained through a label alignment layer, x of normalized maximum position output needs to be output 3 Loss calculations were performed with label. However, this is also a decision operation if (x = = label), and therefore, the present invention proposes an equivalent calculation method. The calculation flow is shown in fig. 4.
Each step of the operation of fig. 4 is a bitwise calculation. Take a pair of samples with wrong scores as an example: normalized maximum position output of x = [0,1 =]The label of the sample manual labeling is label = [1,0]. x = x + label yielding x = [1,1]And then x = x-1 to obtain x = [0,0 =]. At this time, the result of ReLU (ReLU is an activation function often used in deep learning, has a derivative property, and can change numbers smaller than 0 into 0, and numbers larger than 0 do not change) is x = [0,0]. Of the final sample
Wherein x is j Is the j-th class (j-dimension) score of x. Finally, the loss of the error sample is 1, i.e. the loss when the model predicts the error is 1.
If take a pair of samples with correct prediction as an example: normalized maximum position output of x = [1,0 =]The label of the sample manual labeling is label = [1,0]. After x = x + label, x = [2,0 ] is obtained]Then x = x-1 is passed to yield x = [1,0]]. At this time, the ReLU result is x = [1,0]Of the final sample
The calculation is 0, i.e. the loss is 0 if the model prediction is correct.
All of the above operations are guided, so model error prediction is computedThe loss function for the number of samples is derivable. To this end, the present invention is not conducive to using a series of operations
The cost function is computed directly in a derivable manner.
The final calculation formula of the Accloss function value is as follows:
wherein x is the aligned network classification label, label is the label marked manually, i is the ith sample, j is the class of the sample,
for the aligned network classification label of the jth class of the ith sample,
for the artificially labeled label of the jth class of the ith sample, the ReLU is a neural network activation function, has a derivative property, changes the numbers smaller than 0 into 0, and does not change the numbers larger than 0.
(5) Inputting the images of the test set into the trained convolutional neural network model for testing to obtain a power grid fault classification result, outputting a vector x, and judging c in x 1 ，c 2 Determining the prediction result of the network on the input image, c 1 ＞c 2 The network predicts that the input image is of the first classification, i.e. no failure, and vice versa.
According to the embodiment of the invention, the pictures related to the power grid are crawled from the network, and some pictures of the fault power grid are actually shot. The data set for detecting the grid fault is formed by 4000 positive samples (no fault) and 4000 negative samples (fault) in total. It was randomly divided into training and test sets. In the experiment, networks such as resnet, vgg and Alexnet are selected to train on a training set, tests are carried out on a test set, the only variables are different loss functions used in the training process, and Cross entry loss functions are selected to be compared with the Accloss functions of the invention. The test results are shown in table 1.
TABLE 1 grid Fault Classification error Rate
| Cross entropy loss | AccLoss | |
| Alexnet | 15.2％ | 12.1％ |
| VGG-16 | 12.3％ | 9.6％ |
| VGG-19 | 10.1％ | 8.3％ |
| Resnet-18 | 13.2％ | 10.0％ |
| Resnet-34 | 8.5％ | 6.5％ |
| Resnet-50 | 5.4％ | 3.8％ |
As can be seen from table 1, under the same backbone network, the error rate of classification is significantly reduced by the AccLoss function proposed in the present invention.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (2)
1. A power grid fault detection method is characterized by comprising the following steps:
(1) Acquiring a power grid image by using unmanned aerial vehicle cruise shooting, and dividing the acquired image into a training set and a test set;
(2) Inputting the images of the training set into a deep learning convolutional neural network model to obtain a network classification label;
(3) Inputting the obtained network classification label into a label alignment layer to obtain an aligned network classification label with the same format as that of a manually marked label; the label alignment layer performs maximum pooling layer operation, inverse maximum pooling layer operation and normalized maximum position operation on input data;
(4) Calculating AccLoss loss function values of the aligned network classification labels and the manually marked labels, performing back propagation by using the obtained AccLoss loss function values, and continuously adjusting parameters of the convolutional neural network model to finish training of the convolutional neural network model; the calculation formula of the Accloss function value is as follows:
wherein x is the aligned network classification label, label is the manually labeled label, i is the ith sample, j is the sample category,
the aligned network class labels of class j for the ith sample,
for the artificially labeled label of the jth class of the ith sample, the ReLU is a neural network activation function, has a derivable property, changes the numbers less than 0 into 0, and does not change the numbers more than 0;
(5) Inputting the images of the test set into the trained convolutional neural network model for testing to obtain a power grid fault classification result;
in step (3), network classification label x 1 ＝[c 1 ，c 2 ]Obtaining the maximum score c of a certain class of images after passing through the maximum pooling layer max Then the maximum score c is given max Inputting the image into an inverse maximum pooling layer to obtain a network classification label x containing the index of the class to which the image belongs 2 ＝[c max ，0]Or x 2 ＝[0，c max ]Web classification label x containing an index to which an image belongs 2 Obtaining the aligned network classification label x = [1,0] in the same format as the manually marked label after normalization maximum position operation]Or x = [0,1 =]；
The normalized maximum position is calculated as follows:
ε is a very small number; if x =0, then 0 remains after the above calculation; if x is other number, 1 is obtained after the above calculation.
2. The power grid fault detection method according to claim 1, wherein the deep-learning convolutional neural network model in step (2) comprises a resnet, vgg, or Alexnet network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110442697.8A CN113139464B (en) | 2021-04-23 | 2021-04-23 | Power grid fault detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110442697.8A CN113139464B (en) | 2021-04-23 | 2021-04-23 | Power grid fault detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113139464A CN113139464A (en) | 2021-07-20 |
| CN113139464B true CN113139464B (en) | 2022-11-18 |
Family
ID=76812512
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110442697.8A Active CN113139464B (en) | 2021-04-23 | 2021-04-23 | Power grid fault detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113139464B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117853923A (en) * | 2024-01-17 | 2024-04-09 | 山东盛然电力科技有限公司 | Power grid power infrastructure safety evaluation analysis method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034205A (en) * | 2018-06-29 | 2018-12-18 | 西安交通大学 | Image classification method based on the semi-supervised deep learning of direct-push |
| CN111639668A (en) * | 2020-04-17 | 2020-09-08 | 北京品恩科技股份有限公司 | Crowd density detection method based on deep learning |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110745A (en) * | 2019-03-29 | 2019-08-09 | 上海海事大学 | Based on the semi-supervised x-ray image automatic marking for generating confrontation network |
| CN111340123A (en) * | 2020-02-29 | 2020-06-26 | 韶鼎人工智能科技有限公司 | Image score label prediction method based on deep convolutional neural network |
| CN111783616B (en) * | 2020-06-28 | 2024-03-26 | 北京瓦特曼科技有限公司 | Nondestructive testing method based on data-driven self-learning |
-
2021
- 2021-04-23 CN CN202110442697.8A patent/CN113139464B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034205A (en) * | 2018-06-29 | 2018-12-18 | 西安交通大学 | Image classification method based on the semi-supervised deep learning of direct-push |
| CN111639668A (en) * | 2020-04-17 | 2020-09-08 | 北京品恩科技股份有限公司 | Crowd density detection method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113139464A (en) | 2021-07-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| CN112328767B (en) | Question-answer matching method based on BERT model and comparative aggregation framework | |
| CN111126386B (en) | Sequence domain adaptation method based on countermeasure learning in scene text recognition | |
| CN110532900B (en) | Facial expression recognition method based on U-Net and LS-CNN | |
| CN109101938B (en) | Multi-label age estimation method based on convolutional neural network | |
| CN110232395B (en) | Power system fault diagnosis method based on fault Chinese text | |
| CN105138973A (en) | Face authentication method and device | |
| CN111368920A (en) | Quantum twin neural network-based binary classification method and face recognition method thereof | |
| CN110415071B (en) | Automobile competitive product comparison method based on viewpoint mining analysis | |
| CN110969073B (en) | Facial expression recognition method based on feature fusion and BP neural network | |
| CN112732921B (en) | False user comment detection method and system | |
| CN105631477A (en) | Traffic sign recognition method based on extreme learning machine and self-adaptive lifting | |
| CN115205592A (en) | Multi-mode data based rebalance long-tail image data classification method | |
| CN111428481A (en) | Entity relation extraction method based on deep learning | |
| CN111651993A (en) | Chinese named entity recognition method fusing local-global character level association features | |
| CN116342942A (en) | Cross-domain target detection method based on multistage domain adaptation weak supervision learning | |
| CN113139464B (en) | Power grid fault detection method | |
| CN113392191B (en) | Text matching method and device based on multi-dimensional semantic joint learning | |
| CN108388918B (en) | Data feature selection method with structure retention characteristics | |
| CN110751005B (en) | Pedestrian detection method integrating depth perception features and kernel extreme learning machine | |
| CN116468935A (en) | Multi-core convolutional network-based stepwise classification and identification method for traffic signs | |
| Zhu | [Retracted] A Face Recognition System Using ACO‐BPNN Model for Optimizing the Teaching Management System | |
| CN112541081B (en) | Migratory rumor detection method based on field self-adaptation | |
| CN112800959B (en) | Difficult sample mining method for data fitting estimation in face recognition | |
| CN114357166A (en) | Text classification method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |