CN112380985A - Real-time detection method for intrusion foreign matters in transformer substation - Google Patents

Real-time detection method for intrusion foreign matters in transformer substation Download PDF

Info

Publication number
CN112380985A
CN112380985A CN202011266814.1A CN202011266814A CN112380985A CN 112380985 A CN112380985 A CN 112380985A CN 202011266814 A CN202011266814 A CN 202011266814A CN 112380985 A CN112380985 A CN 112380985A
Authority
CN
China
Prior art keywords
image
transformer substation
intrusion
foreign matter
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011266814.1A
Other languages
Chinese (zh)
Inventor
李凯
刘生寒
李波
肖建毅
钟苏生
黄恺彤
梁运德
陈力
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangdong Electric Power Information Technology Co Ltd
Original Assignee
Guangdong Electric Power Information Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guangdong Electric Power Information Technology Co Ltd filed Critical Guangdong Electric Power Information Technology Co Ltd
Priority to CN202011266814.1A priority Critical patent/CN112380985A/en
Publication of CN112380985A publication Critical patent/CN112380985A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/50Context or environment of the image
    • G06V20/52Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00Geometric image transformations in the plane of the image
    • G06T3/40Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4007Scaling of whole images or parts thereof, e.g. expanding or contracting based on interpolation, e.g. bilinear interpolation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/20Image preprocessing
    • G06V10/30Noise filtering
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/56Extraction of image or video features relating to colour

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a transformer substation foreign matter intrusion real-time detection method. The method comprises the following steps: preprocessing a monitoring image of the transformer substation, including image denoising and image enhancement; inputting the preprocessed image into a trained transformer substation intrusion foreign matter detection model, and outputting a detection result; the transformer substation intrusion foreign matter detection model is an improved YoLo network model, the improved YoLo network model utilizes a YoLo network as a basic structure, a batch normalization layer is added behind each convolution layer to normalize the convolved tensor, a modified Dropout layer, namely an R-Dropout layer, is used behind the batch normalization layer to construct uncertain convolution characteristics, and a pooling layer is connected behind the R-Dropout layer to perform down-sampling to complete extraction of uncertain convolution characteristics. The invention can ensure the detection speed, effectively improve the detection precision and effectively realize the real-time detection when the foreign matter invades.

Description

Real-time detection method for intrusion foreign matters in transformer substation
Technical Field
The invention relates to safe operation and maintenance of a transformer substation, in particular to a real-time detection method for intrusion foreign matters in the transformer substation.
Background
Under the conditions of economic growth and urban population increase, power enterprises need to ensure normal electricity utilization in life and work of people and take charge of quality and safety of electricity utilization. The safe operation of the transformer substation is the basic guarantee of the power supply of the power system. The safety production is the most central requirement in the daily production process of the transformer substation. On the one hand, on the construction site of the substation, obstacles which affect the safety of the substation, such as mistakenly-intruding animals, dropped suspended objects (plastic films, kites and the like), detained worker-related equipment and the like, cause very serious accidents in the environment of high-speed operation of equipment when even small foreign matters intrude into the equipment, and for example, the intruded foreign matters can cause line faults and even cause local or even regional power failure. On the other hand, when the constructors enter a transformer substation construction site, a plurality of high-voltage dangerous devices need to be faced, so that a very large potential safety hazard exists, and the conditions that the constructors who do not receive electric power safety training enter the transformer substation by mistake or the workers are left in a dangerous operation area due to alarm errors can cause a very large safety problem, and even death can be caused seriously. Therefore, in order to ensure the safety of a power production construction site of a transformer substation and ensure that the transformer substation can operate safely, reliably, efficiently and stably, automatic monitoring must be effectively carried out aiming at the invasion limit of foreign matters, foreign matters invading the limit can be found in time, the condition of the transformer substation can be accurately known, dangerous situations can be found as soon as possible, but the problem of detection of the invading foreign matters can be solved to a certain extent by the conventional foreign matter detection method, but the detection speed and the detection precision are not high, and the effective real-time detection of the invading foreign matters cannot be well carried out.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the defects of the prior art, the invention provides the transformer substation intrusion foreign matter real-time detection method, which can effectively improve the detection precision and effectively realize the real-time detection during foreign matter intrusion while ensuring the detection speed.
The technical scheme is as follows: a transformer substation foreign matter intrusion real-time detection method comprises the following steps:
preprocessing a monitoring image of the transformer substation, including image denoising and image enhancement;
inputting the preprocessed image into a trained transformer substation intrusion foreign matter detection model, and outputting a detection result; the transformer substation intrusion foreign matter detection model is an improved YoLo network model, the improved YoLo network model utilizes a YoLo network as a basic structure, a batch normalization layer is added behind each convolution layer to normalize the convolved tensor, an R-Dropout layer is used behind the batch normalization layer to construct uncertain convolution characteristics, and a pooling layer is connected behind the R-Dropout layer to perform down-sampling to complete extraction of the uncertain convolution characteristics.
Further, the transformer substation intrusion foreign matter detection model optimizes the prediction frame error in the training process, specifically by increasing the weight lambda of the loss function of the boundary frame coordinate predictioncoordAnd reducing the weight λ of the loss function of the confidence of the bounding box not containing the objectnoobjTo be implemented.
Further, the predicted bounding box error is calculated in the form of the square root of the height h and width w.
Further, the image denoising adopts a wavelet transform mode maximum denoising method to denoise the image.
Further, the image enhancement adopts a contrast adjustment algorithm based on local histogram clipping equalization to improve the contrast of the image and enhance the detail information of the image.
Has the advantages that: the method is mainly based on the YoLo algorithm to carry out real-time detection on the intrusion foreign matters in the dangerous environment of the transformer substation, the YoLo network is applied to the detection on the intrusion foreign matters in the environment of the transformer substation by utilizing the characteristic of high detection speed of the YoLo network, the distribution of input tensors is normalized by adding batch standardization in order to improve the detection accuracy, and meanwhile, the accuracy of the detection network is improved by constructing an uncertain convolution characteristic, so that the characteristic of high detection speed of the YoLo network is maintained, the detection accuracy is also improved, and end-to-end real-time detection is realized.
Drawings
Fig. 1 is a flow chart of a method for detecting intrusion foreign matters in a transformer substation in real time according to an embodiment of the invention;
fig. 2 is a block diagram of an improved YoLo network model according to an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further explained by combining the attached drawings.
Referring to fig. 1, the invention provides a real-time detection method for foreign matter invasion of a transformer substation, which comprises the following steps:
and step S1, preprocessing the monitoring image of the transformer substation area.
Due to the influence of different environmental factors such as weather, temperature and the like, images acquired by the transformer substation monitoring camera have distinct imaging characteristics such as low resolution, low saturation, reduced contrast, noise of different degrees and the like. The characteristics influence each process of transformer substation foreign matter intrusion detection, images under different scene conditions need to be preprocessed before being input into an intrusion foreign matter detection system, and the preprocessing of the images is mainly carried out by image denoising and image enhancement.
(1) Image denoising
Because noise always exists in the image acquisition process and the existence of the noise affects the image processing, the invention mainly adopts a wavelet transform modulus maximum denoising method to denoise the image. According to the difference of the propagation characteristics of the signal and the noise of the image on the wavelet transformation scale, eliminating the mode maximum value points generated by the noise, reserving the mode maximum value points corresponding to the signal, and utilizing the reserved mode maximum value points to reconstruct the signal so as to denoise the image.
The image denoising method comprises the following specific implementation steps:
s1-1, discrete binary wavelet transform is carried out on the detection image containing noise, 4 or 5 scales are selected to make the number of the information number modulus maximum points dominate under the maximum decomposition scale, and the important singular points of the signal are not lost.
S1-2, calculating the wavelet transform coefficient W on each scale2f corresponding to the modulo maximum point.
S13, in the maximum decomposition scale J, the wavelet transform module maximum is almost completely controlled by the signal, a threshold is selected, so that the point with the module maximum smaller than the threshold is taken as noise to be removed, and a new module maximum point on the maximum scale is obtained.
S1-4, starting from each modulo maximum point on the dimension J, the Adhoc algorithm searches up its corresponding modulo maximum curve. Specifically, a propagation point corresponding to each mode maximum value point on the scale J is searched on the scale J-1(J,.., 4,3), the mode maximum value point generated by the signal is reserved, the mode maximum value point caused by noise is removed, points which are not on any mode maximum curve on each scale J are removed, and the step-by-step search is carried out until the scale J is 2.
S1-5, for the scale j equal to 1, the corresponding extreme point when j equal to 1 is retained at the position where j equal to 2 has the extreme point, and the extreme points at the rest positions are set to zero.
And S1-6, reconstructing the signal by an alternative projection method according to the module extreme point reserved on each scale and the position of the extreme point.
By the method, singular point information of the signal is greatly reserved, the oscillation phenomenon of the denoised signal is eliminated, the characteristics of the original signal are effectively reserved, and the denoised image has better picture quality.
(2) Image enhancement
The image enhancement processing is mainly used for improving the quality of the image, strengthening certain characteristics of the image, highlighting useful information and detail information of the image and improving the definition and the interpretability of the image. The invention mainly improves the contrast of the image and enhances the detail information of the image through a contrast adjusting algorithm based on local histogram cutting equalization. Dividing the image into a plurality of sub-blocks according to the selected optimized horizontal and vertical grid numbers, counting the histogram information of each sub-block, mixing the histogram information with the integral histogram information of the original image in a certain way, cutting and equalizing the mixed histogram, obtaining a new mapping histogram of each block, then carrying out multi-point sampling on the data of the mapping table, and carrying out interpolation on sampling points to obtain new pixel values so as to improve the contrast of the image.
The image enhancement is realized by the following steps:
s1-a, reasonably selecting the horizontal and vertical grid number. The reasonable selection of the grid number can have an important influence on the result, the excessive grid number can increase the calculation amount of the algorithm, and the insufficient grid number can reduce the implementation effect of the algorithm. The number of grids is selected according to the following principle, when the mean square deviation of the brightness of the image is smaller, the brightness of the whole image is consistent, more grids (8 × 8) are used, otherwise, less grids (4 × 4) are used.
S1-B, dividing the image according to the specified grid number, and calculating the histogram information of the red, green and blue (R, G, B) color channel of each image.
And S1-c, acquiring histogram information and brightness histogram information of R, G and B color channels of the whole image.
And S1-d, fusing the sub-block histogram and the global histogram.
And S1-e, fusing the fused result with the brightness histogram again.
S1-f, clipping the Histogram according to a method of a Limited Contrast Adaptive Histogram Equalization (CLAHE) algorithm, and then equalizing the clipped Histogram to obtain a mapping table of each small image.
And S1-g, re-interpolating the mapping table of the subblock to obtain a smoothed mapping table.
S1-h, each small block is subjected to bilinear interpolation according to the process of the CLAHE algorithm to obtain the final enhancement effect.
The algorithm not only avoids the excessive discordance of the change after the processing between different channels of the color image, but also inhibits the noise generated in the contrast adjustment process, prevents the image distortion caused by the excessive amplification of the information, and has small calculated amount and high speed.
And dividing the sample image subjected to image denoising and image enhancement into a training set and a test set for training and testing a subsequent detection model.
And step S2, constructing a transformer substation intrusion foreign matter detection model.
The detection of foreign matters in the transformer substation is generally realized by adopting a background modeling method, and the detection and the positioning of a dynamic target are completed by verifying whether each pixel point in the next frame of image of the video data meets a background model. YoLo is a convolutional neural network with a regression function, and can detect the positions and the types of a plurality of target objects in real time at one time. The YoLo network does not train the network in a sliding window selection mode, but directly trains a network model by using the whole image, and effectively distinguishes foreground objects from background areas.
To achieve end-to-end training and real-time detection speed while maintaining high average detection accuracy. The separate components of object detection are integrated into a single neural network, and the YoLo network uses the characteristics of the whole image to predict the position and the category of each frame, and the frames mark the position and the category of the detection target object in the detection process. First, the YoLo network divides the entire input image into S × S grid cells, which are responsible for detecting an object if the center of the object falls within the grid cell. During the training and testing process, each grid unit predicts B bounding boxes and the confidence scores corresponding to each bounding box. The confidence score reflects the likelihood of the presence of the target object within the current bounding box and the accuracy with which the bounding box predicts the location of the target object. The YoLo network defines the confidence as
Figure BDA0002776383170000051
If the target object does not exist in the frame, the confidence pr (object) is 0. If the target object exists in the frame, calculating according to the predicted frame pred and the actual frame truth (the actual frame is the actual position and the type of the target object marked by the ground truth)
Figure BDA0002776383170000052
A confidence score for the bounding box is obtained.
Each frame contains 5 parameters: x, y, w, h and confidence scores. Coordinates (x, y) represent the predicted bounding box position relative to the grid cell boundaries, width w and height h are the predicted bounding box size relative to the entire image, and a confidence score is calculated for each predicted bounding box pred and the actual bounding box truth
Figure BDA0002776383170000053
And (4) obtaining the product.
The YoLo network also sacrifices the detection accuracy when the detection speed is increased, and the distribution of the input of each layer is changed all the time in the network training process, thereby increasing the training difficulty. Therefore, the present invention improves on these two problems. In the embodiment of the invention, YoLo v1 is adopted as a basic network model to be improved.
(1) The distribution of the input tensors of each layer of the YoLo network is changed all the time in the training process, so that the difficulty of the training process of the YoLo network is increased.
The basic idea of Batch Normalization is to change the distribution of an input tensor and an original corresponding input tensor when the input tensor is subjected to convolution operation, and perform correction operation on the current tensor, thereby fixing the mean value and variance of each layer of input tensor. Batch normalization is generally used before a nonlinear activation function, normalized tensors are normalized, the mean values of the output tensors are all 0, the variance is 1, and the normalized tensors have a stable distribution in a network training process.
(2) The YoLo network model is simple in composition, the speed of target object detection is improved, and the accuracy is sacrificed.
In order to improve the accuracy of the YoLo network detection, the uncertainty of a detection system is introduced into a network model, so that the stability and the accuracy of the YoLo network detection are enhanced. Constructing the uncertainty convolution feature using the modified Dropout (R-Dropout) after the convolutional layer, where the constructed structure is: conv + BN + R-Dropout + Pooling.
Suppose that the 3D tensor X ∈ RW×H×C
Figure BDA0002776383170000054
The normalized tensor is pooled for convolution. Tensor normalized to batch
Figure BDA0002776383170000055
Performing a non-linear mapping (activation function), and assuming that g (-) is a non-linear activation function, it is represented linearly
Figure BDA0002776383170000061
Is defined as:
Figure BDA0002776383170000062
in the formula, "" indicates the product of cells, M "is a binary mask matrix with dimension W' × Hi,jSatisfying Bernoulli distribution (M)i,jBernoulli (p)). W 'and H' represent the length and width of a binary mask matrix.
The nonlinear activation function g (-) is shown in trainingWhile Dropout is activated, linearizing the activation function
Figure BDA0002776383170000063
Indicating that Dropout is activated at test time. Many commonly used activation functions such as Sigmod, Tanh, ReLU, and lrefu all have the property that g (0) ═ 0.
Thus, rewritten as the R-Dropout formula:
Figure BDA0002776383170000064
in the formula, the first step is that,
Figure BDA0002776383170000065
still binary, is a learnable tensor, and b is a constant. The formulation introduces a randomness at the input of the activation function, which means that a specific probability distribution is used during each training iteration to generate the learnable tensor S ', each update process of the tensor S' is random, rather than specific.
Extracting uncertain convolution characteristics by constructing a structure of Conv + BN + R-Dropout + Pooling, inputting batch normalized tensor
Figure BDA0002776383170000066
The forward propagation of (c) is expressed as:
Figure BDA0002776383170000067
Figure BDA0002776383170000068
wherein Pooling (. cndot.) represents the maximal Pooling function.
Figure BDA0002776383170000069
The pooling area j at level l is shown,
Figure BDA00027763831700000610
is shown in
Figure BDA00027763831700000611
Activation function of the neural units in the region, flDenotes the convolution operation of the l-th layer CNN, and l denotes the l-th layer. f is the convolution operation of CNN.
Figure BDA00027763831700000612
To represent
Figure BDA00027763831700000613
The number of neural units in the region. To represent the uncertainty convolution characteristic while maintaining the general characteristics, the activation function in pooling region j is set
Figure BDA0002776383170000071
Arranged in a descending order of the first,
Figure BDA0002776383170000072
thus, it is possible to provide
Figure BDA0002776383170000073
The conditions that may be used as an activation function for pooling are:
Figure BDA0002776383170000074
while being retained thereby
Figure BDA0002776383170000075
Is discarded. According to the probability theory description, the probability of the event occurrence is PiThen, there are:
Figure BDA0002776383170000076
p=1-q
where q is the probability that the ith event does not occur. Thus, performing the R-Dropout operation before maximum pooling can be modeled as sampling an index i from the following polynomial, and the pooled activation function is the only one
Figure BDA0002776383170000077
Figure BDA0002776383170000078
Wherein P is0(=qn) Indicating that all activation functions in the pooling layer are discarded. The structure shows the effectiveness of construction uncertainty by introducing R-Dropout into the convolutional layer, and improves the detection accuracy of the YoLo network model. Fig. 2 shows a block diagram of an improved YoLo network model.
According to the method, the Batch Normalization is added behind the middle layer of the YoLo network model, so that the problem that tensor distribution change caused by convolution operation brings difficulty to the training process is solved. In order to improve the accuracy of the YoLo network detection, the invention introduces the uncertainty of the detection system into the network model, and enhances the stability and the accuracy of the YoLo network detection.
And step S3, performing network training and testing according to the improved YoLo network model.
For pre-training, the first 20 convolutional layers in the network model are selected, an average pooling layer and a full-link layer are added, and compiling is performed on the data set. And then, converting the model into an execution detection mode, wherein 4 new convolutional layers and 2 full-connection layers are required to be added after 20 pre-trained convolutional layers, the performance of the network can be improved by adding the convolutional layers and the full-connection layers in the pre-trained network, and the weights of the newly added layers are initialized randomly. To improve the detected visual information, the input resolution of the network model is increased from 224 × 224 to 448 × 448.
The last fully-connected layer of the network model can predict the class probability of the invader and the coordinates of the bounding box (namely the position of the invader). The height and width of the bounding box need to be normalized according to the height h and width w of the image, so that the height h and width w of the bounding box are normalized to be between 0 and 1. Meanwhile, the position coordinate information (x, y) of the bounding box is parameterized as the offset of the specific grid cell position, and also needs to be normalized to be between 0 and 1.
For the fully-connected layer that the last layer in the network model uses to detect, a linear activation function is used, while all other layers use the following modified linear activation function:
Figure BDA0002776383170000081
in order to improve the accuracy of network model detection, the network model optimizes the sum of squares error in the model output. Model instability may result because the optimal sum of squares error does not meet the goal of maximizing average accuracy, and equally weighted localization of classification errors is not ideal. To solve this problem, the weight of the loss function predicted by the bounding box coordinates is added, and the weight of the loss function that does not include the confidence of the bounding box of the object is reduced. Mainly by introducing two parameters lambdacoordAnd λnoobjTo achieve this.
When the sum of squares errors weight the prediction frame errors with different sizes, the error weights of all the prediction frames are set to be consistent, and experimental results show that the errors of the large prediction frames are much smaller than those of the small prediction frames. Therefore, to solve this problem, the error of the predicted bounding box is calculated in the form of the square root of the height h and width w.
During training, the following multi-part loss function needs to be optimized:
Figure BDA0002776383170000082
the loss function has 5 parts, wherein the first two terms represent coordinate errors, the third and fourth terms represent intersection ratio IOU errors, and the last term represents classification errors. In the formula (I), the compound is shown in the specification,
Figure BDA0002776383170000083
the representation object appears in the grid cell i,
Figure BDA0002776383170000084
the jth predicted bounding box representing cell i detects the object,
Figure BDA0002776383170000085
the jth predicted bounding box representing cell i does not detect the object. S is the square root of the total grid cell number, B is the predicted number of frames in a certain grid cell, xi,yiPosition coordinates, ω, representing a bounding boxi、hiRespectively representing the width and height of the predicted frame, CiMinimum bounding rectangle representing two boxes, pi(c) And representing the confidence score of the classification result as the class c. In the expression of the loss function, xi,yi、Ci、ωi、hi、pi(c) In order to predict the value of the network,
Figure BDA0002776383170000086
are labeled values.
If an object appears in the grid cell, the loss function only penalizes classification errors, while the predicted bounding box of the grid cell has a higher value
Figure BDA0002776383170000091
(the actual IOU value of the predictor variable), only penalty the bounding box position coordinates errors.
In the testing process, the images in the data set are selected to test the obtained improved YoLo network, the preprocessed transformer substation images are tested, and the predicted boundary box and class probability are accurate. And one grid unit can directly detect a boundary frame corresponding to one object, but for some objects with larger size or close to the boundary of the transformer substation, a plurality of grid prediction results are needed, the method is processed by a non-maximum suppression algorithm, scores of all frames are sorted, the highest score and the frame corresponding to the highest score are selected, other frames are traversed, if the overlapping area (IOU) of the highest score frame with the current frame is larger than a certain threshold value, the frames are deleted, one frame with the highest score is selected from the unprocessed frames, the process is repeated, redundant repeated boundary frames are eliminated, and the position of the frame detected by the best object is found.
In the embodiment, a sample image is used for testing the model, in practical application, the transformer substation image to be detected is preprocessed and then input into the trained model for detection, the recognition result of the image is output, and the detection and recognition of the invading foreign matter are completed.

Claims (9)

1. A transformer substation foreign matter intrusion real-time detection method is characterized by comprising the following steps:
preprocessing a monitoring image of the transformer substation, including image denoising and image enhancement;
inputting the preprocessed image into a trained transformer substation intrusion foreign matter detection model, and outputting a detection result; the transformer substation intrusion foreign matter detection model is an improved YoLo network model, the improved YoLo network model utilizes a YoLo network as a basic structure, a batch normalization layer is added behind each convolution layer to normalize the convolved tensor, a modified Dropout layer, namely an R-Dropout layer, is used behind the batch normalization layer to construct uncertain convolution characteristics, and a pooling layer is connected behind the R-Dropout layer to perform down-sampling to complete extraction of uncertain convolution characteristics.
2. The method for detecting the intrusion foreign matter in the substation according to claim 1, wherein the training process of the trained substation intrusion foreign matter detection model comprises optimizing the weight of the predicted frame error, specifically by adding the weight λ of the loss function predicted by the coordinate of the boundary framecoordAnd reducing the weight λ of the loss function of the confidence of the bounding box not containing the objectnoobjTo be implemented.
3. The transformer substation foreign object intrusion real-time detection method according to claim 2, wherein the loss function of the foreign object intrusion detection model is as follows:
Figure FDA0002776383160000011
in the formula (I), the compound is shown in the specification,
Figure FDA0002776383160000012
the representation object appears in the grid cell i,
Figure FDA0002776383160000013
the jth predicted bounding box representing cell i detects the object,
Figure FDA0002776383160000014
the jth predicted bounding box representing cell i does not detect the object, S is the square root of the total number of grid cells, B is the number of predicted bounding boxes in a certain grid cell, xi,yiPosition coordinates, ω, representing a bounding boxi、hiTo predict the width and height of the bounding box, pi(c) And scoring the confidence level of the classification result as the class c.
4. The substation intrusion foreign matter real-time detection method according to claim 1, wherein the calculation formula of the R-Dropout layer is as follows:
Figure FDA0002776383160000021
in the formula (I), the compound is shown in the specification,
Figure FDA0002776383160000022
for convolving the normalized tensor in parallel, the nonlinear activation function g (-) indicates activation Dropout during training, the activation function is linearized
Figure FDA0002776383160000023
Indicating that Dropout is activated at test, which indicates the cell product,
Figure FDA0002776383160000024
representing the learning tensor, b is a constant, and M is a binary quantity with the dimension W' × HMask matrix, each element M in Mi,jSatisfying Bernoulli distribution Mi,j~Bernoulli(p)。
5. The method for detecting the foreign matter invasion of the transformer substation in real time according to claim 1, wherein the image denoising adopts a wavelet transform mode maximum denoising method to denoise the image.
6. The method for detecting the foreign matter invasion of the transformer substation in real time according to claim 5, wherein the wavelet transform mode maximum denoising method comprises the following steps:
s1-1, performing discrete binary wavelet transform on the detection image containing the noise, and selecting a decomposition scale;
s1-2, calculating the wavelet transform coefficient W on each scale2f corresponding module maximum value point;
s13, selecting a threshold value in the maximum decomposition scale J, so that the point with the modulus maximum value smaller than the threshold value is taken as noise removal, and a new modulus maximum value point in the maximum scale is obtained;
s1-4, starting from each mode maximum value point on the maximum decomposition scale J, searching the corresponding mode maximum value curve upwards by using an Adhoc algorithm, removing points which are not on any mode maximum value curve on each scale J, and searching step by step until the scale J is 2;
s1-5, for the scale j equal to 1, reserving the corresponding extreme point when j equal to 1 at the position where j equal to 2 has the extreme point, and setting the extreme points at the rest positions to zero;
and S1-6, reconstructing the signal by an alternative projection method according to the module extreme point reserved on each scale and the position of the extreme point.
7. The method for detecting the intrusion foreign matter in the substation according to claim 1, wherein the image enhancement adopts a contrast adjustment algorithm based on local histogram clipping equalization to improve the contrast of the image and enhance the detail information of the image.
8. The transformer substation foreign object intrusion real-time detection method according to claim 7, wherein the image enhancement comprises the following steps:
s1-a, reasonably selecting the number of horizontal and vertical grids;
S1-B, dividing the image according to the specified grid number, and calculating the histogram information of the red, green and blue (R, G, B) color channel of each image;
s1-c, acquiring histogram information and brightness histogram information of R, G and B color channels of the whole image;
s1-d, fusing the sub-block histogram and the global histogram;
s1-e, fusing the fused result with the brightness histogram again;
s1-f, clipping the histogram according to a mode of limiting the contrast self-adaptive histogram equalization CLAHE algorithm, and then equalizing the clipped histogram to obtain a mapping table of each small image;
s1-g, re-interpolating the mapping table of the subblock to obtain a smoothed mapping table;
s1-h, each small block is subjected to bilinear interpolation according to the process of the CLAHE algorithm to obtain the final enhancement effect.
9. The transformer substation foreign object intrusion real-time detection method according to claim 1, further comprising: when some objects obtain the results of multiple grid predictions, the scores of all frames are sorted, the highest score and the frame corresponding to the highest score are selected, the rest frames are traversed, if the overlapping area of the highest score and the current highest score frame is larger than a certain threshold value, the frames are deleted, one frame with the highest score is continuously selected from the unprocessed frames, and the process is repeated, so that redundant repeated boundary frames are eliminated, and the position of the frame detected by the best object is found.
CN202011266814.1A 2020-11-13 2020-11-13 Real-time detection method for intrusion foreign matters in transformer substation Pending CN112380985A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011266814.1A CN112380985A (en) 2020-11-13 2020-11-13 Real-time detection method for intrusion foreign matters in transformer substation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011266814.1A CN112380985A (en) 2020-11-13 2020-11-13 Real-time detection method for intrusion foreign matters in transformer substation

Publications (1)

Publication Number Publication Date
CN112380985A true CN112380985A (en) 2021-02-19

Family

ID=74583746

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011266814.1A Pending CN112380985A (en) 2020-11-13 2020-11-13 Real-time detection method for intrusion foreign matters in transformer substation

Country Status (1)

Country Link
CN (1) CN112380985A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114003900A (en) * 2021-10-25 2022-02-01 广东电网有限责任公司广州供电局 Network intrusion detection method, device and system for secondary system of transformer substation
CN116883648A (en) * 2023-09-06 2023-10-13 南方电网数字电网研究院有限公司 Foreign matter detection method and device, electronic equipment and storage medium
CN118038021A (en) * 2024-02-27 2024-05-14 国网湖北省电力有限公司荆门供电公司 Transformer substation operation site foreign matter intrusion detection method based on improvement yolov4

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016408A (en) * 2017-03-17 2017-08-04 中国南方电网有限责任公司超高压输电公司曲靖局 A kind of empty foreign matter method for inspecting of the extension based on Intelligent Mobile Robot
CN109829429A (en) * 2019-01-31 2019-05-31 福州大学 Security protection sensitive articles detection method under monitoring scene based on YOLOv3
CN110647852A (en) * 2019-09-27 2020-01-03 集美大学 Traffic flow statistical method, terminal equipment and storage medium

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016408A (en) * 2017-03-17 2017-08-04 中国南方电网有限责任公司超高压输电公司曲靖局 A kind of empty foreign matter method for inspecting of the extension based on Intelligent Mobile Robot
CN109829429A (en) * 2019-01-31 2019-05-31 福州大学 Security protection sensitive articles detection method under monitoring scene based on YOLOv3
CN110647852A (en) * 2019-09-27 2020-01-03 集美大学 Traffic flow statistical method, terminal equipment and storage medium

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
ZHANG P P 等: "Learning uncertain convolutional features for accurate saliency detection", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION》, pages 212 - 221 *
一心无二用,本人只专注于基础图像算法的实现与优化: "一种强化的基于局部直方图裁剪均衡化的对比度调节算法", pages 1 - 8, Retrieved from the Internet <URL:https://cloud.tencent.com/developer/article/1011944?areaSource=106005.20> *
付江涛: "基于多特征融合与深度先验信息的图像显著性检测算法研究", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》, pages 138 - 1535 *
曾宇 等: "贝叶斯优化卷积神经网络公共场所异常声识别", 《应用声学》, vol. 39, no. 3, pages 409 - 416 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114003900A (en) * 2021-10-25 2022-02-01 广东电网有限责任公司广州供电局 Network intrusion detection method, device and system for secondary system of transformer substation
CN116883648A (en) * 2023-09-06 2023-10-13 南方电网数字电网研究院有限公司 Foreign matter detection method and device, electronic equipment and storage medium
CN116883648B (en) * 2023-09-06 2024-02-13 南方电网数字电网研究院股份有限公司 Foreign matter detection method and device, electronic equipment and storage medium
CN118038021A (en) * 2024-02-27 2024-05-14 国网湖北省电力有限公司荆门供电公司 Transformer substation operation site foreign matter intrusion detection method based on improvement yolov4

Similar Documents

Publication Publication Date Title
CN111353413B (en) Low-missing-report-rate defect identification method for power transmission equipment
CN110059694B (en) Intelligent identification method for character data in complex scene of power industry
CN106595551B (en) Ice covering thickness detection method in powerline ice-covering image based on deep learning
CN108647655B (en) Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network
CN109359559B (en) Pedestrian re-identification method based on dynamic shielding sample
CN111583198A (en) Insulator picture defect detection method combining FasterR-CNN + ResNet101+ FPN
CN111160407B (en) Deep learning target detection method and system
CN109446925A (en) A kind of electric device maintenance algorithm based on convolutional neural networks
CN114022432A (en) Improved yolov 5-based insulator defect detection method
CN109801297B (en) Image panorama segmentation prediction optimization method based on convolution
CN112380985A (en) Real-time detection method for intrusion foreign matters in transformer substation
CN109886159B (en) Face detection method under non-limited condition
CN109948593A (en) Based on the MCNN people counting method for combining global density feature
CN117495735B (en) Automatic building elevation texture repairing method and system based on structure guidance
CN112419202A (en) Wild animal image automatic identification system based on big data and deep learning
CN111723749A (en) Method, system and equipment for identifying wearing of safety helmet
CN114881286A (en) Short-time rainfall prediction method based on deep learning
CN115239710A (en) Insulator defect detection method based on attention feedback and double-space pyramid
CN116310328A (en) Semantic segmentation knowledge distillation method and system based on cross-image similarity relationship
CN111027542A (en) Target detection method improved based on fast RCNN algorithm
CN114373162A (en) Dangerous area personnel intrusion detection method and system for transformer substation video monitoring
CN112132207A (en) Target detection neural network construction method based on multi-branch feature mapping
CN116152699B (en) Real-time moving target detection method for hydropower plant video monitoring system
CN117789037A (en) Crop growth period prediction method and device
CN117116096A (en) Airport delay prediction method and system based on multichannel traffic image and depth CNN

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