Insulator image pollution feature extraction method based on pulse coupled neural network

Abstract

The application discloses an insulator image pollution feature extraction method based on a pulse coupled neural network, and comprises the following steps: step S1: acquiring an insulator disc surface grayscale image; step S2: performing standardization processing on the insulator disc surface grayscale image; step S3: performing ignition processing on the standardized disc surface grayscale image by using a pulse coupled neural network to obtain a time assignment matrix; step S4: performing normalization processing on the time assignment matrix, and mapping the time assignment matrix to between 0 and 255 to reconstruct a grayscale image; and step S5: reducing the reconstructed grayscale image to obtain an insulator disc surface pollution feature image. The method adopts an image recognition method, extracts the pollution features of the insulator disc surface image from the disc surface image of the insulator, and provides a detection personnel with the pollution degree of the insulator.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment monitoring, in particular to an insulator image contamination feature extraction method based on a pulse coupled neural network. BACKGROUND

[0002] Dust particles in the atmosphere have always been a big problem for the power transmission industry in China. In dusty or air-polluted areas, dust pollutants are easy to deposit on insulators, increasing the surface contamination of insulators. Domestic and foreign research shows that the increase in the surface contamination of insulators is one of the main reasons for the decrease in insulator flashover voltage and the occurrence of insulator pollution flashover accidents. Insulator pollution flashover is an operational accident in the power system that has a large area of destructive power, occurs in a wide area, has a long power outage time, and has a low successful rate of reclosing, which seriously threatens the safe operation of the power system. Therefore, monitoring the surface contamination of insulators in key areas is of great significance for preventing pollution flashover accidents and ensuring the safe and stable operation of the power system.

[0003] The existing method for detecting the surface contamination of insulators is to set an electromagnetic coil at the bottom of the insulator. When the surface contamination of the insulator is high, the leakage current on the surface of the insulator increases. The leakage current on the surface of the insulator is detected through the electromagnetic coil. The detected leakage current is identified and judged after being filtered and amplified. However, the surface contamination of the insulator is only one of the factors that affect the leakage current of the insulator. Environmental humidity, rainwater, and the like can also cause the leakage current on the surface of the insulator to increase.

[0004] Therefore, the existing technology lacks an insulator image contamination feature extraction method based on a pulse coupled neural network. An image recognition method is used to extract the contamination features of the insulator disc surface image from the disc surface image of the insulator for detection personnel to identify the contamination degree of the insulator. SUMMARY

[0005] In view of at least one defect of the prior art, the present application aims to provide an insulator image contamination feature extraction method based on a pulse coupled neural network. An image recognition method is used to extract the contamination features of the insulator disc surface image from the disc surface image of the insulator for detection personnel to identify the contamination degree of the insulator.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: an insulator image contamination feature extraction method based on a pulse coupled neural network, comprising the following steps:

[0007] Step S1: an insulator monitoring device is built on the iron tower, the camera of the insulator monitoring device faces the insulator disc surface, and the insulator monitoring device acquires a gray-scale image of the insulator disc surface;

[0008] Step S2: Standardize the grayscale image of the insulator disc surface;

[0009] Step S3: Use a pulse-coupled neural network to perform ignition processing on the standardized disk grayscale image to obtain the timing matrix;

[0010] Step S4: Normalize the time assignment matrix and map it to the range [0, 255] to reconstruct the grayscale image;

[0011] Step S5: Reduce the size of the reconstructed grayscale image to obtain the contamination feature image of the insulator disk surface.

[0012] In step S1, the grayscale image of the insulator disk is obtained by preprocessing the visible light image of the insulator; the preprocessing includes image segmentation and image grayscale conversion.

[0013] In step S2, the grayscale image standardization process of the insulator disk surface is to unify the grayscale image of the disk surface to a set standard size, which facilitates subsequent feature extraction.

[0014] In step S3, the pulse-coupled neural network (PCNN) technique is used, and the standardized insulator disk grayscale image is ignited based on the FLM model. The expression of the FLM model is as follows:

[0015]

[0016]

[0017] U ij (n)=fU ij (n-1)+F ij (n)(1+βL ij (n)) (3);

[0018] Θ ij (n)=gΘ ij (n-1)+V θ Y ij (n-1) (4);

[0019]

[0020] In the above formula, each neuron is represented by (i,j), and one of its neighboring neurons is represented by (k,l), where n is the number of iterations, and F... ij (n) represents the feedback input, Y kl (n-1) represents the postsynaptic action potential, M ijkl and W ijkl S represents the connection weight matrices for the feedback input domain and the coupled connection domain, respectively. ij As an external incentive, S ij That is, the grayscale value of an image pixel, U ij(n) represents the membrane potential of the neuron, U ij (n-1) represents the membrane potential of the neuron in the previous iteration, d is the global inhibition normal, f is the decay time constant of the membrane potential, β is the connection strength, and Θ ij (n) is the threshold of the neuron, g is the threshold decay time constant, and V θ Y is the amplitude adjustment factor. ij (n) represents the pulse output, Y ij (n) represents the firing state of the neuron; L ij (n) is the input connection; Θ ij (n-1) is the threshold of the neuron in the previous iteration, Y ij (n-1) represents the pulse output from the previous iteration;

[0021] In a PCNN model, the number of neurons equals the number of pixels in the input image. The output of a neuron depends on the pixel information corresponding to that neuron. A time-assignment matrix T is defined, whose size is equal to the external input S and the output Y, where S represents the input matrix. ij Y represents the elements in the input matrix S; Y represents the output matrix. ij (n) represents an element in the output matrix Y; and each element in T corresponds one-to-one with a neuron associated with S and Y. T stores the time or frequency information associated with the firing time of each neuron, i.e., T ij It is the firing time or number of times of the (i,j)th neuron, satisfying the firing information T of the (i,j)th neuron at the nth time. ij (n):

[0022]

[0023] The timing matrix accurately records information related to the firing time of each neuron, reflecting the temporal domain information based on image spatial information.

[0024] In step S4, the time assignment matrix T is normalized and mapped to [0, 255] using the maximum-minimum normalization method: Where T i,j Let T′ be the element in the i-th row and j-th column of the time-assignment matrix. i,j Let x be the element in the i-th row and j-th column of the normalized matrix. j(min) x is the minimum value of the j-th column of the time-assigned matrix. j(max)The maximum value in the j-th column of the time-assignment matrix is ​​1. The maximum value in the normalized matrix T′ is 1, and the minimum value is 0. Multiplying T′ by 255 maps the data in the matrix to [0,255] to obtain a reconstructed grayscale image. This grayscale image reflects the main contamination firing feature information of the image. Through PCNN firing threshold processing, secondary contamination information and interference information are removed. The time-assignment matrix truly records the information related to the firing time of each neuron, which are respectively from the following cases:

[0025] ① If the neuron has never been ignited, the value at the corresponding point in T is 0;

[0026] ②If the first ignition occurs at time n, then the value of the corresponding point in T is n;

[0027] ③ If the neuron has already been fired, the value of the corresponding point in T remains unchanged from the previous value; the firing information of the (i,j)th neuron for the nth time is shown in T. ij (n) equals the previous fire information T. ij (n-1).

[0028] Adjacent pixels in the image are ignited simultaneously through coupling connections, and the iterative process continues until the set number of iterations is reached, forming the PCNN timing matrix; in expression (4), the postsynaptic action potential Y kl (n-1) makes the threshold Θ ij (n) generates an increment that prevents secondary action potentials from being generated within a certain period of time. At the same time, the increased threshold decays with the time constant g. Therefore, the dynamic threshold function in PCNN shows a decreasing trend. Thus, the smaller elements in the time matrix correspond to pixels with smaller gray values ​​in the image, and the larger elements correspond to pixels with larger gray values ​​in the image.

[0029] Pixels that are not ignited within a limited number of iterations are pixels with very light or no contamination on the insulator disk, while pixels with a higher number of ignitions are pixels with more severe contamination. Therefore, the timing matrix can reflect the ignition feature information of insulator contamination. By setting the ignition threshold in PCNN, secondary contamination information and interference information from the disk surface color itself can be removed, and contamination information can be extracted.

[0030] The insulator monitoring device includes a detection box, inside which an image acquisition unit is installed. The detection box has a through hole on one side. The image acquisition unit includes an electrically telescopic lens assembly, which consists of a camera and an irradiance sensor. The electrically telescopic lens assembly can extend out of the detection box through the through hole. The bottom of the detection box is provided with a clamping mechanism for fixed connection with the tower. The image acquisition unit is connected to a controller.

[0031] A sliding door is slidably installed at the position of the detection box corresponding to the through hole. A driver is provided between the sliding door and the detection box. The controller drives the sliding door to close the through hole or expose the through hole via the driver.

[0032] The driver includes a motor installed inside the detection box, a gear installed on the motor output shaft, and a rack installed on the sliding door, wherein the gear meshes with the rack;

[0033] The sliding door is installed on the side wall inside the detection box; the top and bottom of the insulator monitoring device detection box are provided with guide grooves adapted to the thickness of the sliding door, and the lower end of the sliding door is slidably supported in the guide groove by rollers;

[0034] The through hole is a circular hole with a diameter larger than that of the electric telescopic lens assembly.

[0035] The sliding door structure protects the components inside the detection box, especially the electronic components, from moisture. The detection box is conductive and effectively grounded via a steel tower, which also provides shielding. The controller periodically wakes the system during the day, resulting in effective energy savings.

[0036] The controller can also be connected to a humidity sensor, which will pause data collection when the humidity is too high on rainy days.

[0037] The clamping mechanism includes an upper support plate and a lower support plate, which are connected by four sets of bolts. An adjustment assembly is provided between the upper support plate and the detection box, which is used to adjust the installation height and installation angle of the detection box.

[0038] The adjustment assembly includes four screws, which are arranged in a rectangle between the detection box and the upper support plate. Each screw is fitted with a positioning nut and a locking nut. The upper end of the screw is rotatably connected to the bottom of the detection box, and the upper support plate is sandwiched between the positioning nut and the locking nut.

[0039] The upper support plate is provided with an assembly hole (6a1) at the position corresponding to the screw, and the assembly hole (6a1) is a strip hole.

[0040] Depending on the tower structure and insulator installation location, the detection box can be adjusted in the early stages of installation. After adjusting the data acquisition angle, the detection box's acquisition angle and distance can be locked and fixed.

[0041] The detection box is equipped with an antenna. The antenna is connected to the communication module.

[0042] Significant effect: This invention provides a method for extracting contamination features from insulator images based on pulse-coupled neural networks. It uses image recognition to extract contamination features from the insulator disc image, allowing inspectors to identify the degree of contamination of the insulator. Attached Figure Description

[0043] Figure 1 Grayscale images of insulator disks with three different levels of pollution, standardized grayscale images of disks, and grayscale images reconstructed by PCNN are provided.

[0044] Figure 2 Standardized disk grayscale histograms and PCNN-reconstructed grayscale histograms for insulators with three different levels of pollution;

[0045] Figure 3 This is a flowchart of the method of the present invention;

[0046] Figure 4 This is a schematic diagram of the structure of an online monitoring device for insulator pollution.

[0047] Figure 5 A cross-sectional view of the online monitoring device at the position of the motorized telescopic lens 2;

[0048] Figure 6 A cross-sectional view of the online monitoring device at the positions of ball head 6a2 and ball joint seat 1b;

[0049] Figure 7 A partial schematic diagram illustrating the working principle of drive mechanism 5;

[0050] Figure 8 This is a reference diagram showing the usage status of the monitoring device;

[0051] Figure 9 This is a circuit diagram of the monitoring device. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0053] like Figure 4 , 5 As shown in Figures 6, 7, 8, and 9, the monitoring device is fixed on the iron tower so that the camera is kept at a certain distance from the center of the insulator to be measured, such as 3 to 5 meters, and at a certain angle to the central axis of the insulator, such as 45° to 60°, so as to obtain a clear visible light image of the insulator.

[0054] The detection box is used to protect the entire monitoring device and achieve waterproof and dustproof functions. The electric door is controlled by the controller through the motor. The door opens when monitoring data is needed and closes when monitoring data is not needed.

[0055] An irradiance sensor is mounted next to the camera lens to obtain the intensity of photons reflected from the insulator surface at the same level as captured by the camera; a positioning module, such as GPS or BeiDou, is used to obtain the location of the monitoring device when communicating wirelessly with the control center. This allows for accurate feedback on the pollution accumulation on insulators at a specific geographical location. By deploying sampling points at multiple geographical locations, a monitoring network can be formed to achieve regional coverage.

[0056] The controller is the core of the monitoring device. It has built-in control and data processing programs, which are used to control the opening and closing of the protective cover's electric door, control the acquisition of irradiance sensor data, control the camera to capture visible light images of the insulator and calculate the pollution level of the insulator panel, and control the operation of the positioning module and communication module.

[0057] A communication module, such as GPRS, is used for wireless communication with the cloud to send and receive commands and data. The monitoring device measures surface contamination data of the insulator and feeds back the location information and contamination data to the cloud via the communication module. The communication module is equipped with antenna 7.

[0058] like Figure 2 As shown, a grayscale histogram is generated using the reconstructed grayscale image, which can be used as a dirt feature of the disk surface image.

[0059] like Figures 1-3 As shown, a method for extracting pollution features from insulator images based on pulse-coupled neural networks is described below: Figure 1 A represents light pollution, B represents moderate pollution, and C represents heavy pollution.

[0060] A method for extracting contamination features from insulator images based on pulse-coupled neural networks includes the following steps:

[0061] Step S1: An insulator monitoring device is installed on the tower. The camera of the insulator monitoring device faces the insulator disc surface, and the insulator monitoring device acquires a grayscale image of the insulator disc surface. The grayscale image of the insulator disc surface can be obtained by preprocessing the visible light image of the insulator, including image segmentation, image grayscale conversion, etc.; to obtain the grayscale image.

[0062] Step S2: Standardize the grayscale image of the insulator panel; Standardization of the grayscale image of the insulator panel is to unify the grayscale image of the panel to a set standard size (such as 512*512, etc.) to facilitate subsequent feature extraction.

[0063] Step S3: Use a pulse-coupled neural network to perform ignition processing on the standardized disk grayscale image to obtain the timing matrix;

[0064] Step S4: Normalize the time assignment matrix and map it to the range [0, 255] to reconstruct the grayscale image;

[0065] Step S5: Reduce the size of the reconstructed grayscale image to obtain the contamination feature image of the insulator disk surface.

[0066] The pulse-coupled neural network (PCNN) technique and an improved PCNN model (FLM model, Feature-Linking Model) are used to ignite the grayscale image of the standardized insulator disk. The expression of the FLM model is as follows:

[0067]

[0068]

[0069] U ij (n)=fU ij (n-1)+F ij (n)(1+βL ij (n)) (3);

[0070] Θ ij (n)=gΘ ij (n-1)+V θ Y ij (n-1) (4);

[0071]

[0072] In the above formula, each neuron is represented by (i,j), and one of its neighboring neurons is represented by (k,l), where n is the number of iterations, and F... ij (n) represents the feedback input, Y kl (n-1) represents the postsynaptic action potential, M ijkl and W ijkl S represents the connection weight matrices for the feedback input domain and the coupled connection domain, respectively. ij For external stimulus (i.e., image pixel grayscale values), U ij (n) represents the membrane potential of the neuron, d is the global inhibitory normal, f is the decay time constant of the membrane potential, β is the connection strength, and Θ ij (n) is the threshold of the neuron, g is the threshold decay time constant, and V θ Y is the amplitude adjustment factor. ij (n) represents the pulse output, indicating the firing state of the neuron; L ij (n) is the connection input.

[0073] In a PCNN model, the number of neurons equals the number of pixels in the input image. The output of a neuron mainly depends on the pixel information corresponding to that neuron. Therefore, a timing matrix T can be defined, whose size is equal to the external input S and the output Y, where S represents the input matrix. ij Y represents the elements in the input matrix S; here Y represents the output matrix.ij (n) represents an element in the output matrix Y; and each element in T corresponds one-to-one with a neuron associated with S and Y. T stores the time (or number) information associated with the firing time of each neuron, i.e., T ij It is the firing time (or number of times) of the (i,j)th neuron, satisfying:

[0074]

[0075] The timing matrix accurately records information related to the firing time of each neuron, reflecting the temporal domain information based on image spatial information.

[0076] The time-assignment matrix T is normalized and mapped to [0, 255] using the maximum-minimum normalization method: Where T i,j Let T′ be the element in the i-th row and j-th column of the time-assignment matrix. i,j Let x be the element in the i-th row and j-th column of the normalized matrix. j(min) x is the minimum value of the j-th column of the time-assigned matrix. j(max) The maximum value in the j-th column of the timing matrix is ​​1. The normalized matrix T′ has a maximum value of 1 and a minimum value of 0. Multiplying T′ by 255 maps the data in the matrix to [0, 255] to obtain a reconstructed grayscale image. This grayscale image reflects the main contamination firing features of the image. Through PCNN firing threshold processing, secondary contamination and interference information are removed. The timing matrix truthfully records information related to the firing time of each neuron, including the following cases:

[0077] ① If the neuron has never been ignited, the value at the corresponding point in T is 0;

[0078] ②If the first ignition occurs at time n, then the value of the corresponding point in T is n;

[0079] ③ If the neuron has already been ignited, the value at the corresponding point in T will remain unchanged from the previous value.

[0080] Adjacent pixels in the image are ignited simultaneously through coupling connections (depending on the connection strength β), and this iterative process continues until a set number of iterations is reached, forming the PCNN timing matrix. In expression (4), the postsynaptic action potential Y kl (n-1) makes the threshold Θ ij (n) generates an increment that prevents secondary action potentials from being generated within a certain period, while the increased threshold decays with the time constant g. Therefore, the dynamic threshold function in PCNN exhibits a decreasing trend, resulting in smaller elements in the time-mapping matrix corresponding to pixels with smaller image grayscale values, and larger elements corresponding to pixels with larger image grayscale values.

[0081] Pixels that are not ignited within a limited number of iterations are pixels with very light or no contamination on the insulator disk, while pixels with a higher number of ignitions are pixels with more severe contamination. Therefore, the ignition feature information of insulator contamination can be reflected by the timing matrix. By setting the ignition threshold in PCNN, secondary contamination information and interference information from the disk surface color itself can be removed, and the main contamination information can be extracted.

[0082] Fire threshold settings in PCNN: Fire threshold Θ of neurons ij (n) is composed of the threshold decay time constant g and the amplitude adjustment coefficient V. θ The decision is made, where g is typically set to the range of 0 to 1; V θ It is generally set to the range of 0 to 20.

[0083] By using PCNN to reconstruct grayscale images as feature images, it is possible to reflect the pollution characteristics of the insulator disk surface, such as light pollution, moderate pollution, and heavy pollution. This feature is scale-invariant and rotation-invariant.

[0084] There are many methods for identifying whether a grayscale image reconstructed using PCNN is lightly, moderately, or heavily contaminated.

[0085] Automatic identification can be achieved through classification algorithms, such as Naive Bayes (NBC), Logistic Regression (LR), Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Artificial Neural Network (ANN), and so on.

[0086] Scaling the feature image, such as to 64*64, can reduce the data dimensionality without affecting the pollution characteristics, greatly reducing the data volume and thus improving processing speed. The scaled feature image can then be used for pattern recognition to determine the degree of contamination on the insulator panel surface.

[0087] like Figure 4 , 5 As shown, the insulator monitoring device includes a detection box 1, inside which an image acquisition unit 3 is installed. A through hole 1a is provided on one side of the detection box 1. The image acquisition unit 3 includes an electrically telescopic lens assembly 2, which consists of a camera and an irradiance sensor. The electrically telescopic lens assembly 2 can extend out of the detection box 1 through the through hole 1a to acquire images and irradiance of the insulator 8. A clamping mechanism 6 is provided at the bottom of the detection box 1, which is used to fix it to the support 9 of the tower. The image acquisition unit 3 is connected to a controller.

[0088] The detection box 1 is slidably installed with a sliding door 4 at the position corresponding to the through hole 1a. A driver 5 is provided between the sliding door 4 and the detection box 1. The controller drives the sliding door 4 to close the through hole 1a or expose the through hole 1a via the driver 5.

[0089] like Figure 6 , 7 As shown in Figure 8, the driver 5 includes a motor 5a installed inside the detection box 1, a gear 5b installed on the output shaft of the motor 5a, and a rack 5c installed on the sliding door 4. The gear 5b meshes with the rack 5c. The motor 5a drives the sliding door 4 to move left and right through the gear and rack mechanism.

[0090] The sliding door 4 is installed on the inner side wall of the detection box 1; the top and bottom of the insulator monitoring device detection box 1 are provided with guide grooves 1c adapted to the thickness of the sliding door 4, and the lower end of the sliding door 4 is slidably supported in the guide grooves 1c by rollers 4a.

[0091] The through hole 1a is a circular hole with a diameter larger than that of the electric telescopic lens assembly 2.

[0092] The clamping mechanism 6 includes an upper support plate 6a and a lower support plate 6b. The upper support plate 6a and the lower support plate 6b are connected by four sets of bolts 6d. An adjustment component 6c is provided between the upper support plate 6a and the detection box 1. The adjustment component 6c is used to adjust the installation height and installation angle of the detection box 1. The upper support plate 6a and the lower support plate 6b clamp the bracket 9 in the middle and lock it in place by the four sets of bolts 6d.

[0093] The adjustment component 6c includes four screws 6c1, which are arranged in a rectangle between the detection box 1 and the upper support plate 6a. Each screw 6c1 is fitted with a positioning nut 6c2 and a locking nut 6c3. The upper end of each screw 6c1 is rotatably connected to the bottom of the detection box 1. The upper support plate 6a is sandwiched between the positioning nut 6c2 and the locking nut 6c3. The installation height and installation angle of the detection box 1 can be adjusted by adjusting the position of the positioning nut 6c2 and the locking nut 6c3 on the screw 6c1.

[0094] The upper support plate 6a is provided with an assembly hole 6a1 at the position corresponding to the screw 6c1. The assembly hole 6a1 is a strip-shaped hole.

[0095] The upper support plate 6a has an upwardly protruding ball head 6a2 in the middle, and the lower end of the ball head 6a2 is threadedly connected to the upper support plate 6a. The bottom of the detection box 1 has a ball joint seat 1b that matches the ball head 6a2. The detection box 1 is equipped with an antenna 7. With the above structure, after the four screws 6c1 are loosened, the ball head 6a2 and the ball joint seat 1b can hold the detection box 1 in the angle pre-adjustment position, and then the four screws 6c1 can be locked, which has the technical advantage of convenient adjustment.

[0096] The controller can also correct the grayscale image of the disk surface through gamma transformation. The visible light image of the insulator is acquired by a camera or image sensor, and the irradiation power is acquired by an irradiator or irradiation sensor. Image gamma transformation preprocessing methods include: image segmentation, image grayscale conversion, image size normalization, morphological filtering, etc.; the fitting model of irradiation power w and γ value is used... The expression for gamma correction is: I = I γ (where I is the image grayscale value), therefore the expression for image illumination correction is: The first-order color moment of the corrected grayscale image is set to x, where x = 0.4. Gamma correction of the grayscale image can adjust the brightness of the image. For visible light images of insulators taken at the same imaging distance and angle, adjusting the γ value with the first-order color moment of the grayscale image as the standard can unify the acquired insulator disc images to the same brightness range, greatly reducing image processing errors caused by differences in illumination.

[0097] The gamma transformation correction of the grayscale image on the disk can be performed by the controller, and the gamma-transformed grayscale image is sent to the cloud via antenna 7 of the communication module.

[0098] The gamma transform correction of the disk grayscale image can also be performed on a cloud server. The controller can send the disk image 'a' and the irradiation power 'w' to the cloud via the communication module.

[0099] Maintenance personnel can determine the degree of insulator contamination based on the corrected grayscale image and decide whether cleaning and maintenance are necessary.

[0100] Performing pattern recognition on the processed feature images to determine the degree of contamination is a mature technology, and can be done in the following ways:

[0101] Automatic identification is achieved using the Support Vector Machine (SVM) linear classifier algorithm. First, the SVM recognition model is trained using feature images of a certain number of insulator samples. The pollution level is manually marked based on the salt density or ash density value of the insulator surface. After the SVM model is trained, the feature image of the sample to be tested is directly extracted and input into the model to obtain the pollution level of the sample.

[0102] The theoretical basis of the SVM linear classifier algorithm is as follows:

[0103] The first case is that the dataset is linearly separable. Assume a linearly separable dataset is: X = {(x1, y1), (x...} i ,y i ),...,(x n ,y n )}(1≤i≤n), where y iFor each ∈{-1,+1}, the corresponding data points can be classified using the following formula:

[0104]

[0105] In the above formula, w and b are hyperplane parameters, w is the weight vector, and b is the offset.

[0106] During the optimization process, the SVM classifier locates the decision boundary with the maximum margin among all possible hyperplanes. M is defined as the sum of the distances from the hyperplane to the points closest to the hyperplane in both classes. To maximize this distance, the following constraints should be satisfied:

[0107]

[0108] In the above formula, ||w|| is the Euclidean norm of w.

[0109] Introducing positive Lagrange multipliers α i (1≤i≤n) Solving the above problem, we have:

[0110]

[0111] The Karush-Kuhn-Tucker (KKT) condition in the above equation can be expressed as:

[0112]

[0113] Therefore, equation (3) can be transformed into its dual expression:

[0114]

[0115] Solving for the optimal hyperplane in the above equation yields the decision function:

[0116]

[0117] In the above formula, n sv It is the number of support vectors.

[0118] In the second case, the dataset is non-linearly separable. When using a linear classifier to process non-linearly separable datasets, a regularization parameter C is added to balance the maximum margin decision boundary with the minimum number of misclassifications, while a slack variable ξ is introduced. i Change the constraints to:

[0119]

[0120] Solving the above problem using the Lagrange method yields the following decision function:

[0121]

[0122] Finally, it should be noted that the above are only specific embodiments of the present invention. Of course, those skilled in the art can make modifications and variations to the present invention. If these modifications and variations fall within the scope of the claims of the present invention and their equivalents, they should be considered as being within the protection scope of the present invention.

Claims

1. A method for extracting pollution features from insulator images based on pulse-coupled neural networks, characterized in that, Includes the following steps: Step S1: Install an insulator monitoring device on the iron tower. The camera of the insulator monitoring device is facing the insulator disc, and the insulator monitoring device acquires a grayscale image of the insulator disc. Step S2: Standardize the grayscale image of the insulator disc surface; Step S3: Use a pulse-coupled neural network to perform ignition processing on the standardized disk grayscale image to obtain the timing matrix; In step S3, the Pulse Coupled Neural Network (PCNN) technique is used, and the standardized insulator disk grayscale image is ignited based on the FLM model. The expression of the FLM model is as follows: U ij (n)=fU ij (n-1)+F ij (n)(1+βL ij (n)) (3); Θ ij (n)=gΘ ij (n−1)+V θ Y ij (n-1) (4)? In the above formula, each neuron is represented by (i,j), and one of its neighboring neurons is represented by (k,l), where n is the number of iterations, and F... ij (n) represents the feedback input, Y kl (n-1) represents the postsynaptic action potential, M ijkl and W ijkl S represents the connection weight matrices for the feedback input domain and the coupled connection domain, respectively. ij As an external incentive, S ij That is, the grayscale value of an image pixel, U ij (n) represents the membrane potential of the neuron, d is the global inhibitory normal, f is the decay time constant of the membrane potential, β is the connection strength, and Θ ij (n) is the threshold of the neuron, g is the threshold decay time constant, and V θ Y is the amplitude adjustment factor. ij (n) represents the pulse output, Y ij (n) represents the firing state of the neuron; L ij (n) represents the input connection; In a PCNN model, the number of neurons equals the number of pixels in the input image. The output of a neuron depends on the pixel information corresponding to that neuron. A time-assignment matrix T is defined, whose size is equal to the external input S and the output Y, where S represents the input matrix. ij Y represents the elements in the input matrix S; Y represents the output matrix. ij (n) represents an element in the output matrix Y; and each element in T corresponds one-to-one with a neuron associated with S and Y. T stores the time or frequency information associated with the firing time of each neuron, i.e., T ij It is the firing time or number of times of the (i,j)th neuron, satisfying: The timing matrix accurately records information related to the firing time of each neuron, reflecting the temporal domain information based on image spatial information; Step S4: Normalize the time assignment matrix and map it to the range [0, 255] to reconstruct the grayscale image; Step S5: Reduce the size of the reconstructed grayscale image to obtain the contamination feature image of the insulator disk surface.

2. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 1, characterized in that: In step S1, the grayscale image of the insulator disk is obtained by preprocessing the visible light image of the insulator; the preprocessing includes image segmentation and image grayscale conversion.

3. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 1, characterized in that: In step S2, the grayscale image standardization process of the insulator disk surface is to unify the grayscale image of the disk surface to a set standard size, which facilitates subsequent feature extraction.

4. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 1, characterized in that: In step S4, the time assignment matrix T is normalized and mapped to [0, 255] using the maximum-minimum normalization method: Where T i,j Let T′ be the element in the i-th row and j-th column of the time-assigned matrix. i,j Let x be the element in the i-th row and j-th column of the normalized matrix. j(min) x is the minimum value of the j-th column of the time-assigned matrix. j(max) The maximum value in the j-th column of the time-assignment matrix is ​​1. The maximum value in the normalized matrix T′ is 1, and the minimum value is 0. Multiplying T′ by 255 maps the data in the matrix to [0,255] to obtain a reconstructed grayscale image. This grayscale image reflects the main contamination firing feature information of the image. Through PCNN firing threshold processing, secondary contamination information and interference information are removed. The time-assignment matrix truly records the information related to the firing time of each neuron, which are respectively from the following cases: ① If the neuron has never been ignited, the value at the corresponding point in T is 0; ②If the first ignition occurs at time n, then the value of the corresponding point in T is n; ③ If the neuron has already been ignited, the value at the corresponding point in T will remain unchanged from the previous value; Adjacent pixels in the image are ignited at the same time through coupling connections, and the iterative process continues until the set number of iterations is reached, forming the PCNN timing matrix.

5. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 1, characterized in that: The insulator monitoring device includes a detection box (1), an image acquisition unit (3) is installed inside the detection box (1), a through hole (1a) is provided on one side of the detection box (1), the image acquisition unit (3) includes an electric telescopic lens assembly (2), the electric telescopic lens assembly (2) is composed of a camera and an irradiance sensor, the electric telescopic lens assembly (2) can extend out of the detection box (1) through the through hole (1a); a clamping mechanism (6) is provided at the bottom of the detection box (1), the clamping mechanism (6) is used to fix it to the iron tower; the image acquisition unit (3) is connected to a controller.

6. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 5, characterized in that: The detection box (1) is slidably installed with a sliding door (4) at the position corresponding to the through hole (1a). A driver (5) is provided between the sliding door (4) and the detection box (1). The controller drives the sliding door (4) to close the through hole (1a) or expose the through hole (1a) via the driver (5). The driver (5) includes a motor (5a) installed inside the detection box (1), a gear (5b) installed on the output shaft of the motor (5a), and a rack (5c) installed on the sliding door (4), wherein the gear (5b) meshes with the rack (5c); The sliding door (4) is installed on the side wall inside the detection box (1); the top and bottom of the insulator monitoring device detection box (1) are provided with guide grooves (1c) adapted to the thickness of the sliding door (4), and the lower end of the sliding door (4) is slidably supported in the guide groove (1c) by rollers (4a); The through hole (1a) is a round hole with a diameter larger than that of the electric telescopic lens assembly (2).

7. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 5, characterized in that: The clamping mechanism (6) includes an upper support plate (6a) and a lower support plate (6b). The upper support plate (6a) and the lower support plate (6b) are connected by four sets of bolts (6d). An adjustment component (6c) is provided between the upper support plate (6a) and the detection box (1). The adjustment component (6c) is used to adjust the installation height and installation angle of the detection box (1). The adjustment assembly (6c) includes four screws (6c1), which are arranged in a rectangle between the detection box (1) and the upper support plate (6a). Each screw (6c1) is fitted with a positioning nut (6c2) and a locking nut (6c3). The upper end of the screw (6c1) is rotatably connected to the bottom of the detection box (1), and the upper support plate (6a) is sandwiched between the positioning nut (6c2) and the locking nut (6c3). The upper support plate (6a) has an assembly hole (6a1) at the position corresponding to the screw (6c1), and the assembly hole (6a1) is a strip hole.

8. The method for extracting pollution features from insulator images based on pulse-coupled neural networks according to claim 5, characterized in that: The detection box (1) is equipped with an antenna (7).

Images