Multispectral-based insulator smudginess detection and evaluation method

Abstract

The invention discloses a multispectral-based insulator smudginess detection and evaluation method, which comprises the following steps of: firstly, performing multi-angle image acquisition on an insulator by using an imaging system carrying a visible light camera and an infrared camera, and completing preprocessing such as image denoising, enhancement, registration and radiometric calibration; secondly, accurately segmenting an insulator region from the preprocessed visible light image by using a Sobel algorithm and a watershed algorithm, and extracting spectral line features; carrying out spectral line feature screening and fusion by adopting an iterative information variable preserving algorithm, carrying out secondary fusion and dimensionality reduction through kernel principal component analysis in combination with features such as colors and textures, and constructing a fusion feature set for identifying smudginess types; and finally, combining the infrared texture features to construct an integrated learning model, and outputting the smudginess degree.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, specifically to a method for detecting and evaluating contamination in insulators based on multispectral imaging. Background Technology

[0002] Insulators are crucial insulating devices in power transmission lines, and their operational status directly affects the reliability and safety of the power grid. Due to long-term exposure to the natural environment, insulators are susceptible to various defects such as contamination, aging, and cracks, which can lead to decreased insulation performance or even breakdown, causing power outages. Therefore, regular inspection and condition assessment of insulators are of great significance.

[0003] Currently, common methods for insulator inspection include manual inspection and visible light imaging. For example, in "A Method for Insulator Fault Detection Based on UAV Inspection," images of insulators are captured by a visible light camera mounted on a UAV, and image processing algorithms are used to identify anomalies. However, this method is greatly affected by lighting conditions and is insensitive to early insulation defects and internal degradation, making accurate assessment difficult. Another method, such as "An Insulator Detection System Based on Infrared Thermography," detects faults by detecting abnormal insulator temperatures, but it is easily affected by fluctuations in ambient temperature and load current, and cannot effectively identify surface contamination levels and material aging. In recent years, multispectral imaging technology has developed rapidly, enabling the simultaneous acquisition of image information of targets in different wavelength bands, thereby more comprehensively characterizing the physical and chemical state of the inspected object and providing a new technical approach for accurate insulator condition assessment. Therefore, how to combine multispectral technology to achieve comprehensive insulator detection and condition assessment has become a technical problem that needs to be solved. Summary of the Invention

[0004] To address the shortcomings of existing insulator detection methods in terms of defect identification accuracy, early fault warning, and comprehensive condition assessment, this invention provides a multispectral-based method for insulator contamination detection and assessment. This method enables the collection and fusion analysis of multi-dimensional condition information of insulators, thereby improving the accuracy and reliability of detection.

[0005] A method for detecting and assessing contamination in insulators based on multispectral analysis includes the following steps: Step S1: Use a multispectral imaging system equipped with a visible light camera and an infrared camera to acquire images of the target insulator, obtaining visible light images and infrared images; Step S2: Preprocess the acquired multispectral images, including image denoising, enhancement and registration, as well as radiometric calibration before spectral line extraction; Step S3: The insulator region is segmented using an image segmentation algorithm, and spectral features are extracted simultaneously from the visible light and infrared segmented regions. Then, the spectral features are fused using the Iterative Information Retaining Variable (IRIV) algorithm. Step S4: The spectral features after IRIV processing are fused with the image (color, texture) features, and then fused again by kernel principal component analysis (KPCA) to provide a fusion feature basis for dirt type identification; Step S5: Combine infrared texture features to construct an ensemble learning model and output the degree of dirtiness.

[0006] Furthermore, step S1 includes: Step S1-1: Operate the multispectral imaging system equipped with a visible light camera and an infrared thermal imager to acquire images of the target insulator from different angles and distances; Steps S1-2: Acquire and store the visible light images of the target insulator respectively. and infrared images .

[0007] Furthermore, step S2 includes: Step S2-1: To remove Gaussian noise from the original visible light image and infrared image obtained in step S1-2, the visible light image... Infrared images Perform Gaussian filtering to obtain the filtered image:

[0008] in, , These are the coordinates of pixels in the visible light image and the infrared image, respectively. , These represent the pixels in a visible light image. and infrared image pixels and pixels in infrared images Pixel values; , These represent the Gaussian templates used for Gaussian filtering of visible light and infrared images, respectively. This indicates a convolution operation.

[0009] Step S2-2: Perform histogram equalization on the denoised visible light and infrared images from step S2-1 to enhance the contrast of each image and improve image quality.

[0010] The histogram equalization transformation function used is defined as follows:

[0011] Where N is the total number of pixels in the image; L represents the number of pixels at the current gray level; L represents the total number of gray levels in the image. Represents the cumulative distribution function; It is the gray level of the k-th level in the original image. This maps the cumulative probability to a new gray level; Round indicates rounding.

[0012] By mapping the pixels of the original image one-to-one using the cumulative distribution function according to the aforementioned histogram equalization formula, a result image with stronger contrast can be obtained:

[0013] in, , These represent the pixel values ​​of each pixel in the visible light image and the infrared image after histogram equalization, respectively.

[0014] Step S2-3: Using the visible light image as a reference, perform feature point matching-based image registration between the enhanced infrared image from step S2-2 and the reference image to obtain the transformation matrix from the visible light image coordinate system to the infrared image coordinate system. This ensures pixel-level alignment of the same insulator region in two different spectral images.

[0015] Step S2-4: Before spectral line extraction, to eliminate the influence of instrument response, ambient light, and other factors on the multispectral image, the preprocessed image needs to be radiometrically calibrated, including DN correction and black-and-white correction, converting the image grayscale values ​​into reflectance data that can be used for subsequent analysis. The specific operation is as follows: Because changes in ambient light during hyperspectral acquisition can alter the image's DN value (a unitless integer value representing the brightness of the imaging pixels), DN correction is necessary. The correction formula is as follows:

[0016] in, These are the hyperspectral images of each sample after DN correction. These are the DN values ​​for each band of the original hyperspectral image before correction. It corrects the DN values ​​of each band in images at different depths using a whiteboard (placed vertically). It is the DN value of each band of the white-corrected image of the whiteboard, where i is 256 bands.

[0017] After DN correction, the image needs to be calibrated as a reflectance image to identify the spectral characteristics of the substance, and then black and white correction is performed. The correction formula is as follows:

[0018] In the formula, T represents the hyperspectral image of each sample after black and white correction. Indicates a white-corrected image. This represents a black-corrected image.

[0019] Furthermore, step S3 includes: Step S3-1: Extract the preprocessed visible light image The gradient information in the image is used to obtain the preliminary edge contour of the insulator. The edge contour extraction of the insulator uses the Sobel operator to calculate the image gradient. Its core formula includes the convolution calculation of the horizontal and vertical gradients:

[0020] in, The gradient is the horizontal gradient, and the vertical gradient is used to detect edges in the image. is the vertical gradient, used to detect horizontal edges in an image; * represents the convolution operation. Gradient magnitude is an indicator that measures the edge strength of each pixel in an image; the larger the magnitude, the stronger the edge strength.

[0021] By setting a threshold to binarize the gradient magnitude map, a continuous preliminary edge profile of the insulator can be obtained.

[0022] Step S3-2: Based on the edge gradient information obtained in step S3-1, the watershed algorithm is used to segment the insulator region in the visible light image; the edge gradient image is regarded as a topographic map, and the image is segmented into non-overlapping regions by simulating the flooding process, thereby accurately separating the insulator from the background.

[0023] Step S3-3: Using the insulator region segmented from the visible light image in step S3-2 as a mask, and utilizing the registration relationship established in step S2, it is mapped to the infrared image in the same coordinate system, thereby accurately obtaining the corresponding region image of the insulator in the infrared spectrum. This process achieves coordinate mapping between images through affine transformation, and the core formula is:

[0024] in, These are the coordinates of the insulator pixels in the visible light image; It is the transformation matrix from the visible light image coordinate system to the infrared image coordinate system; The coordinates of the insulator pixels in the visible light image The corresponding coordinates in the infrared coordinate system. This transformation accurately projects the insulator outline from the visible light image onto the infrared image, thereby extracting the corresponding region and achieving location-based feature association.

[0025] Step S3-4: For the segmented regions of the preprocessed and radiometrically calibrated visible and infrared images, extract the grayscale values ​​(visible band) or radiance values ​​(infrared band) of each pixel in different wavelengths, and construct the spectral curve of each pixel as the spectral feature of the region. For pixels within the segmented region of the visible image... Its spectral characteristics are expressed as ,in For visible light bands; for pixels within the segmentation area of ​​an infrared image. Its spectral characteristics are expressed as ,in The infrared band is used as the starting point. These spectral features are then integrated to form an initial multispectral feature set. Based on the IRIV algorithm, the information contribution of each spectral feature variable is iteratively calculated, along with the correlation between each spectral feature variable and the degree of contamination, as well as the redundancy between variables. Based on a preset threshold, the optimal subset of spectral features is gradually selected, eliminating redundant and irrelevant spectral features, and retaining P optimal spectral features. This enables the effective fusion of multispectral line features, resulting in fused spectral line features. .

[0026] Furthermore, in step S3-2, a watershed algorithm is proposed to segment the insulator skirt region. First, to effectively highlight the insulator region and suppress noise, the visible light image is... Convert from RGB space to HSV space to obtain the S-component image. The calculation process is as follows:

[0027] in, Indicates color space operations; M is the preprocessed visible light image; M is the mask image used for morphological reconstruction.

[0028] For S-component images Gradient calculation is performed to obtain the gradient image. :

[0029] in, This indicates the output gradient image in coordinates. The gradient magnitude of a pixel; the larger the value, the stronger the edge intensity of that pixel. For S-component image On point Pixel values; Representing an image The first-order partial derivative in the x-direction (horizontal direction) reflects the rate of change of the gray value at that point in the horizontal direction; Representing an image The first-order partial derivative in the y-direction (vertical direction) reflects the rate of change of the gray value at that point in the vertical direction.

[0030] Next, the gradient image is detected. All local minima are identified, and these points are labeled as distinct "basins". The seed. Simulate a seed from arrive The rising water level h, The minimum gradient magnitude in the gradient image. Let be the maximum gradient magnitude in the gradient image. For each water level h, define the set of regions submerged below water level h as . :

[0031] As the water level h increases, The area will continue to expand. When originating from two different basins... and flooded areas and When the merger is about to take place, that is, at a certain water level When they first meet, they satisfy the following formula:

[0032] To prevent the basins from merging, a watershed needs to be constructed at their intersection, eventually forming a closed boundary. Mathematically, this watershed is a set of points that satisfy the following conditions. :

[0033] Finally, when the water level reaches the maximum gradient value of the entire image... At that time, all the basins were filled with water, and the collection of watersheds was constructed. The image was then segmented into non-overlapping regions. :

[0034] in, Represents all regions after segmentation arrive The union of the original image regions is the entire original image region. ; This indicates any two distinct regions. and There can be no overlap between them, and their intersection is an empty set. In this process, the edges of the insulator (high gradient ridges) precisely form these watersheds, thus accurately separating the insulator from the background into different regions.

[0035] Furthermore, step S4 includes: Step S4-1: Based on the insulator region extracted in step S3, calculate the color features and texture features from the corresponding visible light image, and perform preliminary fusion with the IRIV-optimized multispectral spectral features output in step S3 to obtain a preliminary fused feature vector.

[0036] Step S4-2: The preliminary fusion feature vector obtained in step S4-1 is fused a second time using KPCA (Kernel Principal Component Analysis) to obtain a dimensionality-reduced and optimized core feature vector.

[0037] Furthermore, in step S4-1, during the feature extraction process of the insulator region, the visible light image is first extracted. Given the color features, calculate its color feature vector: ; in, Represents a specific color value; This represents the total number of color values. This indicates the proportion of a color value to all color values; It represents the total number of all possible non-repeating color values ​​in the entire image (or insulator region); It is made by all The vectors formed together describe the statistical distribution of color throughout the insulator region.

[0038] Next, texture features of the visible light image are extracted, and gray-level gradient co-occurrence matrix (GGCM) extraction and analysis are performed. Laplacian transform is then applied to the gray-level images of each region to obtain the gray-level gradient maps for each region.

[0039] in, This indicates the grayscale gradient of the contaminated area at location. Pixel value at; Indicates the input image In position The pixel value.

[0040] Then, each gradient map is discretized at the gray level to obtain the gray-level gradient co-occurrence matrix (GGCM) of each contaminated region. , For set The number of elements in the middle. Normalization is performed to obtain the normalized matrix. :

[0041] in for The number of gray levels, for The number of gray levels.

[0042] Based on the above gray-gradient co-occurrence matrix, the texture features of each dirty area can be quantified. Four texture features with significant texture feature changes are selected: gray-level distribution non-uniformity (Gray Asymmetry), gradient distribution non-uniformity (Grads Asymmetry), gray-level average (Gray Mean), and inertia (Inertia).

[0043] Gray-level distribution non-uniformity mainly reflects the degree of unevenness in gray-level changes among pixels in a dirty image. It is used to quantify the degree of dirtiness represented by the dirty area image. The calculation formula is as follows:

[0044] Gradient distribution non-uniformity mainly reflects the degree of unevenness in the contour and edge distribution between pixels and their neighbors in a dirty image. It is used to quantify the edge changes of different degrees of dirtiness in a dirty region image. The calculation formula is as follows:

[0045] Gray-scale average primarily reflects the overall brightness of a region in an image; a higher value indicates greater brightness. It is used to quantify the total amount of dirt represented by a dirty region in an image. The calculation formula is as follows:

[0046] Inertia is a geometric parameter describing an image, and its calculation formula is shown below:

[0047] The above steps are used to fuse gray-level distribution non-uniformity, gradient distribution non-uniformity, gray-level averaging, and inertia to construct a texture feature vector Q:

[0048] The above processing is performed on the visible light image, that is, let Substituting into formula (15-21), we obtain the visible light image texture feature vector. .

[0049] Next, the three types of feature vectors describing dirt from different dimensions—color features, texture features, and IRIV spectral line features—are directly concatenated and integrated into a high-dimensional preliminary fusion feature vector. .

[0050] Furthermore, in step S4-2, kernel principal component analysis (KPCA) is used for secondary fusion and dimensionality reduction. KPCA introduces kernel tricks to nonlinearly map the original feature space to a higher-dimensional feature space, where principal component analysis is performed to effectively extract the nonlinear feature structure. Its basic process is as follows: The initial fused feature vectors are collected to form a feature matrix. Assuming there are t samples, the initial fused feature vector of the b-th sample is denoted as... (Right now The entire feature matrix is ​​represented as:

[0051] The feature matrix X is mapped through a nonlinear mapping. Map from the original space to a high-dimensional feature space F. Calculate the covariance matrix in this space:

[0052] in To initially fuse feature vectors Mapping in high-dimensional space.

[0053] By solving the eigenvalue problem And introduce kernel functions The problem is transformed into finding the eigenvalues ​​and eigenvectors of the kernel matrix H, where The formula for solving the eigenvalue problem is:

[0054] Where C represents the covariance matrix of the sample in the high-dimensional feature space; v represents the eigenvalues ​​of the covariance matrix C; v represents the eigenvalues ​​of the covariance matrix C. eigenvectors.

[0055] The polynomial kernel function is selected as follows:

[0056] Where c is a constant term used to adjust the balance between higher and lower degree terms. In this invention, c=1 is used; d is the polynomial order. In this paper, d=3 is used to enhance the separability of samples in high-dimensional space while maintaining numerical stability.

[0057] In the KPCA process, principal components whose cumulative contribution rate reaches a set threshold are selected as the final fusion features, thereby achieving data dimensionality reduction while retaining most of the effective information. The final fusion feature has an e-dimensional dimension (e≪d), denoted as:

[0058] in This represents the final fusion feature set after kernel principal component analysis; This represents the selected principal components, each of which... The corresponding feature components are representative features in the high-dimensional feature space; e represents the dimension of the final fused feature.

[0059] This fusion feature set integrates key information from three types of features: color, texture, and spectrum. It has good discriminative power and compactness, making it suitable for subsequent tasks of identifying and classifying the degree of insulator contamination.

[0060] Furthermore, step S5 includes: Step S5-1: The "spectral features + image color features" obtained by secondary fusion of Kernel Principal Component Analysis (KPCA) in Step S4 are integrated with the extracted infrared texture features to form a high-dimensional input matrix.

[0061] The infrared texture features are obtained from step 4-1, let Substituting into formula (15-21), we obtain the visible light image texture feature vector. ,Will and By concatenating the features, a high-dimensional input feature vector can be formed. .

[0062] Step S5-2: Using the integrated high-dimensional input feature vector as input, construct a Radial Basis Neural Network (RBFNN) as the base classifier for ensemble learning. The discriminative latent function of the RBFNN is:

[0063] in, These are the hidden layer weight coefficients; The error coefficient; The center of the i-th neuron; is the extension constant of the radial basis function; the number of neurons in the input layer is equal to the dimension d of the high-dimensional feature vector, the number of neurons in the hidden layer is initialized (k=12), and the output layer has 4 neurons (corresponding to the 4 levels of dirtiness: "mild, moderate, moderate, and severe").

[0064] To enable RBFNN to adapt to high-dimensional input features The distribution of dirt is calculated based on the "width criterion". The process is as follows: First, determine the neuron center. The formulas for the spectral lines of each type of contamination level, and the matrix D, are as follows:

[0065] in Represents the center of the i-th neuron With the center of the j-th neuron The Euclidean distance between them.

[0066] The nearest distance is calculated based on matrix D, using the following formula:

[0067] Calculate the distance from each input sample data point to the center of the neuron, and construct a distance array R with coefficient L, as shown in the following formula:

[0068] Finally, the extended constant is defined. The formula is as follows:

[0069] Where n is the total number of samples, each RBFNN classifier can be built through the above steps, and the expansion constant is determined based on the width criterion. This can help capture spectral line distribution features and activate hidden layer neurons to varying degrees. 500 RBFNN classifiers are combined with modules to achieve ensemble learning and output the final judgment result, constructing an insulator pollution level detection model. The 500 RBFNN classifiers are generated using a Bagging ensemble learning method (random bootstrapping with replacement).

[0070] Compared with the prior art, the present invention has the following significant advantages: (1) This invention proposes a multispectral imaging technology that combines visible light and infrared light. The Sobel algorithm and watershed algorithm are used to accurately segment the insulator region, and the IRIV algorithm is used to screen and fuse spectral features. This solves the problem that traditional detection methods are greatly affected by light and environmental interference and are not accurate in identifying the dirt features of insulators. (2) The present invention uses the KPCA algorithm to achieve secondary fusion and dimensionality reduction of multi-dimensional features, and combines the integrated RBFNN model to output the dirt level, thereby realizing non-contact detection and solving the problems of low detection efficiency and ambiguous dirt level assessment of existing methods. Attached Figure Description

[0071] Figure 1 This is an overall flowchart of a multispectral-based insulator contamination detection and evaluation method proposed in this invention; Figure 2 This is a flowchart of the image preprocessing of the acquired multispectral data according to the present invention; Figure 3 This is a schematic diagram illustrating the watershed algorithm used in this invention to segment insulator regions from visible light images; Figure 4 This is a flowchart of the multispectral feature fusion and dirt level assessment model of the present invention. Detailed Implementation

[0072] The following detailed description, in conjunction with the accompanying drawings, of a specific implementation method for detecting and evaluating contamination in insulators based on multispectral methods according to the present invention.

[0073] The present invention adopts the following technical solution, combined with Figure 1 To complete the detection and evaluation, this invention proposes a multispectral-based method for detecting and evaluating insulator contamination, comprising the following steps: Step S1: Use a multispectral imaging system equipped with a visible light camera and an infrared camera to acquire images of the target insulator, obtaining visible light images and infrared images; Step S2: Preprocess the acquired multispectral images, including image denoising, enhancement and registration, as well as radiometric calibration before spectral line extraction; Step S3: The insulator region is segmented using an image segmentation algorithm, and spectral features are extracted simultaneously from the visible light and infrared segmented regions. Then, the spectral features are fused using the Iterative Information Retaining Variable (IRIV) algorithm. Step S4: The spectral features after IRIV processing are fused with the image (color, texture) features, and then fused again by kernel principal component analysis (KPCA) to provide a fusion feature basis for dirt type identification; Step S5: Combine infrared texture features to construct an ensemble learning model and output the degree of dirtiness.

[0074] Preferably, step S1 includes: Step S1-1: Operate the multispectral imaging system equipped with a visible light camera and an infrared thermal imager to acquire images of the target insulator from different angles and distances; Steps S1-2: Acquire and store the visible light images of the target insulator respectively. and infrared images .

[0075] Combination Figure 2 Preferably, step S2 includes: Step S2-1: To remove Gaussian noise from the original visible light image and infrared image obtained in step S1-2, the visible light image... Infrared images Perform Gaussian filtering to obtain the filtered image:

[0076] in, , These are the coordinates of pixels in the visible light image and the infrared image, respectively. , These represent the pixels in a visible light image. and infrared image pixels Pixel values; , These represent the pixels in the visible light image after Gaussian filtering. and pixels in infrared images Pixel values; , These represent the Gaussian templates used for Gaussian filtering of visible light and infrared images, respectively. This indicates a convolution operation.

[0077] Step S2-2: Perform histogram equalization on the denoised visible light and infrared images from step S2-1 to enhance the contrast of each image and improve image quality.

[0078] The histogram equalization transformation function used is defined as follows:

[0079] Where N is the total number of pixels in the image; L represents the number of pixels at the current gray level; L represents the total number of gray levels in the image. Represents the cumulative distribution function; It is the gray level of the k-th level in the original image. This maps the cumulative probability to a new gray level; Round indicates rounding.

[0080] By mapping the pixels of the original image one-to-one using the cumulative distribution function according to the aforementioned histogram equalization formula, a result image with stronger contrast can be obtained:

[0081] in, , These represent the pixel values ​​of each pixel in the visible light image and the infrared image after histogram equalization, respectively.

[0082] Step S2-3: Using the visible light image as a reference, perform feature point matching-based image registration between the enhanced infrared image from step S2-2 and the reference image to obtain the transformation matrix from the visible light image coordinate system to the infrared image coordinate system. This ensures pixel-level alignment of the same insulator region in two different spectral images.

[0083] Step S2-4: Before spectral line extraction, to eliminate the influence of instrument response, ambient light, and other factors on the multispectral image, the preprocessed image needs to be radiometrically calibrated, including DN correction and black-and-white correction, converting the image grayscale values ​​into reflectance data that can be used for subsequent analysis. The specific operation is as follows: Because changes in ambient light during hyperspectral acquisition can alter the image's DN value (a unitless integer value representing the brightness of the imaging pixels), DN correction is necessary. The correction formula is as follows:

[0084] in, These are the hyperspectral images of each sample after DN correction. These are the DN values ​​for each band of the original hyperspectral image before correction. It corrects the DN values ​​of each band in images at different depths using a whiteboard (placed vertically). It is the DN value of each band of the white-corrected image of the whiteboard, where i is 256 bands.

[0085] After DN correction, the image needs to be calibrated as a reflectance image to identify the spectral characteristics of the substance, and then black and white correction is performed. The correction formula is as follows:

[0086] In the formula, T represents the hyperspectral image of each sample after black and white correction. Indicates a white-corrected image. This represents a black-corrected image.

[0087] Combination Figure 3 Preferably, step S3 includes: Step S3-1: Extract the preprocessed visible light image The gradient information in the image is used to obtain the preliminary edge contour of the insulator. The edge contour extraction of the insulator uses the Sobel operator to calculate the image gradient. Its core formula includes the convolution calculation of the horizontal and vertical gradients:

[0088] in, The gradient is the horizontal gradient, and the vertical gradient is used to detect edges in the image. is the vertical gradient, used to detect horizontal edges in an image; * represents the convolution operation. Gradient magnitude is an indicator that measures the edge strength of each pixel in an image; the larger the magnitude, the stronger the edge strength.

[0089] By setting a threshold to binarize the gradient magnitude map, a continuous preliminary edge profile of the insulator can be obtained.

[0090] Step S3-2: Based on the edge gradient information obtained in step S3-1, the watershed algorithm is used to segment the insulator region in the visible light image; the edge gradient image is regarded as a topographic map, and the image is segmented into non-overlapping regions by simulating the flooding process, thereby accurately separating the insulator from the background.

[0091] Step S3-3: Using the insulator region segmented from the visible light image in step S3-2 as a mask, and utilizing the registration relationship established in step S2, it is mapped to the infrared image in the same coordinate system, thereby accurately obtaining the corresponding region image of the insulator in the infrared spectrum. This process achieves coordinate mapping between images through affine transformation, and the core formula is:

[0092] in, These are the coordinates of the insulator pixels in the visible light image; It is the transformation matrix from the visible light image coordinate system to the infrared image coordinate system; The coordinates of the insulator pixels in the visible light image The corresponding coordinates in the infrared coordinate system. This transformation accurately projects the insulator outline from the visible light image onto the infrared image, thereby extracting the corresponding region and achieving location-based feature association.

[0093] Step S3-4: For the segmented regions of the preprocessed and radiometrically calibrated visible and infrared images, extract the grayscale values ​​(visible band) or radiance values ​​(infrared band) of each pixel in different wavelengths, and construct the spectral curve of each pixel as the spectral feature of the region. For pixels within the segmented region of the visible image... Its spectral characteristics are expressed as ,in For visible light bands; for pixels within the segmentation area of ​​an infrared image. Its spectral characteristics are expressed as ,in The infrared band is used as the starting point. These spectral features are then integrated to form an initial multispectral feature set. Based on the IRIV algorithm, the information contribution of each spectral feature variable is iteratively calculated, along with the correlation between each spectral feature variable and the degree of contamination, as well as the redundancy between variables. Based on a preset threshold, the optimal subset of spectral features is gradually selected, eliminating redundant and irrelevant spectral features, and retaining P optimal spectral features. This enables the effective fusion of multispectral line features, resulting in fused spectral line features. .

[0094] Preferably, in step S3-2, a watershed algorithm is proposed to segment the insulator skirt region. First, to effectively highlight the insulator region and suppress noise, the visible light image is... Convert from RGB space to HSV space to obtain the S-component image. The calculation process is as follows:

[0095] in, Indicates color space operations; M is the preprocessed visible light image; M is the mask image used for morphological reconstruction.

[0096] For S-component images Gradient calculation is performed to obtain the gradient image. :

[0097] in, This indicates the output gradient image in coordinates. The gradient magnitude of a pixel; the larger the value, the stronger the edge intensity of that pixel. For S-component image On point Pixel values; Representing an image The first-order partial derivative in the x-direction (horizontal direction) reflects the rate of change of the gray value at that point in the horizontal direction; Representing an image The first-order partial derivative in the y-direction (vertical direction) reflects the rate of change of the gray value at that point in the vertical direction.

[0098] Next, the gradient image is detected. All local minima are identified, and these points are labeled as distinct "basins". The seed. Simulate a seed from arrive The rising water level h, The minimum gradient magnitude in the gradient image. Let be the maximum gradient magnitude in the gradient image. For each water level h, define the set of regions submerged below water level h as . :

[0099] As the water level h increases, The area will continue to expand. When originating from two different basins... and flooded areas and When the merger is about to take place, that is, at a certain water level When they first meet, they satisfy the following formula:

[0100] To prevent the basins from merging, a watershed needs to be constructed at their intersection, eventually forming a closed boundary. Mathematically, this watershed is a set of points that satisfy the following conditions. :

[0101] Finally, when the water level reaches the maximum gradient value of the entire image... At that time, all the basins were filled with water, and the collection of watersheds was constructed. The image was then segmented into non-overlapping regions. :

[0102] in, Represents all regions after segmentation arrive The union of the original image regions is the entire original image region. ; Represents any two distinct regions and There can be no overlap between them, and their intersection is an empty set. In this process, the edges of the insulator (high gradient ridges) precisely form these watersheds, thus accurately separating the insulator from the background into different regions.

[0103] Preferably, step S4 includes: Step S4-1: Based on the insulator region extracted in step S3, calculate the color features and texture features from the corresponding visible light image, and perform preliminary fusion with the IRIV-optimized multispectral spectral features output in step S3 to obtain a preliminary fused feature vector.

[0104] Step S4-2: The preliminary fusion feature vector obtained in step S4-1 is fused a second time using KPCA (Kernel Principal Component Analysis) to obtain a dimensionality-reduced and optimized core feature vector.

[0105] Preferably, in step S4-1, during the feature extraction process of the insulator region, the visible light image is first extracted. Given the color features, calculate its color feature vector:

[0106] in, Represents a specific color value; This represents the total number of color values. This indicates the proportion of a color value to all color values; This represents the total number of all possible, non-repeating color values ​​in the entire image (or insulator region); It is made by all The vectors formed together describe the statistical distribution of color throughout the insulator region.

[0107] Next, texture features of the visible light image are extracted, and gray-level gradient co-occurrence matrix (GGCM) extraction and analysis are performed. Laplacian transform is then applied to the gray-level images of each region to obtain the gray-level gradient maps for each region.

[0108] in, This indicates the grayscale gradient of the contaminated area at location. Pixel value at; Indicates the input image In position The pixel value.

[0109] Then, each gradient map is discretized at the gray level to obtain the gray-level gradient co-occurrence matrix (GGCM) of each contaminated region. , For set The number of elements in the middle. Normalization is performed to obtain the normalized matrix. :

[0110] in for The number of gray levels, for The number of gray levels.

[0111] Based on the above gray-gradient co-occurrence matrix, the texture features of each dirty area can be quantified. Four texture features with significant texture feature changes are selected: gray-level distribution non-uniformity (Gray Asymmetry), gradient distribution non-uniformity (Grads Asymmetry), gray-level average (Gray Mean), and inertia (Inertia).

[0112] Gray-level distribution non-uniformity mainly reflects the degree of unevenness in gray-level changes among pixels in a dirty image. It is used to quantify the degree of dirtiness represented by the dirty area image. The calculation formula is as follows:

[0113] Gradient distribution non-uniformity mainly reflects the degree of unevenness in the contour and edge distribution between pixels and their neighbors in a dirty image. It is used to quantify the edge changes of different degrees of dirtiness in a dirty region image. The calculation formula is as follows:

[0114] Gray-scale average primarily reflects the overall brightness of a region in an image; a higher value indicates greater brightness. It is used to quantify the total amount of dirt represented by a dirty region in an image. The calculation formula is as follows:

[0115] Inertia is a geometric parameter describing an image, and its calculation formula is shown below:

[0116] The above steps are used to fuse gray-level distribution non-uniformity, gradient distribution non-uniformity, gray-level averaging, and inertia to construct a texture feature vector Q:

[0117] The above processing is performed on the visible light image, that is, let Substituting into formula (15-21), we obtain the visible light image texture feature vector. .

[0118] Next, the three types of feature vectors describing dirt from different dimensions—color features, texture features, and IRIV spectral line features—are directly concatenated and integrated into a high-dimensional preliminary fusion feature vector. .

[0119] Preferably, in step S4-2, kernel principal component analysis (KPCA) is used for secondary fusion and dimensionality reduction. KPCA introduces kernel tricks to nonlinearly map the original feature space to a higher-dimensional feature space, where principal component analysis is performed to effectively extract the nonlinear feature structure. Its basic process is as follows: The initial fused feature vectors are collected to form a feature matrix. Assuming there are t samples, the initial fused feature vector of the b-th sample is denoted as... (Right now The entire feature matrix is ​​represented as:

[0120] The feature matrix X is mapped through a nonlinear mapping. Map from the original space to a high-dimensional feature space F. Calculate the covariance matrix in this space:

[0121] in To initially fuse feature vectors Mapping in high-dimensional space.

[0122] By solving the eigenvalue problem And introduce kernel functions The problem is transformed into finding the eigenvalues ​​and eigenvectors of the kernel matrix H, where The formula for solving the eigenvalue problem is:

[0123] Where C represents the covariance matrix of the sample in the high-dimensional feature space; v represents the eigenvalues ​​of the covariance matrix C; v represents the eigenvalues ​​of the covariance matrix C. eigenvectors.

[0124] The polynomial kernel function is selected as follows:

[0125] Where c is a constant term used to adjust the balance between higher and lower degree terms. In this invention, c=1 is used; d is the polynomial order. In this paper, d=3 is used to enhance the separability of samples in high-dimensional space while maintaining numerical stability.

[0126] In the KPCA process, principal components whose cumulative contribution rate reaches a set threshold are selected as the final fusion features, thereby achieving data dimensionality reduction while retaining most of the effective information. The final fusion feature has an e-dimensional dimension (e≪d), denoted as:

[0127] in This represents the final fusion feature set after kernel principal component analysis; This represents the selected principal components, each of which... The corresponding feature components are representative features in the high-dimensional feature space; e represents the dimension of the final fused feature.

[0128] This fusion feature set integrates key information from three types of features: color, texture, and spectrum. It has good discriminative power and compactness, making it suitable for subsequent tasks of identifying and classifying the degree of insulator contamination.

[0129] Combination Figure 4 Preferably, step S5 includes: Step S5-1: The "spectral features + image color features" obtained by secondary fusion of Kernel Principal Component Analysis (KPCA) in Step S4 are integrated with the extracted infrared texture features to form a high-dimensional input matrix.

[0130] The infrared texture features are obtained from step 4-1, let Substituting into formula (15-21), we obtain the visible light image texture feature vector. ,Will and By concatenating the features, a high-dimensional input feature vector can be formed. .

[0131] Step S5-2: Using the integrated high-dimensional input feature vector as input, construct a Radial Basis Neural Network (RBFNN) as the base classifier for ensemble learning. The discriminative latent function of the RBFNN is:

[0132] in, These are the hidden layer weight coefficients; The error coefficient; The center of the i-th neuron; is the extension constant of the radial basis function; the number of neurons in the input layer is equal to the dimension d of the high-dimensional feature vector, the number of neurons in the hidden layer is initialized (k=12), and the output layer has 4 neurons (corresponding to the 4 levels of dirtiness: "mild, moderate, moderate, and severe").

[0133] To enable RBFNN to adapt to high-dimensional input features The distribution of dirt is calculated based on the "width criterion". The process is as follows: First, determine the neuron center. The formulas for the spectral lines of each type of contamination level, and the matrix D, are as follows:

[0134] in Represents the center of the i-th neuron With the center of the j-th neuron The Euclidean distance between them.

[0135] The nearest distance is calculated based on matrix D, using the following formula:

[0136] Calculate the distance from each input sample data point to the center of the neuron, and construct a distance array R with coefficient L, as shown in the following formula:

[0137] Finally, the extended constant is defined. The formula is as follows:

[0138] Where n is the total number of samples, each RBFNN classifier can be built through the above steps, and the expansion constant is determined based on the width criterion. This can help capture spectral line distribution features and activate hidden layer neurons to varying degrees. 500 RBFNN classifiers are combined with modules to achieve ensemble learning and output the final judgment result, constructing an insulator pollution level detection model. The 500 RBFNN classifiers are generated using a Bagging ensemble learning method (random bootstrapping with replacement).

[0139] Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention and not intended to limit it. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the specific embodiments of the present invention, but such modifications or alterations are all within the scope of protection of the pending claims.

Claims

1. A method for insulator contamination detection and assessment based on multi-spectrum, characterized in that, Includes the following steps: Step S1: Use a multispectral imaging system equipped with a visible light camera and an infrared camera to acquire images of the target insulator, obtaining visible light images and infrared images; Step S2: Preprocess the acquired multispectral images, including image denoising, enhancement and registration, as well as radiometric calibration before spectral line extraction; Step S3: The insulator region is segmented using an image segmentation algorithm, and spectral features are extracted simultaneously from the visible light and infrared segmented regions. Then, the spectral features are fused using the Iterative Information Retaining Variable (IRIV) algorithm. Step S4: The spectral features and image features processed by IRIV are fused together, and then fused again by kernel principal component analysis (KPCA) to provide a fusion feature basis for dirt type identification; Step S5: Combine infrared texture features to construct an ensemble learning model and output the degree of dirtiness.

2. A method for detecting and evaluating contamination of insulators based on multi-spectrum according to claim 1, characterized in that, Step S1 includes: Step S1-1: Operate the multispectral imaging system equipped with a visible light camera and an infrared thermal imager to acquire images of the target insulator from different angles and distances; Steps S1-2: Acquire and store the visible light images of the target insulator respectively. and infrared images .

3. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 2, characterized in that, Step S2 includes: Step S2-1: To remove Gaussian noise from the original visible light image and infrared image obtained in step S1-2, the visible light image... Infrared images Perform Gaussian filtering to obtain the filtered image: ; in, , These are the coordinates of pixels in the visible light image and the infrared image, respectively. , These represent the pixels in a visible light image. and infrared image pixels Pixel values; , These represent the pixels in the visible light image after Gaussian filtering. and pixels in infrared images Pixel values; , These represent the Gaussian templates used for Gaussian filtering of visible light and infrared images, respectively. Indicates the convolution operation; Step S2-2: Perform histogram equalization on the denoised visible light and infrared images from step S2-1 to enhance the contrast of each image and improve image quality. The histogram equalization transformation function used is defined as follows: ; Where N is the total number of pixels in the image; L represents the number of pixels at the current gray level; L represents the total number of gray levels in the image. Represents the cumulative distribution function; It is the gray level of the k-th level in the original image. This maps the cumulative probability to a new gray level; Round indicates rounding operation. By mapping the pixels of the original image one-to-one using the cumulative distribution function according to the aforementioned histogram equalization formula, a result image with stronger contrast can be obtained: ; in, , These represent the pixel values ​​of each pixel in the visible light image and the infrared image after histogram equalization, respectively. Step S2-3: Using the visible light image as a reference, perform feature point matching-based image registration between the enhanced infrared image from step S2-2 and the reference image to obtain the transformation matrix from the visible light image coordinate system to the infrared image coordinate system. This ensures pixel-level alignment of the same insulator region in two different spectral images; Step S2-4: Before spectral line extraction, to eliminate the influence of instrument response, ambient light, and other factors on the multispectral image, the preprocessed image needs to be radiometrically calibrated, including DN correction and black-and-white correction, converting the image grayscale values ​​into reflectance data that can be used for subsequent analysis. The specific operation is as follows: Since changes in ambient light during hyperspectral acquisition can alter the image's density (DN) value, DN correction is necessary. The correction formula is as follows: ; in, These are the hyperspectral images of each sample after DN correction. These are the DN values ​​for each band of the original hyperspectral image before correction. It corrects the DN values ​​of each band in the whiteboard images at different depths. This is the DN value of each band in the white-corrected image of the whiteboard, where i represents 256 bands; After DN correction, the image needs to be calibrated as a reflectance image to identify the spectral characteristics of the substance, and then black and white correction is performed. The correction formula is as follows: ; In the formula, T represents the hyperspectral image of each sample after black and white correction. Indicates a white-corrected image. This represents a black-corrected image.

4. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 3, characterized in that, Step S3 includes: Step S3-1: Extract the preprocessed visible light image The gradient information in the image is used to obtain the preliminary edge contour of the insulator. The edge contour extraction of the insulator uses the Sobel operator to calculate the image gradient; its core formula includes the convolution calculation of the horizontal and vertical gradients. ; in, The gradient is the horizontal gradient, and the vertical gradient is used to detect edges in the image. The vertical gradient is used to detect horizontal edges in an image; * represents the convolution operation. Gradient magnitude is an indicator that measures the edge strength of each pixel in an image; the larger the magnitude, the stronger the edge strength. By setting a threshold to binarize the gradient magnitude map, a continuous preliminary edge profile of the insulator can be obtained. Step S3-2: Based on the edge gradient information obtained in step S3-1, the watershed algorithm is used to segment the insulator region in the visible light image; the edge gradient image is regarded as a topographic map, and the image is segmented into non-overlapping regions by simulating the flooding process, thereby accurately separating the insulator from the background; Step S3-3: Using the insulator region segmented from the visible light image in step S3-2 as a mask, and utilizing the registration relationship established in step S2, it is mapped to the infrared image in the same coordinate system, thereby accurately obtaining the corresponding region image of the insulator in the infrared spectrum; this process achieves coordinate mapping between images through affine transformation, and the core formula is: ; in, These are the coordinates of the insulator pixels in the visible light image; It is the transformation matrix from the visible light image coordinate system to the infrared image coordinate system; The coordinates of the insulator pixels in the visible light image The corresponding coordinates in the infrared coordinate system; through this transformation, the insulator outline of the visible light image is accurately projected onto the infrared image, thereby extracting the corresponding area and realizing location-based feature association; Steps S3-4: For the segmented regions of the preprocessed and radiometrically calibrated visible and infrared images, extract the grayscale values ​​or radiance values ​​of each pixel in different wavelength bands, and construct the spectral curve of each pixel as the spectral feature of the region; for pixels within the segmented region of the visible image... Its spectral characteristics are expressed as ,in For visible light bands; for pixels within the segmentation area of ​​an infrared image. Its spectral characteristics are expressed as ,in The infrared band is used as the starting point. These spectral features are then integrated to form an initial multispectral feature set. Based on the IRIV algorithm, the information contribution of each spectral feature variable is iteratively calculated, along with the correlation between each spectral feature variable and the degree of contamination, as well as the redundancy between variables. Based on a preset threshold, the optimal subset of spectral features is gradually selected, eliminating redundant and irrelevant spectral features, and retaining P optimal spectral features. This enables the effective fusion of multispectral line features, resulting in fused spectral line features. .

5. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 4, characterized in that, In step S3-2, a watershed algorithm is proposed to segment the insulator skirt region. First, to effectively highlight the insulator region and suppress noise, the visible light image is... Convert from RGB space to HSV space to obtain the S-component image. The calculation process is as follows: ; in, Indicates color space operations; M is the preprocessed visible light image; M is the mask image used for morphological reconstruction. For S-component images Gradient calculation is performed to obtain the gradient image. : ; in, This indicates the output gradient image in coordinates. The gradient magnitude of a pixel; the larger the value, the stronger the edge intensity of that pixel. For S-component image On point Pixel values; Representing an image The first-order partial derivative in the x-direction reflects the rate of change of the gray value at that point in the horizontal direction; Representing an image The first-order partial derivative in the y-direction reflects the rate of change of the gray value at that point in the vertical direction; Next, the gradient image is detected. All local minima are identified, and these points are labeled as distinct "basins". The seed; simulating a seed from arrive The rising water level h, The minimum gradient magnitude in the gradient image. Let h be the maximum gradient magnitude in the gradient image; for each water level h, define the set of regions submerged below water level h as... : ; As the water level h increases, The area will continue to expand; when it comes from two different basins and flooded areas and When the merger is about to take place, that is, at a certain water level When they first meet, they satisfy the following formula: ; At this point, to prevent the basins from merging, a watershed needs to be constructed at their intersection, eventually forming a closed boundary line; mathematically, this watershed is a set of points that satisfy the following conditions. : ; Finally, when the water level reaches the maximum gradient value of the entire image... At that time, all the basins were filled with water, and the collection of watersheds was constructed. The image was then segmented into non-overlapping regions. : ; in, Represents all regions after segmentation arrive The union of the original image regions is the entire original image region. ; This indicates any two distinct regions. and There can be no overlap between them, and their intersection is an empty set; In this process, the edges of the insulator precisely form these watersheds, thus accurately separating the insulator from the background into different regions.

6. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 1, characterized in that, Step S4 includes: Step S4-1: Based on the insulator region extracted in step S3, calculate the color features and texture features from the corresponding visible light image, and perform preliminary fusion with the IRIV-optimized multispectral spectral features output in step S3 to obtain a preliminary fused feature vector. Step S4-2: Use KPCA to perform secondary fusion on the preliminary fusion feature vector obtained in step S4-1 to obtain the dimensionality-reduced and optimized core feature vector.

7. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 6, characterized in that, In step S4-1, during the feature extraction process of the insulator region, the visible light image is first extracted. Given the color features, calculate its color feature vector: ; in, Represents a specific color value; This represents the total number of color values. This indicates the proportion of a color value to all color values; It represents the total number of all possible, non-repeating color values ​​in the entire image; It is made by all The vectors formed together describe the statistical distribution of color throughout the entire insulator region; Next, texture features of the visible light image are extracted, and gray-level gradient co-occurrence matrix (GGCM) extraction and analysis are performed. Laplacian transform is then applied to the gray-level images of each region to obtain the gray-level gradient maps for each region. ; in, This indicates the grayscale gradient of the contaminated area at location. Pixel value at; Indicates the input image In position Pixel values; Then, each gradient map is discretized at the gray level to obtain the gray-level gradient co-occurrence matrix (GGCM) of each contaminated region. , For set The number of elements in the middle; Normalization is performed to obtain the normalized matrix. : ; in for The number of gray levels, for The number of gray levels; Based on the above gray-gradient co-occurrence matrix, the texture features of each dirty area can be quantified. Four texture feature quantities are selected: gray-level distribution non-uniformity (Gray Asymmetry), gradient distribution non-uniformity (Grads Asymmetry), gray-level average (Gray Mean), and inertia (Inertia). Gray-level distribution non-uniformity mainly reflects the degree of unevenness in gray-level changes among pixels in a dirty image. It is used to quantify the degree of dirtiness represented by the dirty area image. The calculation formula is as follows: ; Gradient distribution non-uniformity mainly reflects the degree of unevenness in the contour and edge distribution between pixels and their neighbors in a dirty image. It is used to quantify the edge changes of different degrees of dirtiness in a dirty region image. The calculation formula is as follows: ; Gray-scale average primarily reflects the overall brightness of a region in an image; a higher value indicates greater brightness. It is used to quantify the total amount of dirt represented by a dirty region in an image. The calculation formula is as follows: ; Inertia is a geometric parameter describing an image, and its calculation formula is shown below: ; The above steps are used to fuse gray-level distribution non-uniformity, gradient distribution non-uniformity, gray-level averaging, and inertia to construct a texture feature vector Q: ; The above processing is performed on the visible light image, that is, let Substituting into formula (15-21), we obtain the visible light image texture feature vector. ; Next, the three types of feature vectors describing dirt from different dimensions—color features, texture features, and IRIV spectral line features—are directly concatenated and integrated into a high-dimensional preliminary fusion feature vector. .

8. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 6, characterized in that, In step S4-2, Kernel Principal Component Analysis (KPCA) is used for secondary fusion and dimensionality reduction. KPCA introduces kernel tricks to nonlinearly map the original feature space to a higher-dimensional feature space, where principal component analysis is performed to effectively extract the nonlinear feature structure. The basic process is as follows: The initial fused feature vectors are collected to form a feature matrix; assuming there are t samples, the initial fused feature vector of the b-th sample is denoted as... (Right now The entire feature matrix is ​​represented as: ; The feature matrix X is mapped through a nonlinear mapping. Map from the original space to a high-dimensional feature space F. Calculate the covariance matrix in this space: ; in To initially fuse feature vectors Mapping in high-dimensional space; By solving the eigenvalue problem And introduce kernel functions The problem is transformed into finding the eigenvalues ​​and eigenvectors of the kernel matrix H, where The formula for solving the eigenvalue problem is: ; Where C represents the covariance matrix of the sample in the high-dimensional feature space; v represents the eigenvalues ​​of the covariance matrix C; v represents the eigenvalues ​​of the covariance matrix C. eigenvectors; The polynomial kernel function is selected as follows: ; Where c is a constant term used to adjust the balance between higher and lower degree terms, and in this invention c=1; d is the polynomial order, and in this paper d=3, so as to enhance the separability of samples in high-dimensional space while maintaining numerical stability. In the KPCA process, principal components whose cumulative contribution rate reaches a set threshold are selected as the final fusion features, thereby achieving data dimensionality reduction while retaining most of the effective information; the final fusion feature dimension is e-dimensional (e≪d), denoted as: ; in This represents the final fusion feature set after kernel principal component analysis; represents the selected principal components, corresponding to representative feature components in the high-dimensional feature space; e represents the dimension of the final fused feature. This fusion feature set integrates key information from three types of features: color, texture, and spectrum. It has good discriminative power and compactness, making it suitable for subsequent tasks of identifying and classifying the degree of insulator contamination.

9. The method for detecting and evaluating insulator contamination based on multispectral analysis according to claim 7, characterized in that, Step S5 includes: Step S5-1: The "spectral features + image color features" obtained by secondary fusion of Kernel Principal Component Analysis (KPCA) in Step S4 are integrated with the extracted infrared texture features to form a high-dimensional input matrix; The infrared texture features are obtained from step 4-1, let Substituting into formula (15-21), we obtain the visible light image texture feature vector. ,Will and By concatenating the features, a high-dimensional input feature vector can be formed. ; Step S5-2: Using the integrated high-dimensional input feature vector as input, construct a Radial Basis Neural Network (RBFNN) as the base classifier for ensemble learning; the discriminative latent function of the RBFNN is: ; in, These are the hidden layer weight coefficients; The error coefficient; The center of the i-th neuron; is the expansion constant of the radial basis function; the number of neurons in the input layer is equal to the dimension d of the high-dimensional feature vector, the number of neurons in the hidden layer is initialized (k=12), and the output layer has 4 neurons; To enable RBFNN to adapt to high-dimensional input features The distribution of dirt is calculated based on the "width criterion". The process is as follows: First, determine the neuron center. The formulas for the spectral lines of each type of contamination level, and the matrix D, are as follows: ; in Represents the center of the i-th neuron With the center of the j-th neuron The Euclidean distance between them; The nearest distance is calculated based on matrix D, using the following formula: ; Calculate the distance from each input sample data point to the center of the neuron, and construct a distance array R with coefficient L, as shown in the following formula: ; Finally, the extended constant is defined. The formula is as follows: ; Where n is the total number of samples, each RBFNN classifier can be built through the above steps, and the expansion constant is determined based on the width criterion. This approach can help capture spectral line distribution characteristics and activate hidden layer neurons to varying degrees. 500 RBFNN classifiers are combined with modules to achieve ensemble learning and output the final judgment result, constructing an insulator pollution level detection model. The 500 RBFNN classifiers are generated through a Bagging ensemble learning method.

Images