Composite insulator aging evaluation method based on combination of improved BAM attention mechanism and multispectral imaging

Abstract

The invention discloses a composite insulator aging evaluation method combining an improved BAM attention mechanism with multispectral imaging. The method comprises the following specific steps: (1) collecting multispectral imaging data of an aged composite insulator; (2) multi-spectral image data set channel screening of the aged composite insulator; (3) preprocessing the multispectral image of the aged composite insulator; (4) dividing a multispectral image sample of the aged composite insulator; (5) extracting multispectral image information features of the aged composite insulator; and (6) constructing a multispectral image recognition model of the aged and damaged composite insulator. According to the method, a multispectral Hash algorithm (MSI-HA) and an improved feature extraction method of a channel space attention mechanism (IBAM) are adopted, and the method is specially used for multispectral image analysis of the aged composite insulator. According to the method, firstly, a multispectral Hash algorithm (MSI-HA) is constructed through combination of an average Hash algorithm (AHA), a perceptual Hash algorithm (PHA) and a differential Hash algorithm (DHA), channel screening of multispectral images is achieved, data redundancy is effectively reduced, and key feature information is reserved. Furthermore, an improved bottleneck attention module (IBAM) is introduced, the module is optimized on the basis of the traditional BAM, and the channel and spatial features of the image can be captured more accurately. By embedding the IBAM into the network model, deep feature extraction can be performed on the multispectral image, so that the capability of identifying the aging features of the insulator is improved. While higher detection precision is realized, the module parameter quantity can be reduced, the calculation efficiency is improved, and the method is very suitable for composite insulator aging state evaluation.

Description

Technical Field

[0001] The present invention relates to the field of composite insulator aging status assessment, and more specifically to a composite insulator aging assessment method that improves the BAM attention mechanism combined with multispectral imaging. Background Art

[0002] The continuous development of power systems has led to the widespread use of composite insulators in transmission lines. However, composite insulators are susceptible to corona and ultraviolet radiation over long-term operation, causing aging that impacts their performance. Accurately assessing the degree of composite insulator aging is crucial. Current methods for assessing the aging state of composite insulators can be divided into two categories: contact and non-contact. However, contact measurement methods are cumbersome and prone to damage to equipment, while existing non-contact inspection methods each have their own limitations. Infrared and ultraviolet imaging have relatively narrow wavelengths and can only indirectly reflect the insulator's condition through electrical and thermal properties within these wavelengths, making it difficult to extract silicone rubber aging characteristics across a wider range. Ultrasonic testing technology needs to be improved in terms of remote sensing and sensitivity in detecting defect scales. Challenges with coupling, attenuation, and ultrasonic transducer performance also need to be overcome. Therefore, an efficient and accurate non-contact method for assessing the surface aging state of composite insulators is urgently needed.

[0003] Multispectral imaging (MSI) is an imaging method that simultaneously captures information from multiple optical spectrum bands. By measuring the reflection and absorption of an object multiple times in the same time and space, it produces images of the object across different wavelengths. In power systems, MSI can be used to detect typical defects in transmission lines, such as field-induced heating in insulators. Furthermore, MSI can fuse visible and infrared spectral information, generating multispectral fusion images through image fusion methods such as wavelet transforms. These images comprehensively reflect equipment details such as temperature and brightness, enabling rapid and accurate fault location, significantly improving the efficiency of power equipment fault detection. Summary of the Invention

[0004] The purpose of the present invention is to provide an improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method to solve the current non-contact composite insulator surface aging status assessment problems of low accuracy and low efficiency.

[0005] In order to solve the above problems, the present invention adopts the following solutions:

[0006] S1: Multispectral imaging data acquisition of aged composite insulators;

[0007] S2: Channel screening of multispectral image dataset of aged composite insulators;

[0008] S3: Multispectral image preprocessing of aged composite insulators;

[0009] S4: multispectral image sample division of aged composite insulator;

[0010] S5: Feature extraction of multispectral image information of aged composite insulators;

[0011] S6: Construction of a multispectral image recognition model for aged composite insulators.

[0012] Preferably, in step S1, multispectral imaging technology is used to image the aged composite insulator to obtain image information thereof in multiple spectral bands, thereby constructing a composite insulator multispectral image dataset containing aging characteristics.

[0013] Preferably, in step S2, a multispectral hash algorithm (MSI-HA) generated by an average hash algorithm (AHA), a perceptual hash algorithm (PHA), and a differential hash algorithm (DHA) is used to calculate mixed hash values of the multispectral image data of the aged composite insulator in 25 channels, and a channel with the smallest similarity is determined therefrom;

[0014] Specifically, the mathematical formula of the average hash algorithm (AHA) is:

[0015] (1) The grayscale conversion formula is:

[0016] Gray(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j)

[0017] Where Gray(i,j) is the grayscale value of pixel (i,j); R(i,j), G(i,j), and B(i,j) are the red, green, and blue component values of pixel (i,j), respectively.

[0018] (2) The calculation formula for the average gray value is:

[0019]

[0020] Where: Average is the arithmetic mean of the grayscale values of all pixels in the entire image; N is the total number of pixels in the image;

[0021] (3) The hash value generation formula is:

[0022]

[0023] Where: Hash(i,j) is a binary hash value (0 or 1), generated by comparing Gray(i,j) with Average (1 if greater than the average, otherwise 0);

[0024] Specifically, the mathematical formula of the Perceptual Hash Algorithm (PHA) is:

[0025] (1) Grayscale conversion formula:

[0026] Gray(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j)

[0027] Where Gray(i,j) is the grayscale value of pixel (i,j); R(i,j), G(i,j), and B(i,j) are the red, green, and blue component values of pixel (i,j), respectively:

[0028] (2) Two-dimensional discrete cosine transform (DCT) formula:

[0029]

[0030] Where: F(u,v) is the frequency domain coefficient after DCT transformation, and the corresponding frequency index is (u,v); C(u) and C(v) are normalized coefficients, defined as: N is the total number of pixels in the image;

[0031] (3) Average value calculation formula:

[0032]

[0033] Where: Average is the arithmetic mean of the grayscale values of all pixels in the entire image;

[0034] (4) Hash value generation formula:

[0035]

[0036] Where: Hash(u,v) is a binary hash value (0 or 1), generated by comparing F(u,v) with Average (1 if greater than the average, otherwise 0);

[0037] Specifically, the mathematical formula of the differential hash algorithm (DHA) is:

[0038] (1) Grayscale conversion formula:

[0039] Gray(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j)

[0040] Where Gray(i,j) is the grayscale value of pixel (i,j); R(i,j), G(i,j), and B(i,j) are the red, green, and blue component values of pixel (i,j), respectively.

[0041] (2) Adjacent pixel difference calculation formula:

[0042] D(i,j)=Gray(i,j)-Gray(i,j+1)

[0043] Where: the range of i is 0≤i<N, the range of j is 0≤j<N-1; N is the total number of pixels in the image;

[0044] (3) Hash value generation formula:

[0045]

[0046] Where: Hash(i,j) is the binary hash value (0 or 1), D(i,j) is 1 when D(i,j)>0, otherwise it is 0;

[0047] The following is the mathematical formula of the multi-spectral hashing algorithm (MSI-HA) which is a combination of the above three hashing algorithms:

[0048] (1) The similarity score calculation formula is:

[0049]

[0050] Where: D k is the Hamming distance of the kth hash algorithm; L is the total length of the hash value; S k is the similarity score of the kth algorithm, ranging from [0,1], where a larger value indicates a more similar result;

[0051] (2) The normalized similarity score formula is:

[0052]

[0053] Where: μ k is the average similarity score of the kth algorithm on the historical data set; σ k is the standard deviation of the kth algorithm on the historical data set; S k ′ is the score after Z-score standardization, eliminating the dimensional differences between algorithms;

[0054] (3) The adaptive weight allocation formula is:

[0055]

[0056] Where: p k,i is the score distribution probability of the i-th sample of the k-th algorithm in the historical data; H k is the information entropy of the kth algorithm, which measures its discrimination ability (the greater the entropy, the higher the weight); w k is the normalized weight, satisfying ∑w k =1;

[0057] (4) The weighted average and final score formula is:

[0058]

[0059] Where: S MSI-HA is the final fusion score, with an unbounded range. A larger value indicates a higher degree of discrimination of the multispectral channels.

[0060] Preferably, in step S3, different types of artificial noise or blur effects are used to preprocess the single-channel multispectral image of the aged composite insulator, wherein the artificial noise includes Gaussian noise and impulse noise, and the blur effects include dynamic blur and atmospheric disturbance blur.

[0061] Preferably, in step S4, a random sampling method is used to divide the pre-processed single-channel multispectral image data of the aged composite insulator into independent training sets and test sets according to a certain ratio.

[0062] Preferably, in step S5, the convolutional neural network (CNN) for extracting multispectral image information features is a network structure including an IBAM attention mechanism, where IBAM is an improved channel space attention module of BAM (Bottleneck Attention Module). The channel attention module generates channel importance weights through adaptive average pooling, two convolutions, and ReLU and Sigmoid activation; the spatial attention module performs channel, horizontal, and vertical pooling, sums the dot product of the feature map and the query vector, and then normalizes it through ReLU6 to obtain a spatial attention map; in the fusion mechanism, the channel and spatial attention maps are added, activated by Sigmoid, and fused with the original feature map through a learnable residual coefficient to output an enhanced feature map; the network structure mainly includes an IBAM module, a convolution layer, a pooling layer, a normalization layer, an activation function layer, a drop layer, a reshape layer, and a fully connected layer.

[0063] Preferably, in step S6, the multispectral image features extracted by CNN are used to construct an IBAM network multispectral image recognition model on the training set, the parameters of the IBAM network multispectral image recognition model are determined, and then the test set is used to test the recognition effect and verify the performance of the model.

[0064] Based on the above technical solution, the beneficial effects of the present invention are as follows: the present invention is an improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method, which utilizes a new channel space attention mechanism IBAM. Compared with the traditional aging composite insulator detection method, this technology detects the imaging pictures of aging composite insulators by improving the BAM attention mechanism combined with multispectral imaging technology, which can achieve higher detection accuracy, effectively extract richer and more accurate feature information, and is simpler and easier to operate. This technology not only has the ability to achieve accurate detection in the multispectral images of aging composite insulators, but also has strong anti-interference performance, and can ensure the reliability of detection results under complex backgrounds. In addition, because it adopts a new channel space attention mechanism module IBAM, the module has fewer parameters and higher computational efficiency. Therefore, this technology can significantly improve the efficiency of extracting key information in the multispectral data processing of large-scale aged composite insulators, and is very suitable for the aging status assessment of composite insulators. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flow chart of the aging status assessment of composite insulators according to an embodiment of the present invention;

[0066] Figure 2 This is a simplified structural diagram of a convolutional neural network used in an embodiment of the present invention to extract information features of multispectral images;

[0067] Figure 3 This is a simplified structural diagram of the IBAM module for implementing feature extraction in the present invention. Specific implementation methods

[0068] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0069] like Figure 1 As shown, the present invention is a composite insulator aging assessment method that improves the BAM attention mechanism combined with multispectral imaging, and the steps are as follows:

[0070] S1: Using a real-time multispectral Mosaic area camera produced by Shanghai Wuling Optoelectronics Technology Co., Ltd. as a multispectral imaging module, we acquired multispectral images and spectral data from multiple samples of aged composite insulators. We obtained 25 sets of multispectral imaging data for aged composite insulators, covering a wavelength range of 600 to 875 nm, encompassing a total of 25 bands.

[0071] S2: For multispectral images of 25 bands, a multispectral hashing algorithm (MSI-HA) is constructed using a combination of the average hashing algorithm (AHA), the perceptual hashing algorithm (PHA), and the differential hashing algorithm (DHA). This algorithm obtains hash values for each band of the multispectral image. Subsequently, the image similarity is quantified by comparing the Hamming distances of the hash values of the multispectral images of each band. It should be noted that image similarity is inversely proportional to the Hamming distance: the larger the Hamming distance, the less similar the images are. Finally, based on the Hamming distance, the band with the largest Hamming distance corresponding to the multispectral image is selected; this band is the selected channel.

[0072] S3: The multispectral image acquisition process for multiple samples of aged composite insulators was completed in a static laboratory environment. Because the laboratory environment is ideal, preprocessing the acquired images is necessary to simulate the imaging environment of real-world aged composite insulators outdoors and enhance the robustness of the IBAM network. This processing method involves adding Gaussian and impulse noise to the acquired images, as well as applying motion blur and atmospheric disturbance blur.

[0073] S4: As Figure 1 As shown in Figure 1, the preprocessed multispectral image dataset of aged composite insulators is divided into two subsets using random sampling method, where 80% of the data is used as the training set and the remaining 20% of the data is used as the test set.

[0074] S5: If Figure 2 As shown, the improved BAM network is used to extract features from the image information of aged composite insulators. The network takes the aged composite insulator image data as input and obtains the output result through the following steps: first, it is processed by the first convolution layer Conv2d, which has an input channel of 1, an output channel of 256, a convolution kernel size of 3×3, a stride of 1, and a padding of 1; then it passes through the pooling layer A: the maximum pooling layer MaxPool2d, with a pooling window size of 2×2 and a stride of 2; then it passes through the normalization layer BatchNorm2d, acting on 256 channels; then the activation function ReLU is applied; then there is the dropout layer A, which is a Dropout layer (dropout rate 0.5), which randomly drops 50% of the neurons; then it passes through the BAM attention module (acting on 256 channels); then it passes through the pooling layer B: the adaptive average pooling layer AdaptiveAvgPool2d(1,1), which compresses the feature map to a size of 1×1; then it enters the reshaping layer to resize the tensor to (batch size, 256), i.e., Reshape(-1, 256); the discard layer A and the Dropout layer (discarding rate 0.5) are applied again; finally, the fully connected layer Linear is passed to output the image information features extracted by the improved BAM network.

[0075] The improved BAM module is the channel spatial attention mechanism module IBAM. Since the original BAM spatial attention branch uses dilated convolution, which may bring redundant parameters and computational complexity, the invented IBAM only makes non-learnable parameterized improvements to the spatial attention branch of BAM, thereby reducing redundant parameters and computational complexity, and improving the efficiency and performance of the model. Figure 3 As shown in the figure, the improved BAM module (IBAM) mainly includes a channel branch, an improved spatial branch and a fusion mechanism. The specific operation process is as follows: in the channel attention module, the input feature map is first adaptively average pooled (pooling layer A) to compress it to a 1×1 size to obtain global channel feature information; then a 1×1 convolution kernel (convolution layer A) is used to reduce the number of channels to the original 1 / reduction ratio, reducing the computational complexity and extracting high-level features between channels; then the activation function ReLU (activation function layer A) is applied to introduce nonlinearity, and the number of channels is restored to the original number of channels through a 1×1 convolution kernel (convolution layer B) to generate preliminary features of the channel attention map; finally, the Sigmoid activation function (activation function layer B) is applied to map the value of the channel attention map to between 0 and 1, indicating the importance weight of each channel. In the improved spatial attention module, the input feature map is first averaged in the channel dimension (pooling layer B: channel average pooling) to obtain a 2D feature map for subsequent spatial attention calculation; then the 2D feature map is averaged in the height dimension (pooling layer C: horizontal average pooling) and the width dimension (pooling layer D: vertical average pooling) to obtain horizontal and vertical query vectors; then the input feature map is element-wise dot convolution with the horizontal and vertical query vectors (convolution layer C), and the sum is taken in the width and height dimensions respectively to obtain the intermediate results in the horizontal and vertical directions; finally, the intermediate results in the horizontal and vertical directions are added through the broadcast mechanism, and the ReLU6 activation function is applied (activation function layer C) and then divided by 6 for normalization to obtain the final spatial attention map. In the fusion mechanism, the channel attention map and the spatial attention map are added to obtain a comprehensive attention map; then the Sigmoid activation function is applied to the comprehensive attention map (activation function layer B) to ensure that the attention weight is between 0 and 1; finally, the original input feature map and the attention-adjusted feature map are weightedly fused through the learnable residual coefficient gamma to output an enhanced feature map, where gamma is a learnable parameter used to balance the contribution of the original features and the attention features.

[0076] S6: First, the IBAM network is used to extract features from the multispectral images of aged composite insulators. The extracted features are then used to construct an aging assessment model for aged composite insulators on the training set. Finally, the model is tested on the test set to evaluate its recognition accuracy and efficiency.

[0077] Obviously, for those skilled in the art, the invention is not limited to the specific details involved in the above-mentioned exemplary embodiments. The invention can be implemented in other specific forms without violating the spirit or basic characteristics of the invention. Therefore, from any perspective, these embodiments should be regarded as exemplary rather than restrictive. The scope of the invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the invention. Any figure mark in the claims should not be considered to limit the claims involved. In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method is for clarity only. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A composite insulator aging assessment method based on an improved BAM attention mechanism combined with multispectral imaging, comprising the following steps: S1: Multispectral imaging data acquisition of aged composite insulators; S2: Channel screening of multispectral image dataset of aged composite insulators; S3: Multispectral image preprocessing of aged composite insulators; S4: multispectral image sample division of aged composite insulator; S5: Feature extraction of multispectral image information of aged composite insulators; S6: Construction of multispectral image recognition model for aged composite insulators; According to the improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method of claim 1, in step S3, different types of artificial noise or blur effects are used to preprocess the single-channel multispectral image of the aged composite insulator, wherein the artificial noise includes Gaussian noise and impulse noise, and the blur effects include dynamic blur and atmospheric disturbance blur.

2. According to the improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method of claim 1, in step S1, multispectral imaging technology is used to image the aged composite insulator to obtain its image information in multiple spectral bands, thereby constructing a composite insulator multispectral image dataset containing aging characteristics.

3. According to the improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method of claim 1, in step S2, a multispectral hash algorithm (MSI-HA) generated by an average hash algorithm (AHA), a perceptual hash algorithm (PHA), and a differential hash algorithm (DHA) is used to calculate the mixed hash values of the multispectral image data of the aged composite insulator in 25 channels, and determine the channel with the smallest similarity. Specifically, the mathematical formula of the average hash algorithm (AHA) is: (1) The grayscale conversion formula is: Gray(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j) Where Gray(i,j) is the grayscale value of pixel (i,j); R(i,j), G(i,j), and B(i,j) are the red, green, and blue component values of pixel (i,j), respectively. (2) The calculation formula for the average gray value is: Where: Average is the arithmetic mean of the grayscale values of all pixels in the entire image; N is the total number of pixels in the image; (3) The hash value generation formula is: Where: Hash(i,j) is a binary hash value (0 or 1), generated by comparing Gray(i,j) with Average (1 if greater than the average, otherwise 0); Specifically, the mathematical formula of the Perceptual Hash Algorithm (PHA) is: (1) Grayscale conversion formula: Gray(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j) Where Gray(i,j) is the grayscale value of pixel (i,j); R(i,j), G(i,j), and B(i,j) are the red, green, and blue component values of pixel (i,j), respectively: (2) Two-dimensional discrete cosine transform (DCT) formula: Where: F(u,v) is the frequency domain coefficient after DCT transformation, and the corresponding frequency index is (u,v); C(u) and C(v) are normalized coefficients, defined as: N is the total number of pixels in the image; (3) Average value calculation formula: Where: Average is the arithmetic mean of the grayscale values of all pixels in the entire image; (4) Hash value generation formula: Where: Hash(u,v) is a binary hash value (0 or 1), generated by comparing F(u,v) with Average (1 if greater than the average, otherwise 0); Specifically, the mathematical formula of the differential hash algorithm (DHA) is: (1) Grayscale conversion formula: Gray(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j) Where Gray(i,j) is the grayscale value of pixel (i,j); R(i,j), G(i,j), and B(i,j) are the red, green, and blue component values of pixel (i,j), respectively. (2) Adjacent pixel difference calculation formula: D(i,j)=Gray(i,j)-Gray(i,j+1) Where: the range of i is 0≤i<N, the range of j is 0≤j<N-1; N is the total number of pixels in the image; (3) Hash value generation formula: Where: Hash(i,j) is the binary hash value (0 or 1), D(i,j) is 1 when D(i,j)>0, otherwise it is 0; The following is the mathematical formula of the multi-spectral hashing algorithm (MSI-HA) which is a combination of the above three hashing algorithms: (1) The similarity score calculation formula is: Where: D k is the Hamming distance of the kth hash algorithm; L is the total length of the hash value; S k is the similarity score of the kth algorithm, ranging from [0,1], where a larger value indicates a more similar result; (2) The normalized similarity score formula is: Where: μ k is the average similarity score of the kth algorithm on the historical data set; σ k is the standard deviation of the kth algorithm on the historical data set; S k ′ is the score after Z-score standardization, eliminating the dimensional differences between algorithms; (3) The adaptive weight allocation formula is: Where: p k,i is the score distribution probability of the i-th sample of the k-th algorithm in the historical data; H k is the information entropy of the kth algorithm, which measures its discrimination ability (the greater the entropy, the higher the weight); w k is the normalized weight, satisfying ∑w k =1; (4) The weighted average and final score formula is: Where: S MSI-HA is the final fusion score, with an unbounded range. A larger value indicates a higher degree of discrimination of the multispectral channels.

4. According to the improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method of claim 1, in step S4, a random sampling method is used to divide the preprocessed single-channel multispectral image data of the aged composite insulator into independent training sets and test sets according to a certain ratio.

5. According to the improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method of claim 1, in step S5, the convolutional neural network (CNN) used to extract multispectral image information features is a network structure that includes an IBAM attention mechanism. IBAM is a channel space attention module improved from BAM (Bottleneck Attention Module), wherein the channel attention module generates channel importance weights through adaptive average pooling, two convolutions, and ReLU and Sigmoid activation; The spatial attention module performs channel, horizontal and vertical pooling, sums the dot product of the feature map and the query vector, and then normalizes it through ReLU6 to obtain the spatial attention map; In the fusion mechanism, the channel and spatial attention maps are added, activated by Sigmoid, and fused with the original feature map through the learnable residual coefficient to output the enhanced feature map; the network structure mainly includes IBAM module, convolution layer, pooling layer, normalization layer, activation function layer, dropout layer, reshape layer, and fully connected layer.

6. According to the improved BAM attention mechanism combined with multispectral imaging composite insulator aging assessment method of claim 1, in step S6, the multispectral image features extracted by CNN are used to construct an IBAM network multispectral image recognition model on the training set, the parameters of the IBAM network multispectral image recognition model are determined, and then the test set is used to test the recognition effect and verify the performance of the model.

Images