High-voltage bushing state diagnosis method and system

Abstract

The invention discloses a high-voltage bushing state diagnosis method and system, relates to the technical field of high-voltage bushing state diagnosis, and is used for solving the technical problem of low accuracy of high-voltage bushing state diagnosis in an existing diagnosis mode. The high-voltage bushing state diagnosis method comprises the following steps: acquiring a visible light image and a thermal image of a high-voltage bushing; performing feature fusion on the visible light image and the thermal image to obtain a fused image; extracting feature data in the fused image by using scale invariant feature transformation; inputting the feature data into a probabilistic neural network optimized by a salp bee colony algorithm, and outputting a high-voltage bushing state diagnosis result represented by a fused image by the probabilistic neural network; the diagnosis results comprise normality, mechanical faults and electrical faults. The technical scheme of the invention is used for providing the high-voltage bushing state diagnosis method and system.

Description

Technical Field

[0001] The present invention relates to the technical field of high-voltage bushing status diagnosis, and more particularly to a high-voltage bushing status diagnosis method and system. Background Art

[0002] High-voltage bushings are a crucial component of substation systems. Faults in these bushings are primarily categorized as mechanical and electrical. While traditional mathematical models based on multi-scale transformations, sparse representations, and low-rank representations can extract some complementary features, their reliance on manual rule-based design makes them difficult to adapt to the demands of complex operating conditions. For example, coupling interference caused by contamination on the bushing surface can easily lead to misjudgments.

[0003] In recent years, deep learning methods, such as convolutional neural networks (CNNs) and generative adversarial networks (GANs), have significantly improved fusion performance in high-voltage bushing diagnosis through semantic feature extraction. Specifically, CNNs can identify patterns associated with micron-scale cracks and local overheating, but they require deep stacking to expand the receptive field for feature extraction, and underlying information in the source image (such as edges and line structures) may be lost during the convolution process. While GANs preserve the authenticity of thermal distribution, they can lead to low fault diagnosis accuracy due to network training instability. Furthermore, Transformer technology can achieve global context modeling through self-attention mechanisms, correlating cross-regional coordinated changes in temperature fields and texture features (such as causal analysis between shed temperature rise and adjacent porcelain corrosion). However, its high computational complexity makes it difficult to meet the real-time processing requirements of high-resolution images, and its accuracy in high-voltage bushing fault diagnosis remains low. It can be seen that, precisely because of these shortcomings, the existing technologies have low accuracy in high-voltage bushing condition diagnosis and are unable to meet practical diagnostic requirements. Summary of the Invention

[0004] The present invention aims to provide a method and system for diagnosing the status of a high-voltage bushing, so as to solve the technical problem of low accuracy of the existing diagnostic methods for high-voltage bushing status. In view of this, the present invention is implemented through the following solutions.

[0005] In a first aspect, the present invention provides a method for diagnosing a high-voltage bushing state, comprising:

[0006] Acquire visible light and thermal images of high-voltage bushings;

[0007] Performing feature fusion on the visible light image and the thermal image to obtain a fused image;

[0008] extracting feature data from the fused image using a scale-invariant feature transform;

[0009] The feature data is input into a probabilistic neural network optimized by the SALP bee colony algorithm, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing state represented by the fused image; the diagnosis result includes normal, mechanical fault and electrical fault.

[0010] Compared with the prior art, in the high-voltage bushing status diagnosis method of the present invention, after obtaining the visible light image and thermal image of the high-voltage bushing, feature fusion processing is performed to obtain the fused image. After extracting feature data in the fused image using the scale-invariant feature transform (SIFT), the feature data is input into the probabilistic neural network optimized by the SALPS bee colony algorithm to obtain the diagnosis result of the high-voltage bushing status. Specifically, the feature fusion process in the above technical solution of the present invention can be implemented using a cross-modal Transformer fusion algorithm, which can capture global interactions in the spatial and channel domains through spatial transformers and channel transformers. Furthermore, by optimizing the probabilistic neural network (PNN) parameters based on the SALPS bee colony algorithm, the correlation between fault features and classification results can be effectively enhanced, and the accuracy of the PNN model for high-voltage bushing fault diagnosis can be improved. At the same time, the PNN model can not only effectively analyze the fault status, but also accelerate the precise positioning of the cause of the high-voltage bushing abnormality through its ideal model size. Its rapid response characteristics can meet the needs of intelligent real-time inspection. The above technical solution of the present invention solves the technical problem of low accuracy of high-voltage bushing status diagnosis in existing diagnostic methods, while reducing the cost of massive data processing and delays in the processing process.

[0011] Furthermore, in the high-voltage bushing status diagnosis method of the present invention, feature fusion of the visible light image and the thermal image is performed, including:

[0012] Given a pair of visible light images and thermal images, a cross-modal Transformer fusion algorithm is used to extract an L-level feature pyramid by stacking convolutional layers.

[0013] The two sets of feature maps at the Lth level are input into the refinement branch to capture the global information of the source image;

[0014] The initial feature map of the current layer is concatenated with the enhanced features of the previous layer and then input into the residual feature distillation block, and the optimized feature map is generated through feature learning;

[0015] The optimized feature maps of each layer are fused using a self-fusion convolution block to obtain the fused image.

[0016] Furthermore, in the high-voltage bushing state diagnosis method of the present invention, after generating the optimized feature graph through feature learning, the method further includes:

[0017] The global context of the source image is captured by mining the complementary information across channels, and redundant information is eliminated from the feature map in the spatial domain and channel domain respectively.

[0018] Furthermore, in the high-voltage bushing status diagnosis method of the present invention, in the process of performing feature fusion on the visible light image and the thermal image, a total loss function is constructed by optimizing data loss, spatial loss, and frequency loss. The expression of the total loss function is:

[0019]

[0020] in, represents the total loss function, represents the fusion data loss function, represents the spatial loss function, represents the perceptual loss function, represents the frequency loss function, λ s ,λ p ,λ f Represents the respective hyperparameters, which are used to control the weights among the four losses.

[0021] Furthermore, in the high-voltage bushing state diagnosis method of the present invention, the step of extracting feature data from the fused image using scale-invariant feature transformation includes:

[0022] Performing multi-scale representation on the input fused image by utilizing the scale-invariant feature transformation to construct a Gaussian pyramid;

[0023] Detecting local extreme points in scale space by using Gaussian difference, optimizing parameters based on the image characteristics of the high-voltage bushing, and selecting extreme points as candidate key points by comparing each pixel with 26 points in its adjacent scale and spatial neighborhoods;

[0024] Taylor expansion is used to perform sub-pixel correction on the positions of candidate key points to eliminate unstable points;

[0025] Accurately locate the high-voltage bushing, divide the key point neighborhood into 4×4 sub-regions, calculate the gradient histogram in 8 directions for each sub-region, and form a 128-dimensional descriptor; complete feature data extraction.

[0026] Furthermore, in the high-voltage bushing status diagnosis method of the present invention, the probabilistic neural network optimized by the SALP bee colony algorithm includes:

[0027] The neurons of the probabilistic neural network receive input signals and pass the input signals to the pattern layer; in the pattern layer, the neurons are grouped by category and used to calculate the matching degree between a specific category and the input signal;

[0028] The neuron output for the kth sequence or feature and the cth category can be obtained by the following formula:

[0029]

[0030] Among them, F C1k (X) represents the neuron output value of the kth sequence or feature and the cth category, π represents the pi, α represents the smoothing limit parameter, n represents the input feature dimension, exp(·) represents the exponential function with the natural constant e as the base, X represents the input sample, X c,k represents the nuclear center of the k-th neuron of the c-th category, and x c,k ∈R n , R n Represents n-dimensional real number space, ||XX c,k || represents the difference between input sample X and training sample X c,k The Euclidean distance of

[0031] Sum the outputs of the pattern layers of the same category and use density to calculate the closest class conditional probability distribution;

[0032] The pattern vector is classified based on the maximum output of the summation unit, and the category with the highest probability is selected as the final classification result according to the Bayesian decision rule;

[0033] Taking the classification accuracy of the probabilistic neural network as the optimization target, the parameters of the SALP bee swarm algorithm are optimized, and the sea bass lead search is updated to minimize the optimal optimization variables of the loss function.

[0034] Furthermore, in the high-voltage bushing state diagnosis method of the present invention, the updated perch lead search minimizes the optimal optimization variable of the loss function, which is expressed as:

[0035]

[0036] in, represents the new position of the first leader in the jth dimension, F j Indicates the position of the food source in the jth dimension, that is, the reference value, ub j represents the upper bound of the j-th dimension, lb jrepresents the lower bound of the j-th dimension variable, c2 and c3 represent random numbers between 0 and 1, which play the role of adjusting the position in the update formula, and c1 represents the disturbance intensity coefficient, which decreases adaptively with the increase of the number of iterations. e represents a natural constant, t represents the number of iterations, t max Indicates the total number of preset iterations; for the position of the follower Then there is, represents the updated position of the i-th follower in the j-th dimension, represents the position of the previous follower in the j-th dimension.

[0037] Furthermore, in the high-voltage bushing state diagnosis method of the present invention, after the updating of the perch lead search for the optimal optimization variable that minimizes the loss function, the method further includes:

[0038] The optimal optimization variables output by the SALPS bee colony algorithm are substituted into the model corresponding to the probabilistic neural network to complete the optimization of the probabilistic neural network.

[0039] Furthermore, in the high-voltage bushing status diagnosis method of the present invention, the mechanical fault includes surface contamination, shed corrosion, and shed damage;

[0040] The electrical faults include dielectric loss, poor contact, oil leakage and partial discharge.

[0041] In a second aspect, the present invention provides a high-voltage bushing status diagnosis system, comprising:

[0042] An image acquisition module, used for acquiring visible light images and thermal images of the high-voltage bushing;

[0043] An image fusion module, configured to perform feature fusion on the visible light image and the thermal image to obtain a fused image;

[0044] A feature data acquisition module, configured to extract feature data from the fused image based on a scale-invariant feature transformation;

[0045] A status diagnosis module is used to input the feature data into a probabilistic neural network optimized by the SALP bee colony algorithm, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing status represented by the fused image; the diagnosis result includes normal, mechanical fault and electrical fault.

[0046] Compared with the prior art, the beneficial effects of the high-voltage bushing status diagnosis system of the present invention are the same as those of the high-voltage bushing status diagnosis method described in the above technical solution, and are not described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0048] Figure 1 Schematic diagram of the cross-modal Transformer fusion algorithm architecture implemented in the present invention by capturing global interactions in the spatial and channel domains;

[0049] Figure 2 Schematic diagram of the self-fusion convolutional block architecture in the cross-modal Transformer fusion algorithm of the present invention;

[0050] Figure 3 Schematic diagram of the hierarchical architecture of the probabilistic neural network optimized based on the SALP bee colony algorithm in the present invention;

[0051] Figure 4 The figure is a flow chart of a high voltage bushing status diagnosis method in the present invention. DETAILED DESCRIPTION

[0052] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] It should be noted that when an element is referred to as being “fixed on” or “disposed on” another element, it may be directly on the other element or indirectly on the other element. When an element is referred to as being “connected to” another element, it may be directly connected to the other element or indirectly connected to the other element.

[0054] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined. "Several" means one or more, unless otherwise specifically defined.

[0055] In recent years, deep learning methods, such as convolutional neural networks (CNNs) and generative adversarial networks (GANs), have significantly improved fusion performance in high-voltage bushing diagnosis through semantic feature extraction. Specifically, CNNs can identify patterns associated with micron-scale cracks and local overheating, but they require deep stacking to expand the receptive field for feature extraction, and underlying information in the source image (such as edges and line structures) may be lost during the convolution process. While GANs preserve the authenticity of thermal distribution, they can lead to low fault diagnosis accuracy due to network training instability. Furthermore, Transformer technology can achieve global context modeling through self-attention mechanisms, correlating cross-regional coordinated changes in temperature fields and texture features (such as causal analysis between shed temperature rise and adjacent porcelain corrosion). However, its high computational complexity makes it difficult to meet the real-time processing requirements of high-resolution images, and its accuracy in high-voltage bushing fault diagnosis remains low. It can be seen that, precisely because of these shortcomings, the existing technologies have low accuracy in high-voltage bushing condition diagnosis and are unable to meet practical diagnostic requirements.

[0056] In order to solve the above technical problems, the present invention provides a high-voltage bushing status diagnosis method, comprising:

[0057] Acquire visible light and thermal images of high-voltage bushings;

[0058] Performing feature fusion on the visible light image and the thermal image to obtain a fused image;

[0059] extracting feature data from the fused image using a scale-invariant feature transform;

[0060] The feature data is input into a probabilistic neural network optimized by the SALP bee colony algorithm, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing state represented by the fused image; the diagnosis result includes normal, mechanical fault and electrical fault.

[0061] When the above technical solution is adopted, in the high-voltage bushing state diagnosis method of the present invention, after obtaining the visible light image and thermal image of the high-voltage bushing, feature fusion processing is performed to obtain the fused image, and after extracting feature data in the fused image using the scale-invariant feature transform (SIFT), the feature data is input into the probabilistic neural network optimized by the SALPS bee colony algorithm to obtain the diagnosis result of the high-voltage bushing state; specifically, the feature fusion process in the above technical solution of the present invention can be implemented using a cross-modal Transformer fusion algorithm, which can capture global interactions in the spatial and channel domains through spatial transformers and channel transformers; further, by optimizing the probabilistic neural network (PNN) parameters based on the SALPS bee colony algorithm, the correlation between fault features and classification results can be effectively enhanced, and the accuracy of the PNN model for high-voltage bushing fault diagnosis can be improved. At the same time, the PNN model can not only effectively analyze the fault state, but also accelerate the precise positioning of the cause of the high-voltage bushing abnormality through its ideal model size. Its rapid response characteristics can meet the needs of intelligent real-time inspection. The above technical solution of the present invention solves the technical problem of low accuracy in diagnosing the status of high-voltage bushings in existing diagnostic methods.

[0062] In order to better understand the present invention, the content of the present invention is further explained below in conjunction with specific examples, but the content of the present invention is not limited to the following examples.

[0063] Example 1

[0064] This embodiment provides a method for diagnosing a high-voltage bushing status, including:

[0065] Step 1: Acquire a visible light image and a thermal image of the high-voltage bushing;

[0066] Step 2: performing feature fusion on the visible light image and the thermal image to obtain a fused image;

[0067] Step 3, extracting feature data from the fused image using Scale Invariant Feature Transform (SIFT);

[0068] Step 4: Using a probabilistic neural network optimized by the SALP bee colony algorithm, the feature data is input into the probabilistic neural network, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing state represented by the fusion image; the diagnosis result includes three types of states: normal, mechanical fault and electrical fault.

[0069] Example 2

[0070] This embodiment provides a method for diagnosing a high-voltage bushing status, including:

[0071] S100, acquiring a visible light image and a thermal image of the high-voltage bushing;

[0072] S200, performing feature fusion on the visible light image and the thermal image to obtain a fused image;

[0073] Furthermore, the performing feature fusion on the visible light image and the thermal image includes:

[0074] S201, given a pair of visible light image and thermal image, refer to Figure 1 , using the cross-modal Transformer fusion algorithm, an L-level feature pyramid is extracted by stacking convolutional layers;

[0075] S202, inputting the two sets of feature maps at the Lth level into the refinement branch to capture the global information of the source image;

[0076] S203, concatenating the initial feature map of the current level with the enhanced features of the previous level and inputting them into the residual feature distillation block, and generating optimized features through feature learning;

[0077] S204 captures the global context of the source image by mining cross-channel complementary information, removes redundant information from the feature maps in both the spatial and channel domains, and further enhances their complementarity. Two visible features and two infrared features are extracted. This embodiment merges these two features to convey important information and improves the interactivity within the spatial channel attention, so that a single 1×1 convolutional layer can be used for further fusion.

[0078] S205, see Figure 2 , using the self-fusion convolution block to fuse the optimized feature maps of each layer to obtain the fused image;

[0079] S206, in the process of performing feature fusion on the visible light image and the thermal image, a total loss function is constructed by optimizing data loss, spatial loss, and frequency loss. The expression of the total loss function is: in, represents the total loss function, represents the fusion data loss function, represents the spatial loss function, represents the perceptual loss function, represents the frequency loss function, λ s ,λ p ,λ f Represents the respective hyperparameters, which are used to control the weights among the four losses;

[0080] S300, extracting feature data from the fused image using a scale-invariant feature transform (SIFT);

[0081] Furthermore, the content of step S300 includes:

[0082] S301, performing multi-scale representation on the input fused image using the scale-invariant feature transform (SIFT) to construct a Gaussian pyramid;

[0083] S302, detecting local extreme points in the scale space by using Gaussian difference, optimizing parameters based on the image characteristics of the high-voltage bushing, and selecting extreme points as candidate key points by comparing each pixel point with its 26 adjacent scale and spatial neighborhood points;

[0084] S303, using Taylor expansion to perform sub-pixel correction on the candidate key point positions to eliminate unstable points;

[0085] S304: Accurately locate the high-voltage bushing, divide the key point neighborhood into 4×4 sub-regions, calculate the gradient histogram in 8 directions for each sub-region, and form a 128-dimensional descriptor; complete feature data extraction;

[0086] S400, using a probabilistic neural network optimized by the SALP bee colony algorithm, inputting the feature data into the probabilistic neural network, and the probabilistic neural network outputting a diagnosis result of the high-voltage bushing state represented by the fused image; the diagnosis result includes normal, mechanical fault, and electrical fault; the mechanical fault includes surface contamination, shed corrosion, and shed damage; the electrical fault includes dielectric loss, poor contact, oil leakage, and partial discharge;

[0087] Furthermore, in step S400, the probabilistic neural network optimized by the SALP bee colony algorithm includes:

[0088] S401, the neurons of the probabilistic neural network receive input signals and transmit the input signals to the pattern layer; in the pattern layer, the neurons are grouped by category and used to calculate the matching degree between a specific category and the input signal;

[0089] S402, the neuron output for the kth sequence or feature and the cth category can be obtained by the following formula:

[0090] Among them, F C1k (X) represents the neuron output value of the kth sequence or feature and the cth category, π represents the pi, α represents the smoothing limit parameter, n represents the input feature dimension, exp(·) represents the exponential function with the natural constant e as the base, X represents the input sample, X c,k represents the nuclear center of the k-th neuron of the c-th category, and x c,k ∈Rn , R n Represents n-dimensional real number space, ||XX c,k || represents the difference between input sample X and training sample X c,k The Euclidean distance of

[0091] S403, summing the pattern layer outputs of the same category and using density to calculate the closest class conditional probability distribution;

[0092] S404, classifying the pattern vector based on the maximum output of the summing unit, and selecting the category with the highest probability as the final classification result according to the Bayesian decision rule;

[0093] S405, taking the classification accuracy of the probabilistic neural network as the optimization target, performing SALP bee swarm algorithm parameter optimization, and updating the sea bass lead search to minimize the optimal optimization variable of the loss function;

[0094] S406, substituting the optimal optimization variables output by the SALP bee colony algorithm into the model corresponding to the probabilistic neural network to complete the optimization of the probabilistic neural network;

[0095] Furthermore, the updated perch lead searches for the optimal optimization variables that minimize the loss function, expressed as:

[0096]

[0097] in, represents the new position of the first leader in the jth dimension, F j Indicates the position of the food source in the jth dimension, that is, the reference value, ub j represents the upper bound of the j-th dimension, lb j represents the lower bound of the j-th dimension variable, c2 and c3 represent random numbers between 0 and 1, which play the role of adjusting the position in the update formula, and c1 represents the disturbance intensity coefficient, which decreases adaptively with the increase of the number of iterations. e represents a natural constant, t represents the number of iterations, t max Indicates the total number of preset iterations; for the position of the follower Then there is, represents the updated position of the i-th follower in the j-th dimension, represents the position of the previous follower in the j-th dimension.

[0098] Example 3

[0099] In a first aspect, this embodiment provides a method for diagnosing a high-voltage bushing status, comprising:

[0100] S100, see Figure 4 , obtain visible light images and thermal images of high voltage bushings;

[0101] Specifically, this embodiment utilizes an ov5640 camera to capture visible light images for system analysis, and a low-cost FLIR Lepton 3.5 infrared sensor to capture thermal images for system analysis. The camera provides a 160×120 pixel auto-calibrated output image in a 14-bit data packet acquired by a Raspberry Pi 4 board by establishing an SPI communication protocol between the devices.

[0102] Furthermore, after obtaining the visible light image and thermal image of the high voltage bushing, the images are preprocessed;

[0103] Specifically, the images received in 14-bit packets through the communication port of the Rasperry Pi 4 board require a pre-processing stage to remap the pixel values of the thermal image to 8-bit packets, i.e., remap the pixel values from the initial range of 0-16383 to the target range of 0-255, and perform noise reduction, contrast enhancement, and grayscale processing on the image.

[0104] S200, performing feature fusion on the visible light image and the thermal image to obtain a fused image;

[0105] Furthermore, the performing feature fusion on the visible light image and the thermal image includes:

[0106] S201, given a pair of visible light image and thermal image, refer to Figure 1 , using the cross-modal Transformer fusion algorithm, an L-level feature pyramid is extracted by stacking convolutional layers;

[0107] S202, inputting the two sets of feature maps at the Lth level into the refinement branch to capture the global information of the source image;

[0108] S203: The initial feature map of the current level is concatenated with the enhanced features of the previous level and then input into the residual feature distillation block, and the optimized feature map is generated through feature learning. and

[0109] S204, by mining the complementary information across channels to capture the global context of the source image, the feature map is processed in the spatial domain and the channel domain respectively. and Eliminate redundant information to further enhance the complementarity between the two; two visible features and two infrared features Extracted, this embodiment merges the two features to convey important information and improve the interactivity within the spatial channel attention so that a single 1×1 convolutional layer can be further fused; then:

[0110]

[0111] in, Represents the optimized feature map features of the visible light modality, Indicates 1×1 convolutional layer fusion dimensionality reduction, Represents the spatial attention feature map features of the visible light modality, Represents the channel attention feature map features of the visible light modality, ⊕ represents the splicing operation in the channel dimension, represents the optimized feature map features of the infrared modality, Represents the spatial attention feature map features of the infrared modality, Channel attention feature map features representing infrared modality;

[0112] S205, see Figure 2 , the optimized feature maps of each layer are fused using the self-fusion convolution block to obtain the fused image

[0113] S206, in the process of performing feature fusion on the visible light image and the thermal image, a total loss function is constructed by optimizing data loss, spatial loss, and frequency loss. The expression of the total loss function is: in, represents the total loss function, represents the fusion data loss function, represents the spatial loss function, represents the perceptual loss function, represents the frequency loss function, λ s ,λ p ,λ f Represents the respective hyperparameters, which are used to control the weights among the four losses.

[0114] S300, extracting feature data from the fused image using a scale-invariant feature transform (SIFT);

[0115] Furthermore, the content of step S300 includes:

[0116] S301, performing multi-scale representation on the input fused image using the scale-invariant feature transform (SIFT) to construct a Gaussian pyramid;

[0117] S302, detecting local extreme points in the scale space by using Difference of Gaussian (DoG), optimizing parameters based on the image characteristics of the high-voltage bushing, and selecting extreme points as candidate key points by comparing each pixel with 26 points in its adjacent scale and spatial neighborhoods;

[0118] S303: Using Taylor expansion, the candidate key point positions are corrected at the sub-pixel level to eliminate points with low contrast (e.g., if |D(x)| < 0.03, the point is considered to have too low contrast and the feature point may be caused by noise. D(x) represents the DoG response value of the corrected sub-pixel precise position) or unstable edge response (determined by the eigenvalue of the Hessian matrix).

[0119] S304 accurately locates the high-voltage bushing, divides the key point neighborhood into 4×4 sub-regions (a total of 16 blocks), calculates the gradient histogram in 8 directions for each sub-region, and forms a 128-dimensional (16×8-dimensional) descriptor; completes feature data extraction.

[0120] S400, see Figure 3 , using a probabilistic neural network optimized by the SALP bee colony algorithm, the feature data is input into the probabilistic neural network, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing status represented by the fused image; please refer to Table 1, the diagnosis results include normal, mechanical fault, and electrical fault; the mechanical fault includes surface contamination, shed corrosion, and shed damage; the electrical fault includes dielectric loss, poor contact, oil leakage, and partial discharge;

[0121] Furthermore, in step S400, the probabilistic neural network optimized by the SALP bee colony algorithm includes:

[0122] S401, the neurons of the probabilistic neural network receive input signals and transmit the input signals to the pattern layer; in the pattern layer, the neurons are grouped by category and used to calculate the matching degree between a specific category and the input signal;

[0123] S402, the neuron output for the kth sequence or feature and the cth category can be obtained by the following formula:

[0124] Among them, F C1k (X) represents the neuron output value of the kth sequence or feature and the cth category, π represents the pi, α represents the smoothing limit parameter, n represents the input feature dimension, exp(·) represents the exponential function with the natural constant e as the base, X represents the input sample, X c,k represents the nuclear center of the k-th neuron of the c-th category, and x c,k ∈R n , R n Represents n-dimensional real number space, ||XX c,k || represents the difference between input sample X and training sample X c,k The Euclidean distance of

[0125] S403, summing the pattern layer outputs of the same category and using density to calculate the closest class conditional probability distribution;

[0126] S404, classifying the pattern vector based on the maximum output of the summing unit, and selecting the category with the highest probability as the final classification result according to the Bayesian decision rule;

[0127] S405, taking the classification accuracy of the probabilistic neural network as the optimization target, performing SALP bee swarm algorithm parameter optimization, and updating the sea bass lead search to minimize the optimal optimization variable of the loss function;

[0128] S406, substituting the optimal optimization variables output by the SALP bee colony algorithm into the model corresponding to the probabilistic neural network to complete the optimization of the probabilistic neural network;

[0129] Furthermore, the updated perch lead searches for the optimal optimization variables that minimize the loss function, expressed as:

[0130]

[0131] in, represents the new position of the first leader in the jth dimension, F j Indicates the position of the food source in the jth dimension, that is, the reference value, ub j represents the upper bound of the j-th dimension, lb j represents the lower bound of the j-th dimension variable, c2 and c3 represent random numbers between 0 and 1, which play the role of adjusting the position in the update formula, and c1 represents the disturbance intensity coefficient, which decreases adaptively with the increase of the number of iterations. e represents a natural constant, t represents the number of iterations, t max Indicates the total number of preset iterations; for the position of the follower Then there is, represents the updated position of the i-th follower in the j-th dimension, represents the position of the previous follower in the j-th dimension.

[0132] Table 1 High voltage bushing status diagnosis

[0133]

[0134] In a second aspect, this embodiment provides a high-voltage bushing status diagnosis system, including:

[0135] An image acquisition module, used for acquiring visible light images and thermal images of the high-voltage bushing;

[0136] An image fusion module, configured to perform feature fusion on the visible light image and the thermal image to obtain a fused image;

[0137] A feature data acquisition module, configured to extract feature data from the fused image based on a scale-invariant feature transform (SIFT);

[0138] A state diagnosis module is used to input the feature data into a probabilistic neural network optimized by the SALP bee colony algorithm, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing state represented by the fused image; the diagnosis result includes three types of states: normal, mechanical fault and electrical fault.

[0139] Example 4

[0140] This embodiment adopts the same high-voltage bushing status diagnosis method and diagnostic system as the above-mentioned embodiment 3. The purpose of this embodiment is to demonstrate the effectiveness of the technical solution of the present invention through a specific example. Specifically, in order to test the method proposed by the present invention, a self-built data set is used for training. In this embodiment, image blocks of size 256×256 are randomly cropped from the original image, and the RGB color space of the visible light image is converted into grayscale space. The Adam optimizer is used in the training process, with parameters set to β1=0.9, β2=0.999, and an initial learning rate of 10 -4 , the training cycle is 20; the hyperparameter λ s ,λ p and λ f are fixed to 0.8, 0.02, and 0.05, respectively, to balance the best overall subjective quality and training stability under the spatial information priority constraint. The weight coefficient ω inf and ω vis Both are set to 0.5 to ensure balanced contributions of infrared and visible light features in the fusion process; the above important weight parameter values are shown in Table 2 below.

[0141] Table 2 Important weight parameter values

[0142]

[0143] Furthermore, in order to verify the effectiveness of the image fusion method mentioned in this invention, it is compared with the traditional fusion methods GAN, RFN-Nest, and IVFusion. For the convenience of comparison, three quality indicators are used to evaluate the image fusion performance: total edge information (Q AB / F ), multi-scale structural similarity (MS-SSIM), and sum of correlation differences (SCD). The scores of these three indicators are calculated between the source image and the fused image, and then the average is taken. A higher score indicates better performance. The final test results are shown in Table 3 below. The CMTs in Table 3 represents the cross-modal Transformer fusion algorithm in the present invention.

[0144] Table 3 Test results

[0145]

[0146]

[0147] From Table 3 above, we can see that the fusion performance on the quantitative comparison dataset is AB / F Indicators, the cross-modal Transformer fusion algorithm (CMTs) mentioned in the present invention scored the highest, indicating that the present invention effectively retains the complementary information in the source image; in addition, by retaining the structure and fine texture in the source image, it outperforms other algorithms in the fidelity-based indicators MS-SSIM and SCD, achieving the best overall performance.

[0148] Furthermore, in order to verify the effectiveness of the probabilistic neural network (PNN) classification module optimized by the SALPS bee colony algorithm mentioned in the present invention, it is compared with the traditional algorithms SVM and ANN, and the data representation is shown in Table 4 below.

[0149] Table 4 Data characterization

[0150]

[0151] As can be seen in Table 4, the PNN classifier outperforms both ANN and SVM classifiers in both accuracy and response speed. The present invention optimizes PNN parameters based on the SALPS bee colony algorithm, effectively strengthening the correlation between fault characteristics and classification results, improving the accuracy of the PNN model for high-voltage bushing fault diagnosis. Furthermore, the PNN model not only effectively analyzes fault conditions but also accelerates the precise location of the cause of high-voltage bushing anomalies through its ideal model size. Its rapid response meets the requirements of intelligent real-time inspections. The PNN model in this invention can be a highly effective model for high-voltage bushing image analysis and state classification tasks.

[0152] In the description of the above embodiments, specific features, structures, materials or characteristics may be combined in an appropriate manner in any one or more embodiments or examples.

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for diagnosing the status of a high-voltage bushing, characterized in that: include: Acquire visible light and thermal images of high-voltage bushings; Performing feature fusion on the visible light image and the thermal image to obtain a fused image; extracting feature data from the fused image using a scale-invariant feature transform; The feature data is input into a probabilistic neural network optimized by the SALP bee colony algorithm, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing state represented by the fused image; the diagnosis result includes normal, mechanical fault and electrical fault.

2. The high voltage bushing status diagnosis method according to claim 1, characterized in that: The performing feature fusion on the visible light image and the thermal image includes: Given a pair of visible light images and thermal images, a cross-modal Transformer fusion algorithm is used to extract an L-level feature pyramid by stacking convolutional layers. The two sets of feature maps at the Lth level are input into the refinement branch to capture the global information of the source image; The initial feature map of the current layer is concatenated with the enhanced features of the previous layer and then input into the residual feature distillation block, and the optimized feature map is generated through feature learning; The optimized feature maps of each layer are fused using a self-fusion convolution block to obtain the fused image.

3. The high voltage bushing status diagnosis method according to claim 2, characterized in that: After generating the optimized feature map through feature learning, it also includes: The global context of the source image is captured by mining the complementary information across channels, and redundant information is eliminated from the feature map in the spatial domain and channel domain respectively.

4. The high voltage bushing status diagnosis method according to claim 3, characterized in that: In the process of performing feature fusion on the visible light image and the thermal image, a total loss function is constructed by optimizing data loss, spatial loss, and frequency loss. The expression of the total loss function is: in, represents the total loss function, represents the fusion data loss function, represents the spatial loss function, represents the perceptual loss function, represents the frequency loss function, λ s ,λ p ,λ f Represents the respective hyperparameters, which are used to control the weights among the four losses.

5. The high voltage bushing status diagnosis method according to claim 4, characterized in that: The extracting feature data from the fused image by using scale-invariant feature transformation includes: Performing multi-scale representation on the input fused image by utilizing the scale-invariant feature transformation to construct a Gaussian pyramid; Detecting local extreme points in scale space by using Gaussian difference, optimizing parameters based on the image characteristics of the high-voltage bushing, and selecting extreme points as candidate key points by comparing each pixel with 26 points in its adjacent scale and spatial neighborhoods; Taylor expansion is used to perform sub-pixel correction on the positions of candidate key points to eliminate unstable points; Accurately locate the high-voltage bushing, divide the key point neighborhood into 4×4 sub-regions, calculate the gradient histogram in 8 directions for each sub-region, and form a 128-dimensional descriptor; complete feature data extraction.

6. The high voltage bushing status diagnosis method according to claim 5, characterized in that: The probabilistic neural network optimized by the SALP bee colony algorithm includes: The neurons of the probabilistic neural network receive input signals and pass the input signals to the pattern layer; in the pattern layer, the neurons are grouped by category and used to calculate the matching degree between a specific category and the input signal; The neuron output for the kth sequence or feature and the cth category can be obtained by the following formula: Among them, F C1k (X) represents the neuron output value of the kth sequence or feature and the cth category, π represents the pi, α represents the smoothing limit parameter, n represents the input feature dimension, exp(·) represents the exponential function with the natural constant e as the base, X represents the input sample, X c,k represents the nuclear center of the k-th neuron of the c-th category, and x c,h ∈R n , R n represents n-dimensional real space, X c,k As training samples, ||XX c,k || represents the difference between input sample X and training sample X c,k The Euclidean distance of Sum the outputs of the pattern layers of the same category and use density to calculate the closest class conditional probability distribution; The pattern vector is classified based on the maximum output of the summation unit, and the category with the highest probability is selected as the final classification result according to the Bayesian decision rule; Taking the classification accuracy of the probabilistic neural network as the optimization target, the parameters of the SALP bee swarm algorithm are optimized, and the sea bass lead search is updated to minimize the optimal optimization variables of the loss function.

7. The high voltage bushing status diagnosis method according to claim 6, characterized in that: The updated perch leads to searching for the optimal optimization variables that minimize the loss function, expressed as: in, represents the new position of the first leader in the jth dimension, F j Indicates the position of the food source in the jth dimension, that is, the reference value, ub j represents the upper bound of the j-th dimension, lb j represents the lower bound of the j-th dimension variable, c2 and c3 represent random numbers between 0 and 1, which play the role of adjusting the position in the update formula, and c1 represents the disturbance intensity coefficient, which decreases adaptively with the increase of the number of iterations. e represents a natural constant, t represents the number of iterations, t max Indicates the total number of preset iterations; for the position of the follower Then there is, represents the updated position of the i-th follower in the j-th dimension, represents the position of the previous follower in the j-th dimension.

8. The high voltage bushing status diagnosis method according to claim 7, characterized in that: After the updated perch leads to searching for the optimal optimization variables that minimize the loss function, it also includes: The optimal optimization variables output by the SALPS bee colony algorithm are substituted into the model corresponding to the probabilistic neural network to complete the optimization of the probabilistic neural network.

9. The high voltage bushing status diagnosis method according to claim 1, characterized in that: The mechanical failures include surface contamination, shed corrosion and shed damage; The electrical faults include dielectric loss, poor contact, oil leakage and partial discharge.

10. A high voltage bushing status diagnosis system, characterized in that: include: An image acquisition module, used for acquiring visible light images and thermal images of the high-voltage bushing; An image fusion module, configured to perform feature fusion on the visible light image and the thermal image to obtain a fused image; A feature data acquisition module, configured to extract feature data from the fused image based on a scale-invariant feature transformation; A status diagnosis module is used to input the feature data into a probabilistic neural network optimized by the SALP bee colony algorithm, and the probabilistic neural network outputs a diagnosis result of the high-voltage bushing status represented by the fused image; the diagnosis result includes normal, mechanical fault and electrical fault.

Images