GIS disconnecting switch on-off position judgment method based on visual identification

Abstract

The invention discloses a GIS isolation switch on-off position judgment method based on visual identification, and the method comprises the following steps: S1, collecting a video image frame, and obtaining RGB format image input; s2, identifying a metal reflective area, and generating a reflective mask graph and a material guide weight graph; s3, constructing an image pair, and inputting the image pair into a contrast attention module to generate a significance response graph; s4, geometric deformation parameters are calculated in combination with the guide map and the image features, and candidate regions of the image blocks are determined; s5, the candidate region is input into an improved Generative Patch Priors model, and a hidden space feature vector is generated; s6, inputting a state recognition network, and outputting an opening and closing state label and a confidence score; s7, feeding back a scoring result to a generation adjustment module, and updating image processing parameters; and S8, repeatedly executing the process to generate a complete state judgment sequence. According to the invention, accurate identification and intelligent tracking of the on-off state of the GIS isolation switch are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent power equipment monitoring, and in particular to a method for judging the opening and closing positions of GIS disconnectors based on visual recognition. Background Art

[0002] Gas-insulated metal-enclosed switchgear (GIS) is widely used in high-voltage power transmission and transformation systems due to its compact structure, small footprint, and high reliability. As a crucial component of GIS disconnectors, accurate determination of their open and closed states is crucial for ensuring safe power system operation, enabling state perception, and intelligent operation and maintenance. Traditional GIS disconnector status determination methods rely on grounding switch auxiliary contact signals, current transformer feedback data, or mechanical displacement sensors to obtain switch position. However, these methods suffer from complex wiring, reliability dependence on hardware consistency, and accuracy degradation after long-term operation, making them difficult to adapt to complex and diverse operating scenarios.

[0003] With the development of computer vision technology, the use of image recognition to detect the status of GIS isolation switches has gradually attracted the attention of researchers. Existing visual recognition methods are mainly based on image classification or target detection algorithms, which extract texture features from images or locate the position of indicator devices to determine the switch status. However, GIS equipment structures are closed and complex, and images are often accompanied by problems such as metal reflections, occlusion interference, and high background similarity. As a result, traditional image classification methods have low accuracy in practical engineering applications. In addition, current methods have shortcomings in image segmentation strategies, state discrimination granularity, and the use of cross-frame image feedback, making it difficult to determine the continuity of dynamic video sequences.

[0004] At the same time, existing technologies generally rely on a single image or a single feature channel for state judgment, ignoring the coupling relationship between the spatial correlation between image blocks and material property characteristics. Reflective areas often interfere with the positioning of the attention mechanism, and most methods do not introduce material guidance strategies that adapt to metal environments, which restricts the model's positioning ability. Some methods attempt to introduce generative models to improve image understanding capabilities, but most of them use general structures such as standard VAE and GAN, lack structural modeling of image block generation patterns, and do not have the ability to restore semantic structures with high precision. In addition, current methods generally use static classification models and do not consider the dynamic feedback adjustment of the recognition results on the processing parameters of the next frame of the image, resulting in a lack of adaptive ability in inter-frame reasoning, affecting the stability and continuity of the discrimination.

[0005] Therefore, how to provide a method for determining the opening and closing positions of GIS disconnectors based on visual recognition is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of the present invention is to propose a method for determining the open / close position of GIS disconnectors based on visual recognition. This method integrates material property recognition, a contrast attention mechanism, image geometric deformation calculation, a state recognition network, and a generative adjustment mechanism to construct a complete state judgment process for video image frame sequences. The improved Generative Patch Priors model is explicitly used to perform structural modeling and feature generation on candidate image block regions. The improved model introduces edge structure constraints, adaptive image block fusion, and a variable-dimensional latent variable mechanism, effectively enhancing the latent space expression capability of image blocks. By inputting latent space features into a state recognition network for open / close state discrimination and combining them with confidence scoring to achieve inter-frame parameter adjustment, the method possesses the advantages of strong resistance to reflective interference, excellent feature modeling capabilities, and high inter-frame adaptability.

[0007] According to an embodiment of the present invention, a method for determining the opening and closing position of a GIS disconnector based on visual recognition includes the following steps:

[0008] S1, collect the video image frame of the GIS disconnector and obtain the current frame image in RGB format as input;

[0009] S2. Perform material property recognition on the current frame image, identify the metal reflective area in the image, generate a reflective mask map, and construct a material guide weight map;

[0010] S3: Combine the current frame image with the sample images marked as split state and combined state to form an image pair, input them into the contrastive attention module, and generate a saliency response map;

[0011] S4. Calculate the geometric deformation parameters of the image block based on the material-guided weight map and the saliency response map, combined with the local gradient and curvature features of the image, and determine the candidate region of the image block;

[0012] S5. Input the candidate image block region into the improved Generative Patch Priors model, model the image structure distribution based on the image block generation method, and output the latent space feature vector corresponding to the image block;

[0013] S6. Input the latent space feature vector into the state recognition network, perform state classification, and output the separation and combination state label and confidence score of the current frame image;

[0014] S7, using the separation and combination state labels and confidence scores as feedback signals to generate an adjustment module, and updating the sampling density, attention response strength, and geometric deformation parameters of the image blocks in the next frame image;

[0015] S8. Repeat steps S1 to S7 to generate a judgment result of the GIS disconnector opening and closing status corresponding to the image frame sequence.

[0016] Optionally, the improved Generative Patch Priors model includes:

[0017] An image block preprocessing module is used to perform a normalization operation on each image block in the image block candidate area to construct an image block pixel vector set;

[0018] An edge map extraction unit, configured to extract an edge feature map of each image block based on an edge operator and generate an edge structure descriptor;

[0019] The structure-aware matching module is used to calculate the structural similarity value based on the edge structure descriptors between image blocks and construct the image block structural similarity matrix;

[0020] Joint feature construction module, which is used to concatenate the image block pixel vector and the corresponding structural similarity vector to form the joint modeling input;

[0021] A variable-length latent variable sampling structure is used to perform parameter compression and transformation on the joint modeling input to generate latent space feature vectors;

[0022] The latent space feature output module is used to aggregate all latent space feature vectors and serve as the input of the state recognition network.

[0023] Optionally, the generation and adjustment module includes:

[0024] State input unit, used to receive state labels and confidence scores;

[0025] a sampling density adjustment unit, configured to update a sampling density parameter in the candidate region of the image block according to the confidence score;

[0026] An attention intensity adjustment unit, used to adjust the response intensity threshold in the saliency response map;

[0027] The geometric deformation parameter adjustment unit is used to update the geometric deformation parameters of the image block, including size, rotation angle and boundary shape control value.

[0028] Optionally, the S1 specifically includes:

[0029] S11. A video acquisition device is arranged opposite the GIS disconnector to capture the status of the device and generate a video image frame sequence arranged in chronological order.

[0030] S12. Extract an image frame with a time index of t from the video image frame sequence as a current frame image. The current frame image is a three-channel color image with a size of H×W, where H represents the image height and W represents the image width.

[0031] S13, perform format standardization processing on the current frame image to construct an RGB format image matrix It ={I t (x,y,c)}, where I t (x, y, c) represents the pixel value of the current frame image at horizontal position x, vertical position y, and color channel c, x∈[1,W], y∈[1,H], c∈{R,G,B}.

[0032] Optionally, the S2 specifically includes:

[0033] S21. Calculate the brightness value L(x,y) for each pixel in the current frame image, where L(x,y) represents the brightness value at the horizontal position x and the vertical position y, x∈[1,W], y∈[1,H];

[0034] S22. Set a brightness threshold Th, aggregate pixels satisfying L(x, y)>Th, form a metal reflective area, and construct a reflective mask M(x, y)∈{0,1}, where M(x, y)=1 indicates that the pixel is in the reflective area, and M(x, y)=0 indicates that the pixel is not in the reflective area.

[0035] S23. Generate a material guidance weight map W(x,y) according to the reflection mask map, where W(x,y)∈[0,1] represents the material guidance weight of each pixel position.

[0036] Optionally, the S3 specifically includes:

[0037] S31, selecting a frame of image from the sample image sets marked as split state and combined state, respectively, and forming an image pair with the current frame of image;

[0038] S32: Input the image pair into the contrast attention module, which includes:

[0039] Feature encoding unit, used to extract the feature map F of the current frame image c (x,y) and the feature map F of the reference image r (x,y), where F c (x,y) and F r (x,y) represents the eigenvalues ​​at the horizontal position x and the vertical position y, respectively, x∈[1,W], y∈[1,H];

[0040] Feature difference calculation unit, used to calculate the difference map ΔF(x,y)=F c (x,y)-F r (x,y);

[0041] S33. Input the difference map ΔF(x,y) into the attention mapping unit to generate a saliency response map A(x,y), where A(x,y) represents the saliency response intensity at the horizontal position x and the vertical position y, x∈[1,W], y∈[1,H].

[0042] Optionally, the S4 specifically includes:

[0043] S41. Read the material guidance weight map W(x,y) and the significance response map A(x,y), and calculate the joint guidance map J(x,y):

[0044] J(x,y)=α·A(x,y)+β·(1-W(x,y));

[0045] Where x is the horizontal pixel position of the image, y is the vertical pixel position of the image, A(x,y) is the saliency response value, W(x,y) is the material guidance weight value, α and β are real weighting coefficients, x∈[1,W], y∈[1,H];

[0046] S42. Based on the current frame image, calculate the local gradient value G(x,y) and curvature response value C(x,y) of each pixel, where G(x,y) represents the grayscale gradient amplitude at the position (x,y), and C(x,y) represents the curvature response value at the position (x,y);

[0047] S43, weighted fusion of the joint guidance map J(x,y), the local gradient map G(x,y) and the curvature map C(x,y) to extract the geometric deformation parameter set P corresponding to each image block position i ={s i ,θ i ,b i}, where P i is the parameter set of the i-th image block, s i represents the size control value of the image block, θ i represents the rotation angle of the image block, b i Represents the boundary shape control value, according to the parameter set P i Generate image patch candidate regions.

[0048] Optionally, the S5 specifically includes:

[0049] S51, perform pixel normalization processing on each image block in the image block candidate area, extract the image block pixel vector, and construct a pixel vector set P = {p1, p2, ..., p N},in, represents the pixel vector of the i-th image block, d represents the image block dimension, and N represents the number of image blocks;

[0050] S52, perform edge feature extraction on each image block, generate an edge feature map based on the gradient operator, and construct an edge structure descriptor g(p i );

[0051] S53. Calculate the structural similarity value based on the edge structure descriptors between the image blocks and construct a structural similarity matrix:

[0052]

[0053] Among them, S ij ∈[0,1] represents the structural similarity between image blocks i and j, <·,·> represents the inner product, and ‖·‖ represents the bi-norm;

[0054] S54, pixel vector p i With the structural similarity vector [S i1 ,S i2 ,…,S iN ] is spliced ​​into joint input features, input into the latent variable sampling structure, and generates the latent space feature vector z i =f(p i ,S i1 ,S i2 ,…,S iN ),in, represents the latent space feature vector of image block i, f(·) represents the joint encoding function, and k represents the vector dimension;

[0055] S55. Gather all latent space feature vectors to form a latent space feature set.

[0056] Optionally, the S6 specifically includes:

[0057] S61. Obtain a set of latent space feature vectors of the image blocks output by the improved Generative Patch Priors model;

[0058] S62, constructing a state recognition network, the state recognition network including a feature splicing unit and a classification mapping unit;

[0059] S63. In the feature splicing unit, a splicing operation is performed on all latent space feature vectors according to the spatial arrangement order of the image blocks in the current frame image to generate a fusion vector.

[0060] S64: Input the fusion vector F into the classification mapping unit, perform linear transformation and normalization operations, and generate the prediction probability vector P = Softmax (W·F + b) of the separation state and the combination state, where represents the weight matrix, represents the bias vector, Indicates the probability value of the current frame image being in the split state and the combined state;

[0061] S65. Select the state with the larger value from the predicted probability vector as the separation state label, and use the value as the confidence score.

[0062] Optionally, the S7 specifically includes:

[0063] S71: Input the separation and combination state label and confidence score outputted by the current frame image into a generation adjustment module;

[0064] S72, adjusting the sampling density parameter in the image block candidate region according to the confidence score, and updating the image block sampling interval;

[0065] S73. Update the response intensity threshold in the significance response graph according to the confidence score;

[0066] S74, selecting a preset control rule according to the split / close state label, and updating the geometric deformation parameters of the image block, wherein the geometric deformation parameters include the image block size, rotation angle, and boundary shape control value;

[0067] S75 , inputting the sampling density parameter, the response strength threshold, and the geometric deformation parameter into the image block candidate region construction process of the next frame image to generate an updated image block candidate region.

[0068] The beneficial effects of the present invention are:

[0069] By constructing a GIS disconnector state recognition process for image frame sequences, this paper addresses the problem of insufficient recognition accuracy in traditional state judgment methods based on static images or hardware signals in scenarios with high reflective interference, structural occlusion, and image blur. By introducing a material attribute recognition mechanism to identify metal reflective areas and generate a material-guided weight map, the targetedness and reliability of subsequent image region screening and saliency response location are effectively improved. Combining a comparative attention mechanism to construct a saliency response map enhances the response expression of structural differences between state images, enabling the model to accurately capture the discriminant region.

[0070] During the image block modeling phase, this paper proposes an improved Generative Patch Priors model, introducing an edge structure extraction module and a structural similarity matching mechanism to construct a structurally perceptual prior relationship between image blocks. Furthermore, a variable-length latent variable sampling structure is designed to jointly model pixel features and structural features, generating discriminative latent space feature vectors and improving the accuracy of local image block representation. Compared to traditional image classification networks, this model is more suitable for image block-level structural modeling and is suitable for identifying reflective interference and structurally complex areas.

[0071] During the state recognition and control phase, a state recognition network is used to classify image frames. Based on the confidence score, an adjustment module is generated to dynamically update the sampling density, response threshold, and geometric parameters of the image blocks, enabling inter-frame adaptive adjustment during subsequent image frame processing. Through multi-stage joint modeling, from image acquisition, feature extraction, structural modeling, to state recognition and feedback adjustment, this invention significantly improves the accuracy, robustness, and continuity of GIS disconnector state recognition in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0073] Figure 1 This is a flow chart of a method for determining the opening and closing positions of GIS disconnectors based on visual recognition proposed by the present invention;

[0074] Figure 2 This is a schematic diagram of image block candidate region generation and geometric deformation calculation for a method for determining the opening and closing positions of GIS disconnectors based on visual recognition proposed by the present invention;

[0075] Figure 3 This is a schematic diagram of the structure of an improved Generative Patch Priors model for a method of determining the opening and closing positions of GIS disconnectors based on visual recognition proposed in the present invention. DETAILED DESCRIPTION

[0076] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0077] refer to Figure 1-3 A method for determining the opening and closing position of a GIS disconnector based on visual recognition comprises the following steps:

[0078] S1, collect the video image frame of the GIS disconnector and obtain the current frame image in RGB format as input;

[0079] S2. Perform material property recognition on the current frame image, identify the metal reflective area in the image, generate a reflective mask map, and construct a material guide weight map;

[0080] S3: Combine the current frame image with the sample images marked as split state and combined state to form an image pair, input them into the contrastive attention module, and generate a saliency response map;

[0081] S4. Calculate the geometric deformation parameters of the image block based on the material-guided weight map and the saliency response map, combined with the local gradient and curvature features of the image, and determine the candidate region of the image block;

[0082] S5. Input the candidate image block region into the improved Generative Patch Priors model, model the image structure distribution based on the image block generation method, and output the latent space feature vector corresponding to the image block;

[0083] S6. Input the latent space feature vector into the state recognition network, perform state classification, and output the separation and combination state label and confidence score of the current frame image;

[0084] S7, using the separation and combination state labels and confidence scores as feedback signals to generate an adjustment module, and updating the sampling density, attention response strength, and geometric deformation parameters of the image blocks in the next frame image;

[0085] S8. Repeat steps S1 to S7 to generate a judgment result of the GIS disconnector opening and closing status corresponding to the image frame sequence.

[0086] The present invention effectively realizes the automation of visual recognition of the open and close positions by collecting video image frames of the GIS disconnector and inputting the images into a specific recognition process. It overcomes the lag and misjudgment problems existing in the traditional reliance on auxiliary contacts and manual visual methods, and improves the intelligence and real-time performance of state recognition.

[0087] In this embodiment, the improved Generative Patch Priors model includes:

[0088] An image block preprocessing module is used to perform a normalization operation on each image block in the image block candidate area to construct an image block pixel vector set;

[0089] An edge map extraction unit, configured to extract an edge feature map of each image block based on an edge operator and generate an edge structure descriptor;

[0090] The structure-aware matching module is used to calculate the structural similarity value based on the edge structure descriptors between image blocks and construct the image block structural similarity matrix;

[0091] Joint feature construction module, which is used to concatenate the image block pixel vector and the corresponding structural similarity vector to form the joint modeling input;

[0092] A variable-length latent variable sampling structure is used to perform parameter compression and transformation on the joint modeling input to generate latent space feature vectors;

[0093] The latent space feature output module is used to aggregate all latent space feature vectors and serve as the input of the state recognition network.

[0094] The present invention improves the image block modeling capability of complex structure areas by setting up an improved Generative Patch Priors model that includes image block preprocessing, edge map extraction, structure-aware matching, joint modeling and latent variable sampling, and achieves higher-quality image block structure understanding and encoding, which is conducive to enhancing the accuracy and robustness of the image state recognition process.

[0095] In this embodiment, the generation and adjustment module includes:

[0096] State input unit, used to receive state labels and confidence scores;

[0097] a sampling density adjustment unit, configured to update a sampling density parameter in the candidate region of the image block according to the confidence score;

[0098] An attention intensity adjustment unit, used to adjust the response intensity threshold in the saliency response map;

[0099] The geometric deformation parameter adjustment unit is used to update the geometric deformation parameters of the image block, including size, rotation angle and boundary shape control value.

[0100] The present invention realizes the dynamic adjustment of the construction method of candidate areas in the image processing process by setting up a generation adjustment module including a state input unit, a sampling density adjustment unit, an attention intensity adjustment unit and a geometric deformation parameter adjustment unit, thereby improving the algorithm's adaptive response ability to changes in inter-frame information and enhancing the convergence and continuity of the recognition process.

[0101] In this embodiment, S1 specifically includes:

[0102] S11. A video acquisition device is arranged opposite the GIS disconnector to capture the status of the device and generate a video image frame sequence arranged in chronological order.

[0103] S12. Extract an image frame with a time index of t from the video image frame sequence as a current frame image. The current frame image is a three-channel color image with a size of H×W, where H represents the image height and W represents the image width.

[0104] S13, perform format standardization processing on the current frame image to construct an RGB format image matrix I t ={I t (x,y,c)}, where I t (x, y, c) represents the pixel value of the current frame image at horizontal position x, vertical position y, and color channel c, x∈[1,W], y∈[1,H], c∈{R,G,B}.

[0105] By introducing time index management, image format standardization, and RGB matrix construction processes into the acquisition step, the present invention ensures the temporal continuity and pixel consistency of the input image frames, providing a stable data foundation for subsequent material property and image block modeling processes.

[0106] In this embodiment, S2 specifically includes:

[0107] S21. Calculate the brightness value L(x,y) for each pixel in the current frame image, where L(x,y) represents the brightness value at the horizontal position x and the vertical position y, x∈[1,W], y∈[1,H];

[0108] S22. Set a brightness threshold Th, aggregate pixels satisfying L(x, y)>Th, form a metal reflective area, and construct a reflective mask M(x, y)∈{0,1}, where M(x, y)=1 indicates that the pixel is in the reflective area, and M(x, y)=0 indicates that the pixel is not in the reflective area.

[0109] S23. Generate a material guidance weight map W(x,y) according to the reflection mask map, where W(x,y)∈[0,1] represents the material guidance weight of each pixel position.

[0110] The present invention constructs metal reflective areas through brightness value calculation and threshold screening, and generates a material guidance weight map based on the reflective areas, so that the subsequent feature extraction module can focus on key areas with high reflective characteristics in the image, effectively enhancing the spatial attention focusing ability in the process of split and combined state recognition.

[0111] In this embodiment, S3 specifically includes:

[0112] S31, selecting a frame of image from the sample image sets marked as split state and combined state, respectively, and forming an image pair with the current frame of image;

[0113] S32: Input the image pair into the contrast attention module, which includes:

[0114] Feature encoding unit, used to extract the feature map F of the current frame image c (x,y) and the feature map F of the reference image r (x,y), where F c (x,y) and F r (x,y) represents the eigenvalues ​​at the horizontal position x and the vertical position y, respectively, x∈[1,W], y∈[1,H];

[0115] Feature difference calculation unit, used to calculate the difference map ΔF(x,y)=F c (x,y)-F r (x,y);

[0116] S33. Input the difference map ΔF(x,y) into the attention mapping unit to generate a saliency response map A(x,y), where A(x,y) represents the saliency response intensity at the horizontal position x and the vertical position y, x∈[1,W], y∈[1,H].

[0117] The present invention constructs a contrast attention module, models the feature differences between the current frame image and the reference image, and generates a saliency response map, thereby achieving a prominent representation of areas with structural differences and enhancing the sensitivity and discriminability of the recognition system under slight structural changes.

[0118] In this embodiment, the S4 specifically includes:

[0119] S41. Read the material guidance weight map W(x,y) and the significance response map A(x,y), and calculate the joint guidance map J(x,y):

[0120] J(x,y)=α·A(x,y)+β·(1-W(x,y));

[0121] Where x is the horizontal pixel position of the image, y is the vertical pixel position of the image, A(x,y) is the saliency response value, W(x,y) is the material guidance weight value, α and β are real weighting coefficients, x∈[1,W], y∈[1,H];

[0122] S42. Based on the current frame image, calculate the local gradient value G(x,y) and curvature response value C(x,y) of each pixel, where G(x,y) represents the grayscale gradient amplitude at the position (x,y), and C(x,y) represents the curvature response value at the position (x,y);

[0123] S43, weighted fusion of the joint guidance map J(x,y), the local gradient map G(x,y) and the curvature map C(x,y) to extract the geometric deformation parameter set P corresponding to each image block position i ={s i ,θ i ,b i}, where P i is the parameter set of the i-th image block, s i represents the size control value of the image block, θ i represents the rotation angle of the image block, b i Represents the boundary shape control value, according to the parameter set P i Generate image patch candidate regions.

[0124] The present invention introduces a joint fusion mechanism of material guidance map and saliency response map, and combines local gradient and curvature features to extract the geometric deformation parameters of image blocks, thereby improving the spatial accuracy and structural stability of candidate region screening and ensuring the quality consistency of subsequent feature modeling input.

[0125] In this embodiment, the S5 specifically includes:

[0126] S51, perform pixel normalization processing on each image block in the image block candidate area, extract the image block pixel vector, and construct a pixel vector set P = {p1, p2, ..., p N},in, represents the pixel vector of the i-th image block, d represents the image block dimension, and N represents the number of image blocks;

[0127] S52, perform edge feature extraction on each image block, generate an edge feature map based on the gradient operator, and construct an edge structure descriptor g(p i );

[0128] S53. Calculate the structural similarity value based on the edge structure descriptors between the image blocks and construct a structural similarity matrix:

[0129]

[0130] Among them, S ij ∈[0,1] represents the structural similarity between image blocks i and j, <·,·> represents the inner product, and ‖·‖ represents the bi-norm;

[0131] S54, pixel vector p i With the structural similarity vector [S i1 ,S i2 ,…,S iN ] is spliced ​​into joint input features, input into the latent variable sampling structure, and generates the latent space feature vector z i =f(p i ,S i1 ,S i2 ,…,S iN ),in, represents the latent space feature vector of image block i, f(·) represents the joint encoding function, and k represents the vector dimension;

[0132] S55. Gather all latent space feature vectors to form a latent space feature set.

[0133] The present invention significantly improves the ability of image block-level structural expression by setting a joint modeling input of image block pixel vectors and edge structure similarity matrices, and generates high-dimensional latent space features through a variable-length latent variable sampling structure, providing more expressive structural embedding information for the state recognition network.

[0134] In this embodiment, S6 specifically includes:

[0135] S61. Obtain a set of latent space feature vectors of the image blocks output by the improved Generative Patch Priors model;

[0136] S62, constructing a state recognition network, the state recognition network including a feature splicing unit and a classification mapping unit;

[0137] S63. In the feature splicing unit, a splicing operation is performed on all latent space feature vectors according to the spatial arrangement order of the image blocks in the current frame image to generate a fusion vector.

[0138] S64: Input the fusion vector F into the classification mapping unit, perform linear transformation and normalization operations, and generate the prediction probability vector P = Softmax (W·F + b) of the separation state and the combination state, where represents the weight matrix, represents the bias vector, Indicates the probability value of the current frame image being in the split state and the combined state;

[0139] S65. Select the state with the larger value from the predicted probability vector as the separation state label, and use the value as the confidence score.

[0140] The present invention constructs a state recognition network, performs splicing, linear mapping and normalization operations on the latent space features of image blocks, and finally outputs state labels and confidence scores, thereby achieving an effective fusion of structural-level information and full-image semantics, and improving the resolution ability and classification accuracy of state recognition.

[0141] In this embodiment, the S7 specifically includes:

[0142] S71: Input the separation and combination state label and confidence score outputted by the current frame image into a generation adjustment module;

[0143] S72, adjusting the sampling density parameter in the image block candidate region according to the confidence score, and updating the image block sampling interval;

[0144] S73. Update the response intensity threshold in the significance response graph according to the confidence score;

[0145] S74, selecting a preset control rule according to the split / close state label, and updating the geometric deformation parameters of the image block, wherein the geometric deformation parameters include the image block size, rotation angle, and boundary shape control value;

[0146] S75 , inputting the sampling density parameter, the response strength threshold, and the geometric deformation parameter into the image block candidate region construction process of the next frame image to generate an updated image block candidate region.

[0147] The present invention effectively constructs an inter-frame feedback closed-loop path by feeding back the recognition result of the previous frame to the generation and adjustment module, and dynamically updates the sampling density, attention response intensity and geometric deformation parameters, thereby achieving continuous optimization of the recognition process and enhancing the stability and adaptability of the overall system.

[0148] Example 1:

[0149] To verify the feasibility of this invention, it was deployed at a 500kV substation in East China. This substation houses 26 three-phase GIS disconnectors, operating in a complex environment characterized by strong metal reflections, structural obstructions, and background interference. The video surveillance system, equipped with a 1080p high-definition industrial camera with a 15fps frame rate, covers key areas such as the conductive arm, operating mechanism, and switch mechanism, providing a representative image.

[0150] Before the system was deployed, traditional identification methods relied primarily on auxiliary contact signals and infrared indicator readings. However, in high-temperature and high-humidity environments, false contact activation rates increased and infrared reflection misalignment became frequent. Particularly during the rainy, foggy, and humid conditions of mid-July 2024, the auxiliary signal recognition accuracy dropped below 85%, far below the 95% or higher required for intelligent operations and maintenance.

[0151] After deploying the system, an RGB image is extracted from a video frame as input. The material attribute recognition module first detects the reflectivity of metal areas, generates a reflectivity mask, and constructs a material guidance weight map based on this. This guides the focus of salient areas on conductive structures. In sample images where the brightness of the metal reflective area is greater than 220, the positioning error is reduced from 14.3 pixels (using traditional methods) to 4.1 pixels, significantly reducing positioning deviations caused by highlight interference.

[0152] Furthermore, by pairing the current frame image with a state sample image and feeding it into the contrastive attention module to generate a saliency response map, the response features that highlight the state differences between the images were enhanced. Across 18 test images, the average discriminant strength of the response area increased by 62.5%. Furthermore, the system was able to stably locate key areas such as operating mechanisms and connectors, resolving the limitation of traditional methods that often experience inability to focus under blurred or occluded conditions.

[0153] Candidate regions for image blocks are determined through the joint screening of four types of information: material guidance map, local gradient, edge curvature, and response intensity. In the 1120-frame image test, the system initially selected an average of 5.7 candidate regions per frame, and ultimately retained 2.3 regions for subsequent recognition, effectively controlling the risk of redundant calculations and misidentification. Subsequently, the candidate image blocks were structurally modeled using the improved Generative PatchPriors model, which introduced edge structure extraction, structure matching, and latent variable sampling mechanisms to enhance the expressive power of local image blocks. In the structured occlusion test set, the state recognition accuracy increased from 71.4% of the traditional method to 89.3%; the average PSNR value of the latent space reconstructed image reached 32.7dB, much higher than the 28.1dB of the conventional convolutional reconstruction model.

[0154] After outputting the state label, the system further feeds the label and confidence score into a generated adjustment module, dynamically adjusting the next frame's sampling density, response threshold, and image block deformation parameters. During testing, the system achieved a response latency of 216ms, significantly lower than the 387ms of traditional methods. State recognition accuracy remained stable at over 96.8%, and inter-frame stability improved by 21.5%, meeting the power system's requirements for rapid response and high reliability.

[0155] In order to further clarify the differences between the present invention and the traditional method in terms of key performance indicators, the key comparative data are summarized as shown in Table 1:

[0156] Table 1 Comparison of key indicators of GIS disconnector visual recognition methods

[0157]

[0158] As shown in Table 1, the state recognition accuracy of this method improved by 12.1 percentage points compared to traditional methods, significantly reducing the false positive rate. In terms of image structure representation, the PSNR increased by 4.6dB, demonstrating the model's enhanced ability to model image details. In complex scenarios such as high reflections and structural occlusion, recognition accuracy increased by over 10%. Response speed was reduced by 171ms, improving processing efficiency. Furthermore, dynamic adjustment of inter-frame continuity enhanced the algorithm's adaptability and recognition stability.

[0159] In summary, this embodiment verifies the adaptability, recognition accuracy, and response performance of the present invention in the complex image environment of GIS disconnectors. It has high practicality and engineering deployability and is suitable for key equipment status identification tasks in remote intelligent monitoring of power systems.

[0160] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for determining the opening and closing positions of GIS disconnectors based on visual recognition, characterized in that: The steps include: S1, collect the video image frame of the GIS disconnector and obtain the current frame image in RGB format as input; S2. Perform material property recognition on the current frame image, identify the metal reflective area in the image, generate a reflective mask map, and construct a material guide weight map; S3: Combine the current frame image with the sample images marked as split state and combined state to form an image pair, input them into the contrastive attention module, and generate a saliency response map; S4. Calculate the geometric deformation parameters of the image block based on the material-guided weight map and the saliency response map, combined with the local gradient and curvature features of the image, and determine the candidate region of the image block; S5. Input the candidate image block region into the improved Generative Patch Priors model, model the image structure distribution based on the image block generation method, and output the latent space feature vector corresponding to the image block; S6. Input the latent space feature vector into the state recognition network, perform state classification, and output the separation and combination state label and confidence score of the current frame image; S7, using the separation and combination state labels and confidence scores as feedback signals to generate an adjustment module, and updating the sampling density, attention response strength, and geometric deformation parameters of the image blocks in the next frame image; S8. Repeat steps S1 to S7 to generate a judgment result of the GIS disconnector opening and closing status corresponding to the image frame sequence.

2. A method for determining the opening and closing position of a GIS disconnector based on visual recognition according to claim 1, characterized in that: The improved Generative Patch Priors model includes: An image block preprocessing module is used to perform a normalization operation on each image block in the image block candidate area to construct an image block pixel vector set; An edge map extraction unit, configured to extract an edge feature map of each image block based on an edge operator and generate an edge structure descriptor; The structure-aware matching module is used to calculate the structural similarity value based on the edge structure descriptors between image blocks and construct the image block structural similarity matrix; Joint feature construction module, which is used to concatenate the image block pixel vector and the corresponding structural similarity vector to form the joint modeling input; A variable-length latent variable sampling structure is used to perform parameter compression and transformation on the joint modeling input to generate latent space feature vectors; The latent space feature output module is used to aggregate all latent space feature vectors and serve as the input of the state recognition network.

3. The method for determining the opening and closing position of a GIS disconnector based on visual recognition according to claim 1 is characterized in that: The generation and adjustment module includes: State input unit, used to receive state labels and confidence scores; a sampling density adjustment unit, configured to update a sampling density parameter in the candidate region of the image block according to the confidence score; An attention intensity adjustment unit, used to adjust the response intensity threshold in the saliency response map; The geometric deformation parameter adjustment unit is used to update the geometric deformation parameters of the image block, including size, rotation angle and boundary shape control value.

4. The method for determining the opening and closing positions of GIS disconnectors based on visual recognition according to claim 1 is characterized in that: Said S1 specifically includes: S11. A video acquisition device is arranged opposite the GIS disconnector to capture the status of the device and generate a video image frame sequence arranged in chronological order. S12. Extract an image frame with a time index of t from the video image frame sequence as a current frame image. The current frame image is a three-channel color image with a size of H×W, where H represents the image height and W represents the image width. S13, perform format standardization processing on the current frame image to construct an RGB format image matrix I t ={I t (x,y,c)}, where I t (x, y, c) represents the pixel value of the current frame image at horizontal position x, vertical position y, and color channel c, x∈[1,W], y∈[1,H], c∈{R,G,B}.

5. The method for determining the opening and closing positions of GIS disconnectors based on visual recognition according to claim 1 is characterized in that: The S2 specifically includes: S21. Calculate the brightness value L(x,y) for each pixel in the current frame image, where L(x,y) represents the brightness value at the horizontal position x and the vertical position y, x∈[1,W], y∈[1,H]; S22. Set a brightness threshold Th, aggregate pixels satisfying L(x, y)>Th, form a metal reflective area, and construct a reflective mask M(x, y)∈{0,1}, where M(x, y)=1 indicates that the pixel is in the reflective area, and M(x, y)=0 indicates that the pixel is not in the reflective area. S23. Generate a material guidance weight map W(x,y) according to the reflection mask map, where W(x,y)∈[0,1] represents the material guidance weight of each pixel position.

6. A method for determining the opening and closing position of a GIS disconnector based on visual recognition according to claim 1, characterized in that: The S3 specifically includes: S31, selecting a frame of image from the sample image sets marked as split state and combined state, respectively, and forming an image pair with the current frame of image; S32: Input the image pair into the contrast attention module, which includes: Feature encoding unit, used to extract the feature map F of the current frame image c (x,y) and the feature map F of the reference image r (x,y), where F c (x,y) and F r (x,y) represents the eigenvalues ​​at the horizontal position x and the vertical position y, respectively, x∈[1,W], y∈[1,H]; Feature difference calculation unit, used to calculate the difference map ΔF(x,y)=F c (x,y)-F r (x,y); S33. Input the difference map ΔF(x,y) into the attention mapping unit to generate a saliency response map A(x,y), where A(x,y) represents the saliency response intensity at the horizontal position x and the vertical position y, x∈[1,W], y∈[1,H].

7. The method for determining the opening and closing positions of GIS disconnectors based on visual recognition according to claim 1 is characterized in that: The S4 specifically includes: S41. Read the material guidance weight map W(x,y) and the significance response map A(x,y), and calculate the joint guidance map J(x,y): J(x,y)=α·A(x,y)+β·(1-W(x,y)); Where x is the horizontal pixel position of the image, y is the vertical pixel position of the image, A(x,y) is the saliency response value, W(x,y) is the material guidance weight value, α and β are real weighting coefficients, x∈[1,W], y∈[1,H]; S42. Based on the current frame image, calculate the local gradient value G(x,y) and curvature response value C(x,y) of each pixel, where G(x,y) represents the grayscale gradient amplitude at the position (x,y), and C(x,y) represents the curvature response value at the position (x,y); S43, weighted fusion of the joint guidance map J(x,y), the local gradient map G(x,y) and the curvature map C(x,y) to extract the geometric deformation parameter set P corresponding to each image block position i ={s i ,θ i ,b i }, where P i is the parameter set of the i-th image block, s i represents the size control value of the image block, θ i represents the rotation angle of the image block, b i Represents the boundary shape control value, according to the parameter set P i Generate image patch candidate regions.

8. The method for determining the opening and closing positions of GIS disconnectors based on visual recognition according to claim 1 is characterized in that: The S5 specifically includes: S51, perform pixel normalization processing on each image block in the image block candidate area, extract the image block pixel vector, and construct a pixel vector set P = {p1, p2, ..., p N },in, represents the pixel vector of the i-th image block, d represents the image block dimension, and N represents the number of image blocks; S52, perform edge feature extraction on each image block, generate an edge feature map based on the gradient operator, and construct an edge structure descriptor g(p i ); S53. Calculate the structural similarity value based on the edge structure descriptors between the image blocks and construct a structural similarity matrix: Among them, S ij ∈[0,1] represents the structural similarity between image blocks i and j, <·,·> represents the inner product, and ‖·‖ represents the bi-norm; S54, pixel vector p i With the structural similarity vector [S i1 ,S i2 ,…,S iN ] is spliced ​​into joint input features, input into the latent variable sampling structure, and generates the latent space feature vector z i =f(p i ,S i1 ,S i2 ,…,S iN ),in, represents the latent space feature vector of image block i, f(·) represents the joint encoding function, and k represents the vector dimension; S55. Gather all latent space feature vectors to form a latent space feature set.

9. The method for determining the opening and closing positions of GIS disconnectors based on visual recognition according to claim 1 is characterized in that: The S6 specifically includes: S61. Obtain a set of latent space feature vectors of the image blocks output by the improved Generative Patch Priors model; S62, constructing a state recognition network, the state recognition network including a feature splicing unit and a classification mapping unit; S63. In the feature splicing unit, a splicing operation is performed on all latent space feature vectors according to the spatial arrangement order of the image blocks in the current frame image to generate a fusion vector. S64: Input the fusion vector F into the classification mapping unit, perform linear transformation and normalization operations, and generate the prediction probability vector P = Softmax (W·F + b) of the separation state and the combination state, where represents the weight matrix, represents the bias vector, Indicates the probability value of the current frame image being in the split state and the combined state; S65. Select the state with the larger value from the predicted probability vector as the separation state label, and use the value as the confidence score.

10. The method for determining the opening and closing positions of GIS disconnectors based on visual recognition according to claim 1, characterized in that: The S7 specifically includes: S71: Input the separation and combination state label and confidence score outputted by the current frame image into a generation adjustment module; S72, adjusting the sampling density parameter in the image block candidate region according to the confidence score, and updating the image block sampling interval; S73. Update the response intensity threshold in the significance response graph according to the confidence score; S74, selecting a preset control rule according to the split / close state label, and updating the geometric deformation parameters of the image block, wherein the geometric deformation parameters include the image block size, rotation angle, and boundary shape control value; S75 , inputting the sampling density parameter, the response strength threshold, and the geometric deformation parameter into the image block candidate region construction process of the next frame image to generate an updated image block candidate region.

Images