A substation pressboard online detection method and system based on trusted WLAN communication

Abstract

The application discloses a substation pressure plate online detection method and system based on trusted WLAN communication and relates to the technical field of substation detection; the method comprises the following steps: collecting an original image of a target pressure plate, performing enhancement on the original image to obtain an enhanced image, inputting the enhanced image into a state detection model to obtain a state detection result, obtaining operation data of a target relay protection circuit corresponding to the state detection result according to an association mapping relationship, and performing consistency verification on the state detection result according to the operation data to obtain a trusted state result; through the collection of the pressure plate image and the enhancement processing, the illumination and reflection interference is effectively weakened, the image quality is improved, the recognition accuracy of the pressure plate on-off, pose and defects is improved, the consistency verification is performed by fusing the relay protection circuit operation data, the misjudgment defects of pure visual detection are avoided, the result credibility is improved, the trusted WLAN is used for safe transmission to the cloud, the real-time monitoring of the pressure plate state is realized, and the intelligence and reliability of the substation operation and maintenance are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of substation testing technology, specifically relating to an online testing method and system for substation pressure plates based on trusted WLAN communication. Background Technology

[0002] Substations are the core hubs in the power grid, transforming voltage, distributing electrical energy, and regulating power transmission, maintaining the stable operation of the regional power system. The control plate is a key secondary control element of the substation's relay protection device; its activation / deactivation, position, and defect status directly determine the reliability of the protection device's operation. Abnormal states can easily lead to maloperation or failure to operate, threatening power grid safety. Therefore, online monitoring of the control plate allows for real-time and accurate monitoring of its activation / deactivation status, timely troubleshooting of anomalies, prevention of maloperation or failure to operate, ensuring reliable relay protection operation, improving the intelligence level of substation operation and maintenance, and effectively reducing potential safety hazards in power system operation.

[0003] In existing substation switchgear image detection technologies, a single image recognition method is typically used to detect the switchgear status, achieving rapid detection under ideal conditions. However, existing detection methods have significant shortcomings: First, the uneven lighting, metal reflection, and shadow interference within substation switchgear make it difficult for existing image enhancement methods to simultaneously suppress illumination and preserve details, easily leading to the loss of text and edge features on the switchgear. Second, existing deep learning models have limited feature extraction capabilities, failing to extract fine-grained features for dense, small switchgear, and exhibiting poor multi-scale adaptability, resulting in low detection accuracy. Third, relying solely on single-source visual data to determine the switchgear status without verification using operational data from related equipment makes it susceptible to image noise and occlusion interference, leading to misjudgments and insufficient reliability of the detection results, resulting in low accuracy in switchgear status detection. Summary of the Invention

[0004] The purpose of this invention is to solve the problems that most existing pressure plate detection methods use a single image recognition detection method, which is difficult to adapt to the complex interference in switch cabinets, lacks relevant equipment operation data for auxiliary verification, and is easily misjudged by image noise and obstruction interference, resulting in low accuracy of pressure plate status detection. Therefore, this invention proposes an online detection method and system for substation pressure plates based on reliable WLAN communication.

[0005] In a first aspect of this invention, a method for online detection of substation pressure plates based on trusted WLAN communication is first proposed, the method comprising: The system periodically acquires original images of the target pressure plate, performs enhancement operations on the original images, and outputs enhanced images; the original images are RGB images. The enhanced image is input into the pressure plate state detection model to obtain the state detection result; the state detection model is based on YOLOv11 and is improved by replacing C3K2 in the YOLOv11 backbone network with the first target module; the state detection result includes at least one of the pressure plate deployment / retraction state, pressure plate pose state, and pressure plate defect state; Based on the association mapping relationship, the operating data of the target relay protection circuit corresponding to the state detection result is obtained, and the state detection result is verified by association consistency based on the operating data to obtain the reliable state result of the pressure plate; the association mapping relationship is the relationship between the substation pressure plate and the corresponding relay protection circuit that is constructed in advance. The trusted status results are sent to the cloud monitoring platform for online monitoring via a trusted WLAN.

[0006] The beneficial effects of this invention are: This invention proposes an online detection method for substation pressure plates based on trusted WLAN communication. By periodically acquiring pressure plate images and performing enhancement processing, environmental interference such as substation illumination and reflection is effectively reduced, improving the quality of the images to be detected and ensuring the accuracy of the detection model in recognizing pressure plate engagement / disengagement, pose, and defect status. By combining relay protection circuit operation data for correlation consistency verification, the method avoids the misjudgment problem of single visual detection and improves the reliability of detection results. Data is securely transmitted to the cloud platform via trusted WLAN to achieve online real-time monitoring of pressure plate status, thereby enhancing the reliability and intelligence level of substation operation and maintenance management.

[0007] Optionally, performing enhancement operations on the original image to output an enhanced image includes: The original image is converted into an HSV color image and used as the first image; The brightness channel V in the first image is extracted, and the brightness channel V is decomposed into a first component and a second component through a preset physical model. The first component is a low-frequency structure used to characterize the lighting at the substation site, and the second component is high-frequency information used to characterize the details of the pressure plate body. First constraint function and second constraint function are constructed for the first component and the second component respectively; the first constraint function is used to constrain the first component to smooth the uneven lighting, shadows and metallic reflections of the original image, and to preserve the edge of the pressure plate structure; the second constraint function is used to constrain the second component to clearly extract and preserve the edge of the pressure plate and text markings, and to prevent details from becoming blurred or lost. A third constraint function for the bright channel is constructed by applying illumination constraints to the original image; the third constraint function is used to constrain the first component to be close to the standard bright channel illumination, so as to avoid overexposure and over-brightness due to deviation from the standard illumination. An enhanced image is obtained by performing preset operations based on the first constraint function, the second constraint function, and the third constraint function.

[0008] By transforming the color space to remove color interference, the brightness channel is decomposed into two components representing ambient light and the details of the pressure plate itself. A triple constraint function is constructed to achieve smooth noise reduction of light, accurate preservation of details, and reasonable control of light, which effectively suppresses interference such as uneven lighting in substations, metal reflection, and shadows, preserves the edges of the pressure plate and text markings, and significantly optimizes image quality.

[0009] Optionally, the first constraint function and the second constraint function are constructed as follows: Centered on each pixel p in the first component, a first local window of fixed size is defined, and horizontal and vertical gradient operators are defined. For each pixel q within the first local window, calculate the spatial weight. Traverse all pixels of the first component and calculate the total variation of the horizontal and vertical windows respectively. Using the luminance channel V as the guide image and the first component as the image to be filtered, a guide filtering operation is performed on the image to be filtered to obtain the filtered image, and its horizontal and vertical gradients are calculated. The first constraint function is calculated based on the total variation of the horizontal window, the total variation of the vertical window, the horizontal gradient, and the vertical gradient. A second local window of fixed size is defined with each pixel m in the second component as the center, and horizontal and vertical gradient operators are defined. The horizontal and vertical gradients of the second component are calculated pixel by pixel based on the horizontal and vertical gradient operators. The horizontal and vertical gradients of all pixels within the second local window are calculated to obtain the mean local gradient. Logarithmic operation is then performed on the mean local gradient to generate adaptive detail weights. The gradients in the corresponding directions are weighted according to the adaptive weights, and the second constraint function of the second component is calculated based on the weighted horizontal and vertical gradients.

[0010] The first constraint function is constructed by using local windows, gradient operators, and window total variation. Guided filtering is then used to achieve smooth noise reduction of the illumination component, effectively suppressing reflections and shadows. The second constraint function is constructed by using gradient calculation, local mean, and logarithmic adaptive weights to specifically enhance the text and edge details on the pressure plate, adaptively suppress noise, achieve differentiated constraints on illumination and details, and improve the accuracy and robustness of image optimization.

[0011] Optionally, an enhanced image is obtained by performing a preset operation based on the first constraint function, the second constraint function, and the third constraint function, including: The first constraint function, the second constraint function, and the third constraint function are weighted and fused to obtain the overall objective function. Using the first component and the second component as initial values ​​for iteration, perform iterative optimization on the total objective function to output the optimal first component and the optimal second component. Adaptive enhancement is performed on the optimal first component to output the final first component, and the final first component is fused with the optimal second component to obtain the optimal luminance channel V; The H channel, S constraint function channel, and optimal luminance channel V of the pressure plate HSV color image are fused to obtain the enhanced constraint function HSV pressure plate image. The enhanced constraint function HSV pressure plate image is then converted into RGB format to obtain the enhanced image of the pressure plate.

[0012] The overall objective function is constructed by weighted fusion of three types of constraint functions. The optimal illumination and detail components are solved by iterative optimization to ensure that the constraints work together. The illumination component is adaptively enhanced and then fused with the detail component to optimize the brightness channel. The image is then reconstructed by combining the original hue and saturation channels, realizing the integration of illumination correction and detail enhancement. This effectively avoids component solution deviations and outputs a high-quality enhanced image of the pressure plate with true color and complete details.

[0013] Optionally, the state detection model is an improvement based on YOLOv11, and the improvements include: All C3K2 modules in the YOLOv11 model backbone network are replaced with the first target module; the first target module is used to enhance the fine-grained feature extraction and multi-scale adaptation capability of dense micro pressure plates in substation switchgear by reparameterizing the heterogeneous multi-scale cascade structure and fusing cross-branch features. The C3K2 modules in the neck network of the YOLOv11 model are all replaced with the second target module. The second target module expands the receptive field and strengthens multi-scale feature fusion while reducing the number of parameters and computational overhead by fusing dual convolutions and dilated grouped shuffling convolutions in stages.

[0014] The backbone network replacement module enhances fine-grained feature extraction and multi-scale adaptation of dense micro pressure plates, while the neck network replacement module expands the receptive field and improves multi-scale feature fusion capabilities while being lightweight and reducing computing power, effectively improving the accuracy of pressure plate state detection and inference efficiency.

[0015] Optionally, the working principle of the first target module is as follows: Obtain the first input feature tensor, split the first input feature tensor according to the number of channels, and convolve the split feature tensors through 4 parallel Conv modules to obtain the first feature tensor, the second feature tensor, the third feature tensor, and the fourth feature tensor respectively. The first feature tensor and the second feature tensor are fused and input into the enhancement module to obtain the first enhanced feature tensor. The first enhanced feature tensor is then input into the enhancement module to obtain the second enhanced feature tensor. The two enhancement modules are connected in series. The first and third enhanced feature tensors are fused and input into the enhancement module to obtain the third enhanced feature tensor, and the third enhanced feature tensor is input into the enhancement module to obtain the fourth enhanced feature tensor; The third and fourth enhanced feature tensors are fused and input into the enhancement module to obtain the fourth enhanced feature tensor, and the fifth enhanced feature tensor is input into the enhancement module to obtain the sixth enhanced feature tensor; The first, second, third, fourth, fifth, and sixth enhanced feature tensors are passed through a fully connected layer to output the seventh enhanced feature tensor. The first feature tensor and the seventh enhanced feature tensor are concatenated to obtain the fourth feature tensor, which is then used as the input to the next module.

[0016] Multi-scale feature extraction is achieved through channel splitting and multi-parallel convolution. The feature expression is enhanced step by step through multi-level serial and cross-branch enhancement modules. After fusing the enhanced features from multiple paths, the output is spliced ​​with the original features. This can fully explore the fine-grained features of the small pressure plates in the switch cabinet, enhance the multi-scale adaptation capability, and effectively improve the feature extraction accuracy of dense pressure plate targets.

[0017] Optionally, the enhancement module works as follows: Obtain the second input feature tensor, and then input the second input feature tensor into two parallel DWConv modules after Conv convolution to obtain the first convolution feature tensor and the second convolution feature tensor. Then, fuse the first convolution feature tensor and the second convolution feature tensor and then convolve them again to obtain the third convolution feature tensor. The third convolutional feature tensor is input into two parallel DWConv modules to obtain the fourth and fifth convolutional feature tensors. The fifth convolutional feature tensor is then convolved and processed by the ReLU activation function to obtain the sixth convolutional feature tensor, which is then used as the input to the next module.

[0018] This enhancement module employs a convolutional structure combined with a two-layer parallel depthwise separable convolutional structure. Through multi-branch feature extraction and stepwise fusion, it reduces the number of parameters while enhancing the capture of local details. It also uses the ReLU activation function to enhance the non-linear expression of features, accurately extracts the minute textures and edge features of the pressure plate, and improves the feature recognition of dense pressure plate targets.

[0019] Optionally, the second target module works as follows: Obtain the third input feature tensor, input the third input feature tensor into the CBS module to obtain the first target feature tensor, and input the first target feature tensor into two parallel CBS modules to obtain the second target feature tensor and the third target feature tensor respectively; The third target feature tensor is input into a CBS module and two cascaded DSConv modules respectively to obtain the fourth and fifth target feature tensors. The fourth and fifth target feature tensors are then fused and input into the sixth target feature tensor. The second and sixth target feature tensors are concatenated and normalized to obtain the seventh target feature tensor, which is then used as the input for the next module.

[0020] The second target module achieves basic feature extraction through a dual-branch parallel CBS module, expands the receptive field by combining serially dilated grouped shuffling convolution, and normalizes the output after multi-branch feature fusion. While reducing the number of model parameters and computational load, it enhances the multi-scale feature fusion effect, adapts to the complex background of switch cabinet pressure plates, and improves detection and inference efficiency.

[0021] Optionally, a reliable state result of the pressure plate is obtained by performing a correlation consistency verification on the state detection result based on the operational data, including: Acquire real-time operating data of the target relay protection circuit corresponding to the target pressure plate; the real-time operating data includes at least one or more of the following: circuit current data, circuit voltage data, input / output status data, protection action flag data, and remote signaling feedback data; Based on the association mapping relationship, the state detection result is converted into the corresponding loop expected state, and the time consistency feature between the loop expected state and the real-time running data is constructed; The temporal consistency characteristics are evaluated using a multi-factor fusion method to obtain an association consistency score that includes state matching degree, temporal synchronization degree, and event correlation degree. When the correlation consistency score meets the preset confidence threshold, the state detection result is determined as the confidence state result of the pressure plate; When the correlation consistency score does not meet the preset confidence threshold, a secondary verification process based on multi-frame image review and loop data resampling is triggered to output the corrected confidence status result.

[0022] By integrating the visual inspection results of the pressure plate with the real-time operation data of the relay protection circuit, a time-series consistency feature is constructed. A multi-dimensional fusion evaluation is carried out, and a preliminary credibility judgment is completed through a credibility threshold. If the standard is not met, a secondary verification process of multi-frame image review and circuit data resampling is initiated, which effectively avoids the misjudgment defects of single visual inspection and greatly improves the reliability and fault tolerance of pressure plate status recognition.

[0023] In a second aspect of this invention, an online detection system for substation pressure plates based on trusted WLAN communication is proposed, the system comprising: The image enhancement module is used to periodically acquire the original image of the target pressure plate, perform enhancement operations on the original image, and output an enhanced image; the original image is an RGB image. A state detection module is used to input the enhanced image into the pressure plate state detection model to obtain the state detection result; the state detection model is based on YOLOv11 and is improved by replacing C3K2 in the YOLOv11 backbone network with the first target module; the state detection result includes at least one of the pressure plate deployment / retraction state, pressure plate pose state, and pressure plate defect state; The result verification module is used to obtain the operating data of the target relay protection circuit corresponding to the state detection result according to the association mapping relationship, and to perform association consistency verification on the state detection result based on the operating data to obtain the reliable state result of the pressure plate; the association mapping relationship is the relationship between the substation pressure plate and the corresponding relay protection circuit that is constructed in advance. The data upload module is used to send the trusted status results to the cloud monitoring platform for online monitoring via a trusted WLAN. Attached Figure Description

[0024] The invention will now be further described with reference to the accompanying drawings.

[0025] Figure 1 A flowchart illustrating an online detection method for substation pressure plates based on trusted WLAN communication, provided in an embodiment of the present invention; Figure 2 This is a network architecture diagram of the state detection model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the first module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the enhancement module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the second module provided in an embodiment of the present invention; Figure 6 This is a framework diagram of an online detection system for substation pressure plates based on trusted WLAN communication, provided as an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0027] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention provides an online detection method for substation pressure plates based on trusted WLAN communication. See also... Figure 1 The method includes the following steps: S1: Periodically acquire the original image of the target pressure plate, perform enhancement operation on the original image, and output the enhanced image; S2, input the enhanced image into the pressure plate state detection model to obtain the state detection result; S3. Based on the correlation mapping relationship, obtain the operating data of the target relay protection circuit corresponding to the state detection result, and perform correlation consistency verification on the state detection result based on the operating data to obtain the reliable state result of the pressure plate. S4 sends trusted status results to a cloud monitoring platform for online monitoring via a trusted WLAN.

[0029] The original image is an RGB image; the state detection model is an improvement based on YOLOv11, replacing C3K2 in the YOLOv11 backbone network with the first target module; The status detection results include at least one of the following: pressure plate deployment / retraction status, pressure plate posture status, and pressure plate defect status; The association mapping relationship is the pre-constructed relationship between substation pressure plates and corresponding relay protection circuits.

[0030] In one implementation, the association mapping relationship between the substation pressure plate and the corresponding relay protection circuit is constructed, including: Obtain the pressure plate identification information, pressure plate bay information, and corresponding relay protection device configuration data for each pressure plate in the substation; Parse the configuration file data and extract the protection function identifier, remote signaling point identifier, and protection circuit status identifier corresponding to each pressure plate; Based on the pressure plate marking information, protection function marking, remote signaling point marking, and protection circuit status marking, establish the status association relationship between the pressure plate and the corresponding relay protection circuit; The corresponding pressure plate identifiers, protection function identifiers, remote signaling point identifiers, and protection circuit status identifiers are uniformly bound to generate a pressure plate status mapping index table. Based on the pressure plate status mapping index table, establish a corresponding target relay protection circuit status access path for each pressure plate; After obtaining the pressure plate status detection result, the corresponding relay protection circuit operation data is called according to the status access path of the corresponding pressure plate to realize the association mapping between the pressure plate status detection result and the relay protection circuit status.

[0031] In one implementation, a trusted communication link is established between the edge detection terminal and the cloud monitoring platform based on a trusted WLAN communication network. The process is as follows: The edge detection terminal connects to a trusted WLAN communication network and sends a terminal access request to the cloud monitoring platform; the terminal access request includes the terminal identity identifier, device fingerprint information, and terminal digital certificate. The cloud-based monitoring platform verifies the legitimacy of the terminal's identity, device fingerprint information, and terminal digital certificate, and sends an identity authentication response to the edge detection terminal after successful verification. The edge detection terminal verifies the identity of the cloud monitoring platform based on the identity authentication response information, and generates random negotiation parameters after the platform identity verification is successful. The edge detection terminal and the cloud monitoring platform negotiate session keys based on random negotiation parameters to generate corresponding dynamic communication session keys. The edge detection terminal establishes an encrypted communication channel based on a dynamic communication session key and performs encrypted encapsulation processing on the pressure plate detection data to be transmitted; The edge detection terminal generates corresponding data summary information based on the pressure plate detection data, and sends the data summary information and the encrypted pressure plate detection data to the cloud monitoring platform. The cloud-based monitoring platform decrypts the received encrypted data and performs integrity verification on the decrypted pressure plate detection data based on the data digest information. When the integrity verification passes, it is determined that the trusted communication link between the edge detection terminal and the cloud monitoring platform has been established. Trusted status results are sent to the cloud monitoring platform via a trusted communication link.

[0032] In one embodiment, performing an enhancement operation on the original image to output an enhanced image includes: The original image is converted to an HSV color image and used as the first image; The brightness channel V in the first image is extracted, and the brightness channel V is decomposed into a first component and a second component through a preset physical model. The first component is a low-frequency structure used to characterize the lighting at the substation site, and the second component is high-frequency information used to characterize the details of the pressure plate body. First constraint functions and second constraint functions corresponding to the first component and the second component are constructed respectively. The first constraint function is used to constrain the first component to smooth the uneven lighting, shadows and metallic reflections of the original image and preserve the edge of the pressure plate structure. The second constraint function is used to constrain the second component to clearly extract and preserve the edge of the pressure plate and text markings, and prevent the details from being blurred or lost. A third constraint function for the bright channel is constructed by applying illumination constraints to the original image; the third constraint function is used to constrain the first component to be close to the standard bright channel illumination, so as to avoid overexposure and over-brightness due to deviation from the standard illumination. An enhanced image is obtained by performing a first preset operation based on the first constraint function, the second constraint function, and the third constraint function.

[0033] In one implementation, the original image is converted into an HSV color image, the process of which is as follows: The original image is split into three independent channels based on physical channels: red channel R, green channel G, and blue channel B. The pixel values ​​of the R, G, and B channels are normalized from integers of 0 to 255 to floating-point numbers of 0 to 1 to eliminate the influence of numerical magnitude. Calculate the maximum and minimum values ​​of R / G / B in a single pixel, and use the maximum value of the pixel as the luminance V channel data of the HSV color image; If the maximum value of a pixel is 0, then the saturation channel S of the HSV color image is 0. If the maximum value of a pixel is not 0, then the saturation channel S = (maximum value - minimum value) / maximum value; The angle value (0~360°) is calculated based on the relative size of R / G / B, which uniquely represents the color of the pressure plate, such as red for input and green for output.

[0034] In one implementation, the luminance channel V is decomposed using a preset physical model (e.g., Retinex, Retinex-Net) to obtain the first component and the second component.

[0035] In one implementation, a third constraint function for the bright channel is constructed by applying illumination constraints to the original image. The process is as follows: A fixed-size neighborhood window is defined with each pixel in the original image as the center. For each pixel within the neighborhood window, extract the pixel values ​​of the R, G, and B color channels, and take the maximum value of the three channels for a single pixel; Within the entire neighborhood window, the maximum value among the three channels of all pixels is taken again as the initial illumination value of the center pixel. This process is repeated for all pixels in the original image to obtain the initial bright channel illumination map. Since the initial bright channel illumination map calculated directly has block effects and jagged edges, especially in the pressure plate metal area, hard bright spots will appear. Therefore, the brightness channel V is used as the guide map to perform guided filtering on the initial bright channel illumination map to obtain an optimized bright channel illumination map. Using the optimized bright channel lighting map as the standard reference lighting, a third constraint function is constructed;

[0036] in, The third constraint function is defined as follows: Y is the initial bright channel lighting pattern, and B is the optimized bright channel lighting pattern. It represents the squared L2 norm, which measures the deviation of the illumination component from the standard illumination.

[0037] In one implementation, both the first constraint function and the third constraint function are used to constrain the first component, but they have different constraint dimensions. The first constraint function focuses on constraining the illumination component from the perspective of local gradient changes to smooth the uneven illumination and metal reflection of the substation pressure plate image and preserve the edge of the pressure plate structure. The third constraint function focuses on constraining the illumination component from the perspective of overall brightness rationality to fit the standard bright channel illumination calculated from the original RGB image, so as to avoid the illumination component from being too bright or overexposed and having an imbalance between light and dark. The two functions work together to complete the constraint optimization of the first component from the two dimensions of local details and overall brightness.

[0038] In one embodiment, the first constraint function and the second constraint function are constructed as follows: Centered on each pixel p in the first component, a first local window of fixed size is defined, and horizontal and vertical gradient operators are defined. For each pixel q within the first local window, calculate the spatial weight. Traverse all pixels of the first component and calculate the total variation of the horizontal and vertical windows respectively. Using the luminance channel V as the guide image and the first component as the image to be filtered, a guide filtering operation is performed on the image to be filtered to obtain the filtered image, and its horizontal and vertical gradients are calculated. The first constraint function is calculated based on the total variation of the horizontal window, the total variation of the vertical window, the horizontal gradient, and the vertical gradient. A second local window of fixed size is defined with each pixel m in the second component as the center, and horizontal and vertical gradient operators are defined. The horizontal and vertical gradients of the second component are calculated pixel by pixel based on the horizontal and vertical gradient operators. The horizontal and vertical gradients of all pixels within the second local window are calculated to obtain the mean local gradient. Logarithmic operation is then performed on the mean local gradient to generate adaptive detail weights. The gradients in the corresponding directions are weighted according to the adaptive weights, and the second constraint function of the second component is calculated based on the weighted horizontal and vertical gradients.

[0039] In one implementation, the horizontal gradient operator and vertical gradient operator corresponding to the first component and the second component are defined as follows: Set up a pixel coordinate system with the row direction of the image as the y-axis and the column direction as the x-axis. Record the pixel position of the first component / second component as (i,j), where i represents the row of the pixel and j represents the column of the pixel. The horizontal gradient operator (in the left-right direction, used to capture vertical light and dark variations) adopts a discrete central difference form: F represents the first or second component to be calculated. The vertical gradient operator (vertical direction, used to capture horizontal light and dark variations) adopts a discrete central difference form: .

[0040] In one implementation, constructing the first constraint function corresponding to the first component includes: Centered on each pixel p in the first component, a first local window of fixed size is defined (the size can be a 3×3 window, completely covering the plate body, edge, text and background area). Horizontal gradient operator and vertical gradient operator are defined to extract the pixel change amplitude of the first component. For each pixel q within the local window, a spatial weight is calculated using a Gaussian function. The spatial weight decreases as the distance between pixel q and the center pixel p increases, with closer pixels having higher weights and farther pixels having lower weights, ensuring that the lighting is smoother and better matches the local features of the pressure plate. Iterate through all pixels of the first component and calculate the total variation of the horizontal and vertical windows respectively: The absolute gradient values ​​of all pixels within the local window are multiplied by their corresponding spatial weights and then summed to obtain the total amount of illumination change of pixel p in the horizontal and vertical directions, which is used to quantify the degree of texture fluctuation in the illumination of the pressure plate.

[0041]

[0042] in, and These represent the total variation of the first component pixel p in the horizontal and vertical directions of the window, respectively; This represents the Gaussian space weights of pixel p and pixel q within the window; and These are the horizontal gradient operator and the vertical gradient operator, respectively. Let be the value of the first component corresponding to pixel q within the window. This represents a local window centered at pixel p; Using the luminance channel V as the guide image and the first component as the image to be filtered, a guided filtering operation is performed. After filtering, the overall structure of the pressure plate illumination is preserved and the fine texture fluctuations are eliminated. Then, the horizontal and vertical gradients of the first component after filtering are calculated to distinguish the structural edges of the pressure plate illumination from noise textures. The first constraint function is calculated based on the total variation of the horizontal window, the total variation of the vertical window, and the filtered horizontal and vertical gradients.

[0043] in, This represents the first constraint function used to constrain the first component. and These represent the horizontal and vertical gradients of the light emitted from the pressure plate after the filter is applied. Let I be the L1 norm and I be the first component; Minimal constant, to prevent the denominator from being 0, takes the value of .

[0044] In one implementation, the second constraint function corresponding to the second component is constructed as follows: Centered on each pixel p in the second component, a fixed 3×3 second local window is set, which covers the surrounding area of ​​the pixel to adapt to the detailed size of small text and narrow edges on the pressure plate. The horizontal and vertical gradient operators are used to calculate the horizontal and vertical gradients of the second component pixel by pixel; the magnitude of the gradient value represents the strength of the brightness and darkness change of the edge and text at that position on the pressure plate. For each 3×3 local window, calculate the arithmetic mean of the horizontal gradient and the arithmetic mean of the vertical gradient of all pixels within the window to obtain the local detail gradient mean; the gradient mean is used to characterize the overall saliency of the pressure plate details in this region. Logarithmic operation on the local gradient mean generates detailed adaptive weights: Low-light dark areas: The gradient mean is small, and logarithmic operations will amplify the weights, highlighting the faint text and edge details on the plate. Background noise region: Gradients are messy and irregular. Logarithmic operations can weaken the noise weights and suppress invalid noise points; The gradients in the corresponding directions are weighted using adaptive weights, and the L1 norms of the weighted horizontal and vertical gradients are calculated and then summed to finally obtain the second constraint function of the second component of the pressure plate.

[0045] in, This represents the second constraint function used to constrain the second component. It is a 3x3 local window. R is the total number of pixels within the window, a is the second component, and a is the logarithmic coefficient used to adjust the sensitivity of the pressure plate detail extraction.

[0046] In one implementation, due to issues such as color interference, uneven lighting, metallic reflection, local shadows, and the easy loss of subtle details at the edges and text of the substation pressure plate in the original RGB image, color information is stripped away by color space conversion, retaining only the luminance channel. Then, a preset model is used to split the luminance information into a first component representing the external environment and a second component representing the characteristics of the pressure plate itself. Subsequently, constraint functions are constructed to locally constrain the two components. The luminance channel is introduced as a priori constraint on the overall luminance rationality of the first component. From multiple dimensions such as color stripping, component decoupling, local smoothing to preserve details, and overall luminance control, a complete set of constraints is established for subsequent optimization, solving the defects of abnormal lighting and blurred details in the pressure plate image.

[0047] In one embodiment, performing a preset operation based on a first constraint function, a second constraint function, and a third constraint function to obtain an enhanced image includes: The first constraint function, the second constraint function, and the third constraint function are weighted and fused to obtain the overall objective function. Using the first and second components as initial values ​​for iteration, perform iterative optimization on the overall objective function and output the optimal first and second components. Adaptive enhancement is performed on the optimal first component to output the final first component. The final first component is then fused with the optimal second component to obtain the optimal luminance channel V. The H channel, S constraint function channel, and optimal luminance channel V of the pressure plate HSV color image are fused to obtain the enhanced constraint function HSV pressure plate image. The enhanced constraint function HSV pressure plate image is then converted into RGB format to obtain the enhanced image of the pressure plate.

[0048] In one implementation, the first constraint function, the second constraint function, and the third constraint function are weighted and fused to obtain the overall objective function, including: Based on Retinex theory, a system was constructed. The data fidelity term constrains the pixel-wise product of the first component I and the second component R to be close to the original brightness V, ensuring that the optimized components do not deviate from the true brightness information of the pressure plate and maintain the basic observation features of the image. The constraint strength of the first constraint function, the second constraint function, and the third constraint function are adjusted according to the preset weight coefficients to adapt to the scene characteristics of low light, reflection, and weak details of substation pressure plates, and to balance the effects of lighting smoothing, detail enhancement and brightness control. The data fidelity term is added to the weighted three types of constraint terms and integrated into a unified overall objective function; this achieves the overall joint optimization of the first component and the second component R, while incorporating all conditions such as local smoothing, detail preservation, and overall brightness constraint.

[0049] In one implementation, iterative optimization is performed on the overall objective function, the process of which is as follows: Step 1: Set the maximum number of iterations and the convergence error threshold; Step 2: Fix the second component of the current iteration, and only retain the data fidelity term, the first constraint function and the third constraint function related to the first component in the objective function. By minimizing this part of the function, the updated optimal first component is obtained, so as to achieve smooth lighting of the pressure plate, suppression of metal reflection and overall brightness and darkness balance. Step 3: Fix the optimal first component, retain only the data fidelity term and the second constraint function related to the second component in the objective function, and obtain the updated optimal second component by minimizing this part of the function;

[0050]

[0051] in, and Let these represent the first and second components of the pressure plate in the k-th iteration, respectively. and These are the regularization weights of the first constraint function. and These are the regularization weights of the second constraint function; S is the original image of the pressure plate. Step 4: Calculate the pixel error between the first and second components of the current round and the previous round. If the error is less than the preset convergence error threshold or the maximum number of iterations is reached, stop the iteration. If it does not converge, return to step 2 and continue to update alternately until the convergence condition is met.

[0052] In one implementation, adaptive enhancement is performed on the optimal first component to output the final first component, including: The average brightness of the optimal first component is statistically analyzed, and adaptive enhancement parameters are obtained based on the changes in average brightness. Construct an arctangent transform function, and perform nonlinear adjustment on the optimal first component to obtain the final first component. This component is used to compress the brightness and suppress overexposure in high-brightness areas such as metal reflections on the pressure plate, and to raise the brightness in low-brightness areas such as cabinet shadows and dark areas on the pressure plate, thereby achieving adaptive balance between bright and dark areas.

[0053] in, Indicates the final first component, The optimal first component, For adaptive enhancement parameters.

[0054] In one implementation, to address the issues of scattered constraints and difficulty in solving the first and second components obtained from the initial decomposition, and the fact that the original pressure plate image still suffers from weak details in dark areas, overexposure of highlights, and uneven distribution of light and dark areas, a unified optimization objective function is constructed by integrating multiple constraints such as illumination smoothing, detail preservation, and brightness control. The overall objective function is iteratively optimized to accurately solve for the optimal first and second components. The optimal illumination component is nonlinearly adaptively adjusted using arctangent transform to balance brightness and enhance details in dark areas. The optimized brightness channel is then fused with the original unchanged hue and saturation channels to finally generate a complete pressure plate enhancement image with realistic colors, uniform illumination, and clear details, which meets the needs of pressure plate activation / deactivation status recognition in complex substation environments.

[0055] In one embodiment, the state detection model is based on an improvement of YOLOv11, and the improvements include: All C3K2 modules in the backbone network of the YOLOv11 model are replaced with the first target module. The first target module is used to enhance the fine-grained feature extraction and multi-scale adaptation capability of dense micro pressure plates in substation switchgear by reparameterizing the heterogeneous multi-scale cascade structure and fusing cross-branch features. The C3K2 module in the neck network of the YOLOv11 model is completely replaced with the second target module. The second target module expands the receptive field and strengthens the fusion of multi-scale features by fusing double convolution and dilated grouped shuffling convolution in stages, which reduces the number of parameters and computational overhead while making the model lightweight.

[0056] In one implementation, the state detection model is trained using a self-built dataset; the training environment is: operating system: Windows 11, GPU: NVIDIA GeForce RTX 4070 Ti Super, 16GB video memory; CUDA version: 12.1, Python version: 3.9.19, deep learning framework: PyTorch 2.5.1.

[0057] In one implementation, see [link to implementation details]. Figure 2After replacing all C3K2 modules in the YOLOv11 backbone network with the first target module, the model can significantly enhance the fine-grained feature extraction capability of small, dense, and obstructed pressure plates in substation switchgear by leveraging the advantages of reparameterized heterogeneous multi-scale cascaded structure and cross-branch feature fusion. This adapts to the multi-scale morphological differences of pressure plates, effectively avoids the loss of small target features caused by complex backgrounds and messy lines, and reduces the probability of missed or false detections of pressure plates. After replacing the neck network C3K2 with the second target module, the model can significantly expand the feature receptive field and strengthen the interaction and fusion of high and low-level multi-scale features by using the synergistic effect of staged double convolution and dilated grouped mixed convolution. This allows the improved model to balance inference speed and pressure plate status recognition accuracy in real-time detection scenarios of embedded equipment in substations, and has a stronger generalization ability in complex cabinet environments.

[0058] In one implementation method, the results of ablation experiments are compared as follows:

[0059] Where A represents replacing only C3K2 in the YOLOv11 backbone network with the first target module, B represents replacing only C3K2 in the YOLOv11 neck network with the second target module, and A+B represents the state detection model; FLOPs represents floating-point operations, Param represents the number of parameters (the smaller the value, the lower the memory usage of the model), and AP represents average precision. As can be seen from the table above, among the ablation experiment results under the three indicators, the state detection model is the best, followed by model A, while model B performs the worst.

[0060] In one embodiment, the first target module works as follows: Obtain the first input feature tensor, split the first input feature tensor according to the number of channels, and convolve the split feature tensors through 4 parallel Conv modules to obtain the first feature tensor, the second feature tensor, the third feature tensor, and the fourth feature tensor respectively. The first and second feature tensors are fused and input into the enhancement module to obtain the first enhanced feature tensor. The first enhanced feature tensor is then input into the enhancement module to obtain the second enhanced feature tensor. The two enhancement modules are connected in series. The first and third enhanced feature tensors are fused and input into the enhancement module to obtain the third enhanced feature tensor, and the third enhanced feature tensor is input into the enhancement module to obtain the fourth enhanced feature tensor; The third and fourth enhanced feature tensors are fused and input into the enhancement module to obtain the fourth enhanced feature tensor, and the fifth enhanced feature tensor is input into the enhancement module to obtain the sixth enhanced feature tensor; The first, second, third, fourth, fifth, and sixth enhanced feature tensors are passed through a fully connected layer to output the seventh enhanced feature tensor. The first feature tensor and the seventh enhanced feature tensor are concatenated to obtain the fourth feature tensor, which is then used as the input to the next module.

[0061] In one implementation, channel splitting and four-way parallel convolution lay the foundation for multi-scale basic feature extraction: The input feature tensor is split by channel and generated into four parallel Conv modules to produce the first, second, third, and fourth feature tensors. The four parallel convolutions can simultaneously extract the basic features of the pressure plate at different scales: the first feature tensor focuses on shallow, fine-grained information such as the edge of the pressure plate and the subtle texture of opening and closing; the second feature tensor captures the local morphological features of a single pressure plate; the third feature tensor extracts the regional layout features of dense pressure plates; and the fourth feature tensor obtains the global environment of the switchgear and the overall position information of the pressure plates. Through the multi-branch parallel structure, the limitation of a single convolution with a fixed receptive field is broken, and the multi-scale features of the pressure plate from local to global are covered at one time, which is suitable for the detection characteristics of small-sized and densely arranged pressure plates.

[0062] In one implementation, a hierarchical cross-tensor fusion and concatenated enhancement module is used to enhance fine-grained features and suppress background noise in a layered manner. The module adopts a hierarchical structure of progressive feature fusion and dual enhancement module concatenation to achieve refined feature enhancement from shallow to deep: After fusing the first and second feature tensors, the first and second enhanced feature tensors are obtained through two-stage series enhancement modules. The basic outline of the pressure plate and the local texture are fused together. The subtle state features of the pressure plate opening and closing are amplified step by step through two enhancements, filtering out the cabinet reflection and shallow noise of messy lines, and solving the problem that the small pressure plate features are easily submerged by the background. The third feature tensor is introduced and fused with the first enhanced feature tensor to generate the third and fourth enhanced feature tensors, which connect the local fine-grained features of the pressure plate with the regional arrangement features, strengthen the spatial correlation information between dense pressure plates, and alleviate feature confusion caused by mutual occlusion of pressure plates. The fourth feature tensor is introduced and fused with the third enhanced feature tensor to generate the fifth and sixth enhanced feature tensors, thus opening up a multi-scale feature link from local to regional to global, improving the global positioning information of the pressure plate under complex cabinet background, and expanding the feature receptive field.

[0063] In one implementation, the fully connected layer aggregates multi-level enhanced features to achieve effective feature selection and unified representation. The 1st, 2nd, 3rd, 4th, 5th, and 6th enhancement feature tensors are input into the fully connected layer, and the 7th enhancement feature tensor is output. The fully connected layer can adaptively assign weights to the enhancement features of multiple scales and dimensions, eliminate redundant and invalid features such as cabinet and wiring, and filter and aggregate the core state features of the pressure plate to achieve a unified representation of multi-scale features, which greatly improves the anti-interference ability of features to complex cabinet environments.

[0064] In one implementation, basic features and deep enhancement features are concatenated to achieve complementary fusion of shallow and deep features: The first feature tensor and the seventh enhanced feature tensor are concatenated to generate the final output features. The first feature tensor fully preserves basic fine-grained information such as shallow edges and textures of the pressure plate, avoiding the loss of small shallow features of the pressure plate; the seventh enhanced feature tensor carries the global semantic and state features after depth enhancement. The concatenation of the two achieves bidirectional complementarity between shallow detail features and deep semantic features, ultimately achieving the core functions of enhancing fine-grained feature extraction of substation pressure plates, adapting to multi-scale targets, and resisting interference from complex backgrounds.

[0065] In one implementation, the comparison results of downsampling experiments conducted in the backbone network using different modules are shown in the table below:

[0066] As can be seen from the table above, among the three indicators of the first target module in the backbone network, although it does not have the smallest number of parameters, it is the module with the best performance when considered in combination with mAP50(%) and Flops / G.

[0067] In one implementation, see [link to implementation details]. Figure 3 The first target module first splits the input feature channels and extracts multi-scale basic features of the substation pressure plate through four-way parallel convolution. Through the hierarchical tensor fusion and serial enhancement module, the fine-grained texture, regional arrangement and global position features of the pressure plate are enhanced in layers to alleviate interference such as cabinet occlusion and reflection. Relying on the fully connected layer to aggregate multi-level enhanced features, effective target information is screened. Finally, the shallow basic features and deep enhanced features are spliced ​​to achieve complementary deep and shallow features, effectively improve the feature extraction capability of small and dense pressure plates, reduce the probability of missed detection and false detection, and adapt to complex cabinet detection scenarios.

[0068] In one embodiment, the enhancement module works as follows: Obtain the second input feature tensor, and then input the second input feature tensor into two parallel DWConv modules after Conv convolution to obtain the first convolution feature tensor and the second convolution feature tensor. Then, fuse the first convolution feature tensor and the second convolution feature tensor and then convolve them again to obtain the third convolution feature tensor. The third convolutional feature tensor is input into two parallel DWConv modules to obtain the fourth and fifth convolutional feature tensors. The fifth convolutional feature tensor is then convolved and processed by the ReLU activation function to obtain the sixth convolutional feature tensor, which is then used as the input to the next module.

[0069] In one implementation, firstly, the second input feature tensor undergoes basic feature normalization via Conv, and is then split into two parallel DWConv (depth convolution) paths to generate the first and second convolutional feature tensors. DWConv can extract local details in multiple directions, such as the edge of the pressure plate and the fine texture of the opening and closing of the gate, in parallel with a low number of parameters. After the two are fused, they are passed through Conv to obtain the third convolutional feature tensor, which integrates the detailed information of the two branches and initially filters the shallow background noise of the cabinet. Secondly, the third convolutional feature tensor is again processed by dual-path DWConv to generate the fourth and fifth convolutional feature tensors, further mining the spatial arrangement and deep correlation features of the dense pressure plates; after fusion convolution, the sixth convolutional feature tensor is obtained by ReLU activation, nonlinearly screening effective target features, suppressing interference such as messy wiring and metal reflection, and finally achieving refined enhancement of the features of the small pressure plates.

[0070] In one implementation, see [link to implementation details]. Figure 4 The enhancement module adopts a two-stage parallel depthwise separable convolutional structure. First, the input features are regularized through Conv, and two DWConv channels are used in parallel to extract multi-directional detailed features of the pressure plate. After fusion, the initial noise reduction is completed. The second dual-branch convolution further mines deep correlation features, and the effective information is filtered by ReLU activation. Under the premise of lightweight design, the fine-grained features of the small pressure plate are enhanced, effectively suppressing interference such as cabinet reflection and messy lines, and improving the robustness of feature extraction.

[0071] In one embodiment, the second target module works as follows: Obtain the third input feature tensor, input the third input feature tensor into the CBS module to obtain the first target feature tensor, and input the first target feature tensor into two parallel CBS modules to obtain the second target feature tensor and the third target feature tensor respectively; The third target feature tensor is input into a CBS module and two cascaded DSConv modules respectively to obtain the fourth and fifth target feature tensors. The fourth and fifth target feature tensors are then fused and input into the sixth target feature tensor. The second and sixth target feature tensors are concatenated and normalized to obtain the seventh target feature tensor, which is then used as the input for the next module.

[0072] In one implementation, the first target feature tensor is generated from the third input feature tensor via the CBS module, completing feature standardization and extraction of the basic contour and edge texture of the pressure plate; the features are split to obtain the second and third target feature tensors, the second tensor retains the shallow basic details, and the third tensor enters the deep feature mining branch to achieve feature layer processing; wherein, the CBS module is composed of Conv (ordinary convolution) + BN (batch normalization) + SiLU (activation function); The third target feature tensor extracts features in two ways. The fourth target feature tensor generated by CBS retains the local subtle state features of the pressure plate. The fifth target feature tensor obtained by concatenating DSConv (dilated grouped shuffling convolution) expands the receptive field in a lightweight way and captures the global arrangement information of the pressure plate. The two are fused to obtain the sixth target feature tensor, which takes into account both detailed extraction and global correlation, while reducing the number of model parameters and computational cost. The shallow second target feature tensor and the deep sixth target feature tensor are concatenated and normalized to obtain the seventh target feature tensor. This integrates multi-scale information from both shallow and deep layers, suppresses cabinet reflection and line obstruction interference, stabilizes feature distribution, and achieves feature enhancement under lightweight conditions.

[0073] In one implementation, see [link to implementation details]. Figure 5 The input features are standardized and extracted using CBS. The feature layering is achieved through a dual-branch splitting process. The second target feature tensor retains the shallow details of the pressure plate. Local textures are extracted and the receptive field is expanded by parallel extraction of CBS and serial DSConv. The fusion is then used to obtain the sixth target feature tensor. Finally, the deep and shallow features are spliced ​​and normalized to achieve multi-scale feature fusion. While reducing computational overhead, this method captures the global layout features of dense pressure plates, suppresses reflections inside the cabinet and line interference, and adapts to the real-time detection requirements at the edge of the substation.

[0074] In one embodiment, the reliable state result of the pressure plate is obtained by performing a correlation consistency verification on the state detection results based on the operating data, including: Acquire real-time operating data of the target relay protection circuit corresponding to the target pressure plate; the real-time operating data shall include at least one or more of the following: circuit current data, circuit voltage data, input / output status data, protection action flag data, and remote signaling feedback data; Based on the association mapping relationship, the state detection results are converted into the corresponding loop expected state, and the time consistency characteristics between the loop expected state and the real-time running data are constructed. A multi-factor fusion evaluation of temporal consistency characteristics is performed to obtain an association consistency score that includes state matching degree, temporal synchronization degree, and event correlation degree. When the association consistency score meets the preset confidence threshold, the state detection result is determined as the confidence state result of the pressure plate. When the correlation consistency score does not meet the preset confidence threshold, a secondary verification process based on multi-frame image review and loop data resampling is triggered to output the corrected confidence status result.

[0075] In one implementation, the state detection result is converted into the corresponding desired loop state based on the association mapping relationship. The process is as follows: Step 1: Establish a mapping table between pressure plate and circuit function Based on the substation's primary wiring diagram, secondary circuit diagram, and switchboard activation / deactivation logic, a mapping relationship is established between "switching board number - corresponding relay protection circuit - circuit functional status". Each switching board corresponds to at least one target circuit, and the electrical response rules that the corresponding circuit should exhibit when the switching board is in "activated, deactivated, abnormal position, defect" or other states are recorded.

[0076] Step 2: Obtain the pressure plate status detection results The pressure plate state detection model outputs the state detection result of the target pressure plate, such as "engaged", "deactivated" or "abnormal pose"; this state detection result serves as the input for the generation of the expected state of the subsequent loop. Step 3: Generate the desired state of the loop based on the association mapping relationship. The status detection results are matched with the mapping table to determine the target relay protection circuit corresponding to the pressure plate. Then, according to the preset rules, the pressure plate status is converted into the expected status that the circuit should present at the current moment. For example: When the pressure plate is in the "engaged" state, the circuit is expected to be in one of the preset logic states of "locked / connected / enabled"; When the pressure plate is in the "exit" state, the circuit is expected to be one of the preset logic states of "disconnect / release / disable"; When the pressure plate is in an "abnormal position" or "defective state", the circuit is expected to be further marked as an "uncertain state" or an "abnormal state to be verified".

[0077] In one implementation, the process of constructing the timing consistency characteristics between the desired loop state and the real-time running data is as follows: Step 1: Collect real-time operating data of the target circuit.

[0078] On the relay protection circuit corresponding to the pressure plate, real-time operating data is collected according to a preset cycle. The data may include input quantity, output quantity, contact status, remote signaling status, action flag, circuit voltage, current, etc. At the same time, the timestamp of the collection corresponding to each set of data is recorded to form a time series data stream. Step 2: Step S5: Perform time alignment Align the desired state of the loop with the collected real-time operating data using a unified time reference; For example, by taking the pressure plate detection time as the center, the loop operation data within the time window before and after can be selected to form the same analysis interval, thus avoiding misjudgment due to sampling delay.

[0079] Step 3: Extract temporal consistency features The aligned expected state and real-time running data are compared to construct time-series consistency features; This feature may include at least the following: State matching feature: Whether the real-time running data is consistent with the expected state; Response delay characteristics: Whether the loop state changes and responds within a preset time; Stability maintenance characteristics: Whether the state remains stable after a change, without jitter or repeated jumps; Event synchronization characteristics: Whether the change in the pressure plate status and the circuit action flag appear synchronously; The above characteristics together reflect whether there is consistency between the visual condition of the pressure plate and the electrical condition of the circuit.

[0080] Step 4: Output the timing consistency judgment result Generate consistency scores or consistency labels based on temporal consistency characteristics; If the real-time running data and the expected state of the loop are continuously matched in time, they are considered to be consistent. If there are state mismatches, response timeouts, event asynchrony, or frequent jitter, it is judged as inconsistency or abnormality.

[0081] In one implementation, by acquiring real-time operating data of the relay protection circuit and combining it with the pressure plate status detection results to construct the timing consistency characteristics between the expected state and the real-time operating state of the circuit, electrical logic verification of the pressure plate visual recognition results can be achieved, avoiding misjudgment of the state due to light interference, obstruction, or misidentification. Furthermore, by using multi-factor fusion evaluation of state matching degree, timing synchronization degree, and event correlation degree, the reliability and accuracy of pressure plate status determination can be improved. When the consistency is insufficient, a secondary verification process is triggered, which can effectively reduce the false alarm rate and the missed alarm rate, thereby improving the safety and reliability of substation pressure plate online monitoring.

[0082] Based on the same inventive concept, this invention also provides an online detection system for substation pressure plates based on trusted WLAN communication. See also Figure 6 The system includes: The image enhancement module is used to periodically acquire the original image of the target pressure plate, perform enhancement operations on the original image, and output the enhanced image; the original image is an RGB image. The state detection module is used to input the enhanced image into the pressure plate state detection model to obtain the state detection result; the state detection model is based on YOLOv11 and is improved by replacing C3K2 in the YOLOv11 backbone network with the first target module; the state detection result includes at least one of the pressure plate deployment / retraction state, pressure plate pose state, and pressure plate defect state; The result verification module is used to obtain the operating data of the target relay protection circuit corresponding to the state detection result according to the association mapping relationship, and to perform association consistency verification on the state detection result based on the operating data to obtain the reliable state result of the pressure plate; the association mapping relationship is the relationship between the substation pressure plate and the corresponding relay protection circuit that is constructed in advance. The data upload module is used to send trusted status results to the cloud monitoring platform for online monitoring via a trusted WLAN.

[0083] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A method for online detection of substation pressure plates based on trusted WLAN communication, characterized in that, The method includes: The system periodically acquires original images of the target pressure plate, performs enhancement operations on the original images, and outputs enhanced images; the original images are RGB images. The enhanced image is input into the pressure plate state detection model to obtain the state detection result; the state detection model is based on YOLOv11 and is improved by replacing C3K2 in the YOLOv11 backbone network with the first target module; the state detection result includes at least one of the pressure plate deployment / retraction state, pressure plate pose state, and pressure plate defect state; Based on the association mapping relationship, the operating data of the target relay protection circuit corresponding to the state detection result is obtained, and the state detection result is verified by association consistency based on the operating data to obtain the reliable state result of the pressure plate; the association mapping relationship is the relationship between the substation pressure plate and the corresponding relay protection circuit that is constructed in advance. The trusted status results are sent to the cloud monitoring platform for online monitoring via a trusted WLAN.

2. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 1, characterized in that, Perform enhancement operations on the original image to output an enhanced image, including: The original image is converted into an HSV color image and used as the first image; The brightness channel V in the first image is extracted, and the brightness channel V is decomposed into a first component and a second component through a preset physical model. The first component is a low-frequency structure used to characterize the lighting at the substation site, and the second component is high-frequency information used to characterize the details of the pressure plate body. First constraint function and second constraint function are constructed for the first component and the second component respectively; the first constraint function is used to constrain the first component to smooth the uneven lighting, shadows and metallic reflections of the original image, and to preserve the edge of the pressure plate structure; the second constraint function is used to constrain the second component to clearly extract and preserve the edge of the pressure plate and text markings, and to prevent details from becoming blurred or lost. A third constraint function for the bright channel is constructed by applying illumination constraints to the original image; the third constraint function is used to constrain the first component to be close to the standard bright channel illumination, so as to avoid overexposure and over-brightness due to deviation from the standard illumination. An enhanced image is obtained by performing preset operations based on the first constraint function, the second constraint function, and the third constraint function.

3. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 2, wherein the first constraint function and the second constraint function are constructed, the process is as follows: Centered on each pixel p in the first component, a first local window of fixed size is defined, and horizontal and vertical gradient operators are defined. For each pixel q within the first local window, calculate the spatial weight. Traverse all pixels of the first component and calculate the total variation of the horizontal and vertical windows respectively. Using the luminance channel V as the guide image and the first component as the image to be filtered, a guide filtering operation is performed on the image to be filtered to obtain the filtered image, and its horizontal and vertical gradients are calculated. The first constraint function is calculated based on the total variation of the horizontal window, the total variation of the vertical window, the horizontal gradient, and the vertical gradient. A second local window of fixed size is defined with each pixel m in the second component as the center, and horizontal and vertical gradient operators are defined. The horizontal and vertical gradients of the second component are calculated pixel by pixel based on the horizontal and vertical gradient operators. The horizontal and vertical gradients of all pixels within the second local window are calculated to obtain the mean local gradient. Logarithmic operation is then performed on the mean local gradient to generate adaptive detail weights. The gradients in the corresponding directions are weighted according to the adaptive weights, and the second constraint function of the second component is calculated based on the weighted horizontal and vertical gradients.

4. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 2, characterized in that, An enhanced image is obtained by performing preset operations based on the first constraint function, the second constraint function, and the third constraint function, including: The first constraint function, the second constraint function, and the third constraint function are weighted and fused to obtain the overall objective function. Using the first component and the second component as initial values ​​for iteration, perform iterative optimization on the total objective function to output the optimal first component and the optimal second component. Adaptive enhancement is performed on the optimal first component to output the final first component, and the final first component is fused with the optimal second component to obtain the optimal luminance channel V; The H channel, S constraint function channel, and optimal luminance channel V of the pressure plate HSV color image are fused to obtain the enhanced constraint function HSV pressure plate image. The enhanced constraint function HSV pressure plate image is then converted into RGB format to obtain the enhanced image of the pressure plate.

5. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 1, characterized in that, The state detection model is based on an improvement of YOLOv11, and the improvements include: All C3K2 modules in the YOLOv11 model backbone network are replaced with the first target module; the first target module is used to enhance the fine-grained feature extraction and multi-scale adaptation capability of dense micro pressure plates in substation switchgear by reparameterizing the heterogeneous multi-scale cascade structure and fusing cross-branch features. The C3K2 modules in the neck network of the YOLOv11 model are all replaced with the second target module. The second target module expands the receptive field and strengthens multi-scale feature fusion while reducing the number of parameters and computational overhead by fusing dual convolutions and dilated grouped shuffling convolutions in stages.

6. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 4, characterized in that, The working principle of the first target module is as follows: Obtain the first input feature tensor, split the first input feature tensor according to the number of channels, and convolve the split feature tensors through 4 parallel Conv modules to obtain the first feature tensor, the second feature tensor, the third feature tensor, and the fourth feature tensor respectively. The first feature tensor and the second feature tensor are fused and input into the enhancement module to obtain the first enhanced feature tensor. The first enhanced feature tensor is then input into the enhancement module to obtain the second enhanced feature tensor. The two enhancement modules are connected in series; The first and third enhanced feature tensors are fused and input into the enhancement module to obtain the third enhanced feature tensor, and the third enhanced feature tensor is input into the enhancement module to obtain the fourth enhanced feature tensor; The third and fourth enhanced feature tensors are fused and input into the enhancement module to obtain the fourth enhanced feature tensor, and the fifth enhanced feature tensor is input into the enhancement module to obtain the sixth enhanced feature tensor; The first, second, third, fourth, fifth, and sixth enhanced feature tensors are passed through a fully connected layer to output the seventh enhanced feature tensor. The first feature tensor and the seventh enhanced feature tensor are concatenated to obtain the fourth feature tensor, which is then used as the input to the next module.

7. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 6, characterized in that, The working principle of the enhancement module is as follows: Obtain the second input feature tensor, and then input the second input feature tensor into two parallel DWConv modules after Conv convolution to obtain the first convolution feature tensor and the second convolution feature tensor. Then, fuse the first convolution feature tensor and the second convolution feature tensor and then convolve them again to obtain the third convolution feature tensor. The third convolutional feature tensor is input into two parallel DWConv modules to obtain the fourth and fifth convolutional feature tensors. The fifth convolutional feature tensor is then convolved and processed by the ReLU activation function to obtain the sixth convolutional feature tensor, which is then used as the input to the next module.

8. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 5, characterized in that, The working principle of the second target module is as follows: Obtain the third input feature tensor, input the third input feature tensor into the CBS module to obtain the first target feature tensor, and input the first target feature tensor into two parallel CBS modules to obtain the second target feature tensor and the third target feature tensor respectively; The third target feature tensor is input into a CBS module and two cascaded DSConv modules respectively to obtain the fourth and fifth target feature tensors. The fourth and fifth target feature tensors are then fused and input into the sixth target feature tensor. The second and sixth target feature tensors are concatenated and normalized to obtain the seventh target feature tensor, which is then used as the input for the next module.

9. The online detection method for substation pressure plates based on trusted WLAN communication according to claim 1, characterized in that, The reliable state result of the pressure plate is obtained by performing a correlation consistency verification on the state detection results based on the operational data, including: Acquire real-time operating data of the target relay protection circuit corresponding to the target pressure plate; the real-time operating data includes at least one or more of the following: circuit current data, circuit voltage data, input / output status data, protection action flag data, and remote signaling feedback data; Based on the association mapping relationship, the state detection result is converted into the corresponding loop expected state, and the time consistency feature between the loop expected state and the real-time running data is constructed; The temporal consistency characteristics are evaluated using a multi-factor fusion method to obtain an association consistency score that includes state matching degree, temporal synchronization degree, and event correlation degree. When the correlation consistency score meets the preset confidence threshold, the state detection result is determined as the confidence state result of the pressure plate; When the correlation consistency score does not meet the preset confidence threshold, a secondary verification process based on multi-frame image review and loop data resampling is triggered to output the corrected confidence status result.

10. An online detection system for substation pressure plates based on trusted WLAN communication, characterized in that, The system includes: The image enhancement module is used to periodically acquire the original image of the target pressure plate, perform enhancement operations on the original image, and output an enhanced image; the original image is an RGB image. A state detection module is used to input the enhanced image into the pressure plate state detection model to obtain the state detection result; the state detection model is based on YOLOv11 and is improved by replacing C3K2 in the YOLOv11 backbone network with the first target module; the state detection result includes at least one of the pressure plate deployment / retraction state, pressure plate pose state, and pressure plate defect state; The result verification module is used to obtain the operating data of the target relay protection circuit corresponding to the state detection result according to the association mapping relationship, and to perform association consistency verification on the state detection result based on the operating data to obtain the reliable state result of the pressure plate; the association mapping relationship is the relationship between the substation pressure plate and the corresponding relay protection circuit that is constructed in advance. The data upload module is used to send the trusted status results to the cloud monitoring platform for online monitoring via a trusted WLAN.

Images