Method for monitoring position state of circuit breaker switch based on video identification technology

Abstract

The invention relates to the technical field of circuit breaker switch monitoring, and provides a method for monitoring the position state of a circuit breaker switch based on a video recognition technology, and the method comprises the steps: collecting a real-time image of a circuit breaker through a camera device, judging the physical state of the circuit breaker, and determining a first judgment result; acquiring real-time data of the relay protection device, recording the operation state of the circuit breaker, and determining a second judgment result; inputting the first judgment result and the second judgment result into a preset signal response device to generate a corresponding response action; wherein the response action comprises an opening action, a fault action and a closing action.

Description

Technical Field

[0001] The present invention relates to the field of circuit breaker switch monitoring, and particularly to a method for monitoring the position state of a circuit breaker switch based on video recognition technology. Background Art

[0002] With the rapid development of smart grids, circuit breakers, as key protection and control devices in power systems, the real-time monitoring of their operating states and the ability to respond quickly directly affect the security and reliability of the power grid. Traditional circuit breaker state monitoring mainly relies on mechanical position sensors (such as auxiliary contacts, micro switches) or electrical quantity measurements (such as closing and opening coil currents), but there are the following technical defects:

[0003] Sensor limitations: Mechanical sensors are easily affected by factors such as environmental corrosion, vibration aging, etc., resulting in contact oxidation or poor contact, generating false signals; while pure electrical quantity monitoring cannot intuitively reflect the physical position of the switch (such as hidden faults like virtual contact of contacts, mechanical jamming, etc.).

[0004] Poor environmental adaptability: Existing optical detection methods (such as laser ranging, infrared temperature measurement) have a significant decrease in recognition accuracy under complex lighting, rainy and foggy weather, or metal reflection interference, and it is difficult to meet the all-weather monitoring requirements of outdoor substations.

[0005] Information island problem: The circuit breaker state data lacks deep coordination with the relay protection system. The protection action decision only depends on electrical parameters (such as overcurrent, differential protection), and it is impossible to combine the physical state of the switch (such as incomplete separation of contacts) for compound fault judgment, resulting in misoperation or refusal of protection.

[0006] Insufficient intelligence level: Traditional image processing algorithms (such as threshold segmentation, edge detection) rely on artificial feature design, are difficult to adapt to dynamic scenarios such as equipment fouling and uneven lighting, and lack self-optimization ability, requiring frequent manual calibration.

[0007] In recent years, related technical improvement attempts include:

[0008] Patent CN202010730692.0 proposes circuit breaker state monitoring based on infrared thermal imaging, and judges poor contact of contacts through abnormal temperature, but does not solve the problem of image noise caused by metal shell reflection.

[0009] The literature "Switchgear State Recognition Based on Deep Learning" (Automation of Electric Power Systems, 2021) uses a convolutional neural network (CNN) to identify the closing and opening states of circuit breakers, but its model does not fuse multi-spectral data and has limited generalization ability in low-light environments.

[0010] In summary, the prior art has not yet achieved the deep integration of high-robustness visual monitoring and intelligent decision-making for relay protection. There are significant technical bottlenecks, especially in terms of complex environment adaptability, multi-modal data collaboration, and automated closed-loop control. Therefore, there is an urgent need for an innovative technical solution to improve the accuracy, real-time performance, and reliability of circuit breaker status judgment through the deep integration of video recognition and protection systems, providing a higher level of active defense capabilities for smart grids. Summary of the Invention

[0011] The present invention proposes a method for monitoring the switch position status of a circuit breaker based on video recognition technology, aiming to solve the problem of poor monitoring results in traditional circuit breaker status monitoring, which mainly relies on mechanical position sensors such as auxiliary contacts and electrical quantity measurements of micro switches such as closing and opening coil currents.

[0012] The present invention proposes a method for monitoring the switch position status of a circuit breaker based on video recognition technology, including:

[0013] Collecting real-time images of the circuit breaker through a camera device, judging the physical state of the circuit breaker, and determining the first judgment result;

[0014] Obtaining real-time data of the relay protection device and recording the operating state of the circuit breaker to determine the second judgment result;

[0015] Inputting the first judgment result and the second judgment result into a preset signal response device to generate corresponding response actions; where the response actions include: opening action, fault action, and closing action.

[0016] Further, the real-time images include visible light images, near-infrared images, depth thermal maps, and images after supplementary lighting;

[0017] Further, collecting real-time images of the circuit breaker through a camera device further includes:

[0018] Pre-constructing a multi-clarity scene database and performing adaptive optical analysis to determine the clarity state of the current scene;

[0019] Among them, the adaptive optical analysis includes multi-spectral imaging analysis, dynamic polarization analysis, and interference light source array analysis;

[0020] According to the multi-spectral imaging analysis, calculating the inter-frame error between the visible light image and the near-infrared feature image in the real-time image under alternating acquisition;

[0021] According to the dynamic polarization analysis, calculating the real-time polarization angle of the rotatable linear polarizer array and the annular phase retarder of the camera device;

[0022] According to the interference light source array analysis, generating a structured light distribution array of the current scene and determining the area to be supplemented with light that does not meet the preset clarity;

[0023] Build an image enhancement network for real-time images based on the inter-frame error, polarization angle, and the area to be filled with light, and perform enhancement processing on the real-time images.

[0024] Furthermore, the image enhancement network includes an input layer, a data alignment sub-network, a physical feature encoder, a conditional modulation enhancement unit, and a cascaded enhancement strategy unit; among them,

[0025] The input layer includes an image data for receiving real-time images through a signal channel; among them, the signal channel includes a visible light image channel, a near-infrared image channel, a polarization angle tensor channel, and a structured light mask channel;

[0026] The data alignment sub-network is used to perform the first enhancement processing on the image data through a spatio-temporal alignment network to eliminate the inter-frame error;

[0027] The physical feature encoder is used to extract different modal features in the image data through a multi-branch parallel structure to generate a multi-modal feature map;

[0028] The conditional modulation enhancement unit is used to use the multi-modal feature map and the inter-frame error as conditional vectors to determine the image area to be enhanced;

[0029] The cascaded enhancement strategy unit is used to stack multiple dynamic enhancement blocks to perform enhancement processing on image areas of different scales through the multiple dynamic enhancement blocks.

[0030] Furthermore, the multi-branch parallel structure includes: a spectral branch, a polarization branch, and a fill-light branch;

[0031] Among them, the spectral branch is used to extract and separate the specular reflection feature and the texture feature through channel attention;

[0032] The polarization branch is used to extract the rotation weight feature guided by the polarization angle of the imaging device and suppress the specular reflection feature of the metal through directionally separable convolution;

[0033] The fill-light branch is used to determine the image gray deviation feature through mask-guided local contrast enhancement.

[0034] Furthermore, the determining the physical state of the circuit breaker and obtaining the first judgment result includes:

[0035] Build a multi-modal feature fusion network, input the visible light image, near-infrared image, depth thermal map, and the image after fill-light into the spatio-temporal alignment module, and fuse the texture feature, thermodynamic feature, and polarization suppression feature through a cross-modal attention mechanism to generate a high-confidence circuit breaker state feature map;

[0036] Based on a preset physical constraint state decision tree, perform hierarchical judgment on the key areas in the high-confidence circuit breaker state feature map:

[0037] The first layer of judgment: Identify the closed / separated state of the contact according to the near-infrared reflectance and temperature rise gradient in the contact area;

[0038] The second layer of judgment: Detect mechanical jamming or deformation based on the contour integrity of the mechanical interlock device in the image after polarization suppression;

[0039] The third layer of judgment: Determine the surface contamination level of the insulating component through the abnormal ratio of the local contrast between the supplementary light area and the original image;

[0040] Output a circuit breaker physical state report that fuses multi-modal judgment results, and determine the first judgment result. The first judgment result includes a three-dimensional state code of the switch position, mechanical health, and insulation status.

[0041] Furthermore, obtain the real-time data of the relay protection device, record the operating state of the circuit breaker, and determine the second judgment result, including:

[0042] Based on the relay protection device, configure a multi-source data acquisition device, and obtain the real-time electrical parameters of the relay protection device and the multi-dimensional operating state data of the circuit breaker;

[0043] Construct a multi-source data fusion model with time series alignment, and perform dynamic collaborative analysis on the real-time electrical parameters and multi-dimensional operating state data to obtain a fault-action mapping matrix;

[0044] Perform multi-dimensional anomaly detection on the fault-action mapping matrix through an autoregressive anomaly propagation model:

[0045] Generate a second judgment result including the root cause tracing of the fault, and output it in a four-tuple coding format.

[0046] Furthermore, the real-time electrical parameters include the waveform of the protection action signal, the amplitude of the fault current, the impedance of the tripping circuit, and the harmonic distortion rate;

[0047] The multi-dimensional operating state data includes the mechanical vibration spectrum of the circuit breaker, the closing and opening coil current curves, the pressure value of the energy storage mechanism, and the wear degree index of the switch contact.

[0048] Furthermore, the autoregressive anomaly propagation model includes a deviation warning layer, a risk identification layer, and a fault location layer; among them,

[0049] The deviation warning layer is used to compare the deviation between the moment when the tripping instruction is issued and the actual opening time of the circuit breaker, and trigger a mechanical response delay warning when the deviation exceeds the preset threshold;

[0050] The risk identification layer is used to analyze the non-linear relationship between the number of operations of the protection device and the contact electrical life loss curve, and identify the hidden fault risk caused by abnormal frequent operations;

[0051] The fault location layer is used to locate the fault positions of mechanism jamming or spring fatigue based on the coupled analysis of vibration spectrum entropy value and energy storage pressure volatility.

[0052] Further, the inputting the first judgment result and the second judgment result into a preset signal response device to generate corresponding response actions includes:

[0053] Receiving the three-dimensional state encoding of the first judgment result and the quaternion encoding of the second judgment result;

[0054] Constructing a multi-modal response decision model to dynamically fuse the encoded data and generate a response decision set;

[0055] Generating a composite response strategy based on the quaternion encoding of the fault root cause traceability;

[0056] Generating response actions according to the composite response strategy and the triggered response decisions in the response action set.

[0057] Other features and advantages of the present invention will be described in the subsequent specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structures specifically pointed out in the written specification and the accompanying drawings.

[0058] The technical solution of the present invention will be further described in detail below through the accompanying drawings and embodiments. Description of the Drawings

[0059] The 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 to the present invention. In the drawings:

[0060] Figure 1 is the flowchart of a method for monitoring the position state of a circuit breaker switch based on video recognition technology in an embodiment of the present invention;

[0061] Figure 2 is the process flowchart of collecting real-time images of a circuit breaker by a camera device in an embodiment of the present invention;

[0062] Figure 3 is the function block diagram of an image enhancement network in an embodiment of the present invention;

[0063] Figure 4 is the association diagram of a multi-branch parallel structure in an embodiment of the present invention;

[0064] Figure 5 is the implementation flowchart of the first judgment result in an embodiment of the present invention;

[0065] Figure 6It is the implementation flowchart of the second determination result in the embodiment of the present invention;

[0066] Figure 7 It is the function diagram of real-time electrical parameters and multi-dimensional operating state parameters in the embodiment of the present invention;

[0067] Figure 8 It is the function diagram of the autoregressive anomaly propagation model in the embodiment of the present invention;

[0068] Figure 9 It is the flowchart for generating corresponding response actions in the embodiment of the present invention. Detailed implementation manners

[0069] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0070] Embodiment 1:

[0071] Based on the defects in the traditional technical solutions, this application proposes a method for monitoring the position state of a circuit breaker switch based on video recognition technology. Refer to Figure 1 :

[0072] First, for the traditional method that relies on limit switches or current / voltage sensors and is prone to misjudgment due to mechanical wear or electromagnetic interference. This application collects real-time images of the circuit breaker through a camera device and judges the physical state of the circuit breaker to determine the first judgment result; for example: collecting real-time images of the circuit breaker through a camera, such as: object detection, state classification, image segmentation, etc., to analyze the physical state of the circuit breaker, such as opening / closing position, etc.; judging the physical state of the circuit breaker through non-contact video monitoring, avoiding the problems of traditional mechanical / electrical sensors being vulnerable to environmental interference and complex installation.

[0073] For the traditional method that only relies on a single electrical data source and cannot cope with extreme scenarios such as sensor failures or communication interruptions. This application is based on the real-time data of the relay protection device and records the operating state of the circuit breaker to determine the second determination result; aiming to combine the electrical parameters during the operation of the circuit breaker, such as: opening / closing current, protection signal, to provide double verification and enhance the reliability of state judgment.

[0074] In traditional technology cases, an independent judgment logic is adopted, lacking the ability of multi-dimensional collaborative analysis. In this application, the first judgment result and the second judgment result are input into a preset signal response device to generate corresponding response actions. Among them, the response actions include: opening action, fault action, and closing action. That is, by fusing visual and electrical data, more accurate response actions (such as fault alarms, automatic locking) are generated, reducing the risks of misoperation and refusal to operate. The purpose is to reject the unreliability problem of a single data source, thereby improving the misjudgment rate, misjudgments of mechanical jamming and electrical faults. At the same time, whether it is visual detection or electrical detection, each has its own high efficiency effect.

[0075] The focus of this application is to form a dual-redundant detection logic. The signal response device needs to rely on two types of data inputs at the same time and generate response actions through logical operations. There is a contradiction between the real-time requirement of video recognition, which requires high-frame-rate processing, and the high-reliability requirement of relay protection data, which requires strict synchronization timing. This application solves the data synchronization problem through a timestamp alignment algorithm; by combining the frozen frames of the video capture mechanism and the abnormal coil current in the electrical data, the fault type can be identified.

[0076] During actual implementation:

[0077] Through the real-time data of the relay protection device. For example: current, voltage, protection signal, combined with logical rules or timing analysis to judge the operating state of the circuit breaker.

[0078] By presetting logic for visual judgment and sensor data, such as: weighted voting fusion, Bayesian network fusion, to solve the limitations of a single data source and improve the robustness of state judgment.

[0079] Identify the mechanical position of the circuit breaker through video recognition, such as: the opening and closing angle of the disconnecting switch, and cross-verify it with the electrical state recorded by the relay protection device, such as: whether the current is zero, to identify potential abnormalities, such as misjudgment of electrical signals caused by mechanical jamming.

[0080] During actual implementation, a decision tree can be designed based on dual-source data input, such as:

[0081] Both vision and sensor determine "switch off" → generate a closing action;

[0082] Vision shows "switch on" but there is no current in the sensor → trigger a fault action;

[0083] The sensor alarms but the visual state is normal → start a secondary verification process.

[0084] The beneficial effects of the above technical solutions are as follows:

[0085] This application adopts a dual verification mechanism to reduce the misjudgment rate and achieve the integration of the two. Traditional sensors are vulnerable to electromagnetic interference, resulting in false alarms, while video verification can eliminate false alarms. When the vision system is blocked / stained, sensor data can provide redundant backup. When a fault occurs, the video record is aligned with the electrical data timestamp, enabling the backtracking of key states before the fault and assisting in analyzing the reasons for circuit breaker refusal / malfunction. Compared with manual inspection, it realizes 7×24-hour automatic monitoring; through early anomaly detection, such as action delay caused by contact oxidation, it avoids unplanned power outages.

[0086] Embodiment 2:

[0087] The real-time images include visible light images, near-infrared images, depth heat maps, and images after supplementary lighting.

[0088] In actual implementation of this application, different technical observation details can be obtained based on different real-time images. Visible light images provide high-resolution visual information on the switch position of the circuit breaker, directly observing the physical states of mechanical components, such as the physical states of contacts and connecting rods. Traditional solutions only rely on visible light images and are easily interfered by environmental light changes (such as strong reflections and shadows), resulting in misjudgments. Near-infrared images capture the circuit breaker state in low-light or nighttime environments, penetrating partial obstructions, such as dust and fog, which can enhance environmental adaptability. Depth heat maps detect abnormal heat generation at the contact points of the circuit breaker through temperature distribution to give early warnings of faults. Traditional infrared temperature measurement is mainly based on single-point temperature and cannot globally perceive temperature and local overheating. The images after supplementary lighting eliminate shadow or reflection interference through controllable light sources, improving image consistency and ensuring algorithm stability.

[0089] In actual implementation, technical solutions can also be integrated. When the visible light image shows that the contact is closed, but the heat map shows high temperature at the contact point, a "poor contact" alarm is triggered.

[0090] Embodiment 3:

[0091] This application addresses the problems of the clarity and anti-interference of circuit breaker images in complex environments in traditional technical solutions, and proposes a method of collecting real-time images of circuit breakers through camera devices; refer to Figure 2 :

[0092] In view of the fact that in traditional methods, optical filtering relying on fixed parameters and the inability to adapt to dynamic complex scenes, such as oil stain scenes, this application determines the clarity state of the current scene by pre-constructing a multi-clarity scene database and performing adaptive optical analysis; by pre-collecting circuit breaker images under different lighting, occlusion, and pollution scenes, a training data set is constructed to enable the image enhancement network to have environmental self-adaptability and reduce the risk of model overfitting.

[0093] In this process, the means of adaptive optical analysis include multispectral imaging analysis, dynamic polarization analysis, and interfering light source array analysis;

[0094] For traditional single-spectrum, the superposition effect of natural light and device supplementary light cannot be distinguished. According to multispectral imaging analysis, this application calculates the inter-frame error between the visible light image and the near-infrared characteristic image of the real-time image under alternating acquisition; by alternately acquiring visible light and near-infrared images, the difference between the two is calculated, such as calculating the edge offset and contrast change to identify environmental light interference;

[0095] For the polarization filter with a fixed angle that cannot adapt to multi-angle specular reflection scenarios, according to dynamic polarization analysis, this application calculates the real-time polarization angle of the rotatable linear polarizer array and the annular phase retarder of the imaging device, that is, by dynamically adjusting the polarizer array and the phase retarder, the specular reflection on the metal surface is suppressed, such as the specular reflection interference on the circuit breaker housing, to improve the clarity of image details.

[0096] For traditional supplementary light that causes light pollution or insufficient brightness in key areas due to global uniform illumination, according to interfering light source array analysis, this application generates a structured light distribution array for the current scene, determines the area to be supplemented with light that does not meet the preset clarity, that is, generates a programmable supplementary light mode (such as a grid-shaped light spot), identifies the area to be enhanced (such as the contact position covered by the shadow), and realizes local precise supplementary light.

[0097] For traditional supplementary light that uses a fixed architecture and cannot adapt to scenarios with sudden changes in the light mask or local occlusion, an image enhancement network for real-time images is built based on the inter-frame error, polarization angle, and area to be supplemented with light, and the real-time image is enhanced. This application combines multispectrum (spectral dimension), dynamic polarization (polarization dimension), and structured supplementary light (spatial dimension) to form a multi-physical field collaborative optimization mechanism. The image enhancement network is dynamically adjusted through real-time analysis results to form a closed-loop perception, analysis, and measurement to solve timing synchronization and calculation delay. In terms of anti-interference, specular reflection is suppressed through dynamic polarization, combined with structured supplementary light shadow elimination, and the identification of key areas to improve contrast. The multispectral inter-frame error is used to locate the interference area and limit the operation range of the image enhancement network, which can also reduce the calculation amount. The optimized light distribution of this application can penetrate oil stains or dust, such as near-infrared supplementary light, and cooperate with the local repair of the enhancement network, resulting in a great improvement in the recognition accuracy.

[0098] In actual implementation, this application can improve the image analysis effect, that is, clarity, in complex lighting or dynamic environments such as strong specular reflection, low light, and multi-light source interference by pre-constructing a multi-clarity scene database.

[0099] The database needs to cover various scenarios, such as different times, weather conditions, and equipment aging status, to prevent overfitting. The inter-frame error between visible light and near-infrared images is used to capture dynamic changes, such as equipment heating and surface oxygen, while multi-spectral data enhances the feature separation ability, such as distinguishing metal reflection from stains. The dynamic adjustment of the rotating polarizer and phase retarder can effectively suppress the reflection on the metal surface, which is commonly found on the circuit breaker housing, and improve the visibility of key areas, such as switch contacts. The structured light distribution can actively compensate for areas with insufficient ambient light, such as shadows and backlighting, to ensure uniform illumination in the monitoring area. By combining multi-modal inputs of inter-frame error, polarization angle, and light compensation area, the network can optimize the image quality specifically, such as denoising, deblurring, and enhancing contrast.

[0100] Embodiment 4:

[0101] The image enhancement network of this application has a multi-modal input function. Refer to Figure 3 , although the prior art does not encode optical parameters and input them into the network, this application proposes that the image enhancement network includes an input layer, a data alignment sub-network, a physical feature encoder, a conditional modulation enhancement unit, and a cascading enhancement strategy unit; among them,

[0102] For the single image channel in the traditional solution, which cannot fuse polarization and structured light information, the input layer of this application includes image data for receiving real-time images through signal channels; among them, the signal channels include visible light image channels, near-infrared image channels, polarization angle tensor channels, and structured light mask channels; through the multi-channel input of visible light, near-infrared, polarization tensor, and structured light mask, the parallel fusion of multi-modal data (spectrum, polarization, spatial light compensation) is realized, and the completeness of feature expression is improved.

[0103] For the traditional solution that uses fixed registration parameters, resulting in the inability to dynamically adapt to the jitter of the camera and environmental jitter, the data alignment sub-network of this application is used to perform the first enhancement processing on the image data through the spatio-temporal alignment network, so that the inter-frame error is eliminated. This application ensures the accuracy of feature acquisition by eliminating the spatio-temporal offset of multi-modal images. For example, the polarization angle reflection suppression feature and the near-infrared penetration feature.

[0104] For the traditional encoder that uses a serial structure, which may cause mutual interference of multi-modalities. The physical feature encoder of this application is used to extract different modality features in the image data through a multi-branch parallel structure, generate multi-modal feature maps, and solve the problem of extracting different modality physical features through independent analysis. Retain modality-specific information.

[0105] Since the traditional enhancement unit uses global parameters and cannot optimize for local interference, the conditional modulation enhancement unit of this application is used to take the multi-modal feature map and the inter-frame error as conditional vectors to determine the image regions that need to be enhanced, that is, dynamically adjust the enhancement weights according to the inter-frame error (reflecting the environmental interference intensity) and preferentially enhance the high-error regions.

[0106] Regarding the fixed hierarchy of the traditional network, which cannot adapt to the problem of enhancing the depth with complex dynamic adjustment, the cascaded enhancement strategy unit of this application stacks multiple dynamic enhancement blocks to perform enhancement processing on image regions of different scales through multiple dynamic enhancement blocks, and then realizes hierarchical optimization through multi-scale enhancement blocks (such as small-scale capture of contact details and large-scale recognition of overall positions) to avoid information loss under a single scale.

[0107] This application uses the polarization tensor (physical optical parameter) and the structured light mask (spatial supplementary light requirement) as network input conditions to achieve the collaborative optimization of physical rules and deep learning. Through the collaboration of the conditional modulation unit (local optimization) and the cascaded strategy unit (multi-scale optimization), it solves the global and local problems in complex scenarios. The dynamic enhancement block adaptively selects the enhancement level according to the scene complexity, and while processing in real time, maintains high precision.

[0108] In actual implementation, this application maps multi-spectral imaging visible light + near-infrared, dynamic polarization angle, and structured light supplementary light regions into multi-channel inputs, not just relying on RGB images.

[0109] Based on the results of adaptive optical analysis, for example: inter-frame error, polarization angle, dynamically adjust network parameters, and the scene adaptability will be enhanced.

[0110] Introduce physically interpretable constraints, such as specular reflection suppression and supplementary light consistency, to obtain an adversarial network (GAN), which can avoid overfitting of pure data-driven.

[0111] In the specific process, the input layer is a visible light image (3 channels), a near-infrared image (1 channel), a polarization angle tensor (2 channels), and a structured light mask (1 channel) → a total of 7-channel input. The data alignment sub-network uses a lightweight spatio-temporal alignment network (such as deformable convolution based on optical flow) to eliminate the inter-frame offset between multi-spectral and polarization data. The conditional modulation enhancement unit input: multi-modal feature map + real-time polarization angle / inter-frame error parameter (as a conditional vector). Stack multiple dynamic enhancement blocks, and each layer focuses on features of different scales (from global illumination compensation to local detail restoration).

[0112] When actually implemented, it is improved based on U-Net and introduces:

[0113] Polarization physical loss constraint to generate the consistency of the specular reflection distribution between the image and the polarization angle prediction.

[0114] Multi - spectral consistency loss, cross - modal alignment of visible light and near - infrared enhancement results. A multi - scale PatchGAN structure, with an additional input of a structured light mask as a condition to distinguish the rationality of fill light for real / generated images.

[0115] Example 5:

[0116] This application proposes a method for determining the characteristics affecting gray - scale deviation. Refer to Figure 4 , during the implementation process, the multi - branch parallel structure includes: a spectral branch, a polarization branch, and a fill - light branch;

[0117] In view of the fact that in traditional technical solutions, using a fixed filter may not be able to distinguish between specular reflection and real texture, the spectral branch of this application is used to extract and separate specular reflection features and texture features through channel attention. When implemented, the specular reflection features and real texture features are separated through channel attention, thereby improving the recognition accuracy of key areas. Specular reflection features include the highlights on the metal surface, and real texture features are unearthed oxidation traces, etc.

[0118] In view of the fact that the direction of the traditional convolution kernel is fixed and only applicable to specific polarization angles, with poor generalization ability, the polarization branch of this application is used to extract the rotation weight features guided by the polarization angle of the imaging device and suppress the metal specular reflection features through direction - separable convolution; furthermore, based on the polarization angle, the direction of the convolution kernel is dynamically adjusted to suppress multi - angle metal specular reflection and retain edge details. The polarization branch embeds the physical parameter of the polarization angle into the direction - separable convolution to achieve the collaborative design of optical characteristics and network structure;

[0119] In view of the fact that the global operation of traditional contrast enhancement may cause noise enhancement in non - visible areas, the fill - light branch of this application is used to determine the gray - scale deviation characteristics of the image through mask - guided local contrast enhancement. Furthermore, a structured light mask is used to locate low - contrast areas, and the gray - scale distribution can be adjusted specifically to prevent over - exposure caused by global enhancement.

[0120] The polarization branch suppresses specular reflection, and it is easy to separate real texture in combination with the spectral branch. The fill - light branch accurately enhances the local contrast, causing the error of any link to be amplified step by step, thereby judging the error and enhancing the recognition of error areas.

[0121] During actual implementation, the physical feature encoder uses a multi - branch parallel structure to extract different - modality features:

[0122] For example: Spectral branch: 3×3 convolution + channel attention (ECA - Net) → separate specular reflection and texture.

[0123] Polarization branch: Direction - separable convolution + rotation weight guided by polarization angle → suppress metal specular reflection.

[0124] Fill - light branch: Mask - guided local contrast enhancement → restore shadow areas.

[0125] Example 6:

[0126] This application will also judge the physical state of the circuit breaker to determine the first judgment result. In this process, hierarchical judgment needs to be carried out. Refer to Figure 5 :

[0127] In view of the fact that the traditional method relies on a single modality and cannot cope with complex interferences, such as misjudgment caused by the superposition of reflection and high temperature, this application constructs a multi-modal feature fusion network. The visible light image, near-infrared image, depth thermal map and post-supplementary light image are input into the spatio-temporal alignment module. Through the cross-modal attention mechanism, the texture features, thermodynamic features and polarization suppression features are fused to generate a high-confidence circuit breaker state feature map. In this process, through the cross-modal attention mechanism, the texture of visible light, the penetrability of near-infrared light, the temperature of the thermal map and the anti-interference of the supplementary light image are fused to generate a high-confidence feature map, covering multi-dimensional information of machinery, thermodynamics and electricity, which can improve the comprehensiveness of state judgment. The spatio-temporal alignment module fixes the registration parameters in the traditional technology and cannot dynamically adapt to equipment vibration and environmental vibration, and can eliminate the spatio-temporal offset caused by the acquisition time difference or perspective difference of multi-modal images, ensuring the accuracy of feature fusion.

[0128] In view of the single-threshold judgment in the traditional scheme, which cannot distinguish mechanical faults and electrical abnormalities, based on the preset physical constraint state decision tree, hierarchical judgment is carried out on the key areas in the high-confidence circuit breaker state feature map:

[0129] First-layer judgment: According to the near-infrared reflectivity and temperature rise gradient in the contact area, identify the contact closing / separation state. This application adopts the combination of infrared reflectivity, that is, the contact resistance abnormality of the contact material characteristics and temperature rise gradient, which can accurately distinguish the closing / separation state;

[0130] Second-layer judgment: Based on the integrity of the mechanical interlock device contour in the post-polarization suppression image, detect mechanical jamming or deformation. In the second-layer judgment of this application, the contour of the interlock device is analyzed through the post-polarization suppression image. For example, in the case of deformation and jamming, metal reflection interference is avoided.

[0131] Third-layer judgment: Determine the surface contamination level of the insulating component through the local contrast abnormality ratio between the supplementary light area and the original image. This application uses the local contrast abnormality between the supplementary light area and the original image to detect that the surface insulation performance will decrease due to dirt.

[0132] Regarding the traditional output which is either on or off state and cannot provide guiding opinions, this application will output a breaker physical state report that integrates multi-modal judgment results, determine the first judgment result, and the first judgment result includes a three-dimensional state code of switch position, mechanical health, and insulation state. Through this application, the codes of the switch position, mechanical health, and insulation state can be unified as decision vectors to achieve multi-dimensional fault diagnosis.

[0133] In actual implementation, this application is based on multi-spectral data of visible light, near-infrared (NIR), and depth thermal maps, covering different physical characteristics, such as

[0134] Capture the surface texture and oxidation traces of the breaker through visible light. For example, the oxidized area may show color changes or spots under visible light;

[0135] Detect the contact state of the contacts by near-infrared penetrating the surface stains. For example, the near-infrared reflectivity of metal oxides is significantly lower than that of normal metals;

[0136] Monitor the temperature rise at the contact point through the thermal map. For example, the Joule heat generated by the contact resistance in the closed state.

[0137] Reduce motion blur through the alternate exposure mechanism by dynamically adjusting the exposure time. For example, short exposure of visible light + long exposure of near-infrared, and use the inter-frame error matrix, that is, calculate the difference between adjacent frames, and then extract the oxidized area. The oxidized area has abnormal heat conduction due to oxidation, and the temperature rise gradient is significantly different from the normal area.

[0138] The reflection of the metal surface of the breaker will cover up key details. For example, contact wear and micro-cracks. Through the coordinated control of the rotatable linear polarizer array and the phase retarder, the polarization direction is switched at a preset frequency to suppress specular reflections at different angles; generate a polarization angle - reflectivity curve, and the reflectivity of the metal oxide or rough surface changes more sensitively with the polarization angle, so as to distinguish the normal and abnormal areas.

[0139] Benshenqing dynamically generates a structured light distribution according to the inter-frame error matrix, such as the oxidized area, and the polarization reflectivity mapping, such as the requirements for specular reflection suppression, fills light in the shadow area, fills light for the near-infrared band, and more clearly shows the details of the contact surface of the contacts.

[0140] Fill light for the oxidized area, and use a specific wavelength of visible light to enhance the contrast of oxidation spots.

[0141] Calculate the local contrast anomaly ratio through the comparison of the images before and after filling light. For example, the brightness increase ratio of the shadow area after filling light, so as to quantify the degree of pollution or insulation deterioration.

[0142] This application uses timestamp synchronization and feature point matching, such as the SIFT algorithm, to ensure the spatio-temporal consistency of multi-modal data;

[0143] Design a multi-head attention layer to dynamically allocate weights. For example, the weight ratio of thermal features in the contact state judgment is 70%, and the weight ratio of polarization features in the mechanical interlock detection is 60%;

[0144] Furthermore, by fusing texture (visible light), thermodynamics (thermal map), and polarization suppression (reflectivity) features, a high-confidence feature map is generated. Based on the near-infrared reflectivity threshold, the normal contact reflectivity > 80%, and after oxidation < 50%, as well as the temperature rise gradient, the temperature rise in the closed state > 3 °C / s, logical rules are constructed.

[0145] In specific implementation, through the contour integrity analysis of the image after polarization suppression:

[0146] Edge detection: Use the Canny algorithm to extract the contour of the interlock device;

[0147] Deformation determination: Calculate the Hausdorff distance between the contour curvature and the standard template (if > 2 mm, it is determined as deformed).

[0148] Insulation pollution level classification;

[0149] According to the abnormal contrast ratio of the supplementary light area:

[0150] Pollution level 1 (clean): CR ≤ 1.2;

[0151] Pollution level 2 (mild): 1.2 < CR ≤ 1.5;

[0152] Pollution level 3 (severe): CR > 1.5.

[0153] Example 7:

[0154] This application proposes a determination method for fault traceability. Refer to Figure 6 , obtain the real-time data of the relay protection device, record the operating state of the circuit breaker, and determine the second determination result. Specifically:

[0155] Aiming at the problem that traditional single data sources cannot identify mechanical and electrical correlation faults, this application uses a relay protection device, configures a multi-source data acquisition device, and obtains the real-time electrical parameters of the relay protection device and the multi-dimensional operating state data of the circuit breaker. By integrating the real-time electrical parameters of the relay protection device, such as current, voltage, and trip signals, cross-domain data fusion of the multi-dimensional motion state data of the circuit breaker, such as mechanical vibration, temperature, and operation times, can be achieved, improving the comprehensiveness of fault detection.

[0156] In view of the data misalignment problem caused by using fixed-time series hooks, this application constructs a multi-source data fusion model with time series alignment to dynamically analyze the real-time electrical parameters and multi-dimensional operating state data in a coordinated manner. For the fault-action mapping matrix, this application uses the dynamic time warping algorithm to align electrical and mechanical data with different sampling rates, thereby eliminating the error of time series offset and maintaining the accuracy of data collaborative analysis.

[0157] In view of the fact that the traditional solution relies on a static rule matrix and cannot adapt to all fault types, this application performs multi-dimensional anomaly detection on the fault-action mapping matrix through an autoregressive anomaly propagation model. That is, by means of time series autoregression, the propagation path of abnormal signals is traced, and then a quadruple code is generated to generate encoded data that can be traced. Moreover, the output is not a simple text warning, but a decision priority ranking based on the structured encoding of faults, locations, times, and confidence levels. The autoregressive model introduces time-dimensional causal relationship analysis (such as the abnormal propagation path), while the traditional solution only statically associates faults with actions.

[0158] Generate a second determination result including the root cause tracing of faults, and the output is in the format of a quadruple code.

[0159] In the actual implementation process, the real-time electrical parameters in this application include the waveform of the protection action signal, the amplitude of the fault current, the impedance of the tripping circuit, and the harmonic distortion rate; the multi-dimensional operating state data includes the mechanical vibration spectrum of the circuit breaker, the current curve of the closing and opening coils, the pressure value of the energy storage mechanism, and the wear degree index of the switch contacts; the sliding window mechanism is used to extract the change gradient of the mechanical characteristics of the circuit breaker within the ±100 ms time window before and after the protection action event; based on the dynamic weight adjustment algorithm, the collaborative correlation degree between the electrical fault characteristics and the mechanical state parameters is calculated to generate a spatio-temporally synchronized fault-action mapping matrix; after multi-dimensional anomaly detection, the coding dimensions include the electrical protection matching degree, the mechanical action reliability, the probability of hidden defects, and the recommended maintenance priority; the blockchain evidence storage module generates an immutable time series correlation hash value for the determination result and the original data.

[0160] Example 8:

[0161] This application is directed to multi-micro cloud state data. Refer to Figure 7 , specifically, the real-time electrical parameters include the waveform of the protection action signal, the amplitude of the fault current, the impedance of the tripping circuit, and the harmonic distortion rate;

[0162] The multi-dimensional operating state data includes the mechanical vibration spectrum of the circuit breaker, the current curve of the closing and opening coils, the pressure value of the energy storage mechanism, and the wear degree index of the switch contacts.

[0163] In the traditional solution, only the action status is recorded, and the hidden faults caused by waveform distortion cannot be analyzed. This application captures the timing characteristics of protection actions (such as tripping delay, signal jitter) to identify misoperations or refusals to operate. The traditional peak detection ignores the current change trend. This application quantifies the severity of short circuits / overloads and combines waveform analysis to distinguish instantaneous overcurrent from persistent faults. In the traditional solution, the impedance is only manually measured during regular maintenance. This application detects abnormal contact resistance in the circuit to prevent tripping failures. For the harmonic distortion rate, it will identify abnormal heating and misoperations of circuit breakers caused by grid harmonics. Multidimensional status data can detect hidden faults such as mechanical jamming and connecting rod deformation through vibration frequency analysis, analyze the rise time / peak value of coil current to judge the health of electromagnetic mechanisms, monitor the spring / hydraulic energy storage status in real time, prevent slow opening and closing caused by insufficient pressure, and estimate the remaining life of contacts based on the number of arcs and current integration to achieve predictive maintenance.

[0164] Example 9:

[0165] For deviation calculation, risk identification, and fault location, this application adopts an autoregressive anomaly propagation model including a deviation warning layer, a risk identification layer, and a fault location layer; in specific implementation, refer to Figure 8 ;

[0166] Because the traditional technical solution only records the opening time and cannot associate the difference between the instruction and the action timing, the deviation warning layer of this application is used to compare the deviation between the moment when the tripping instruction is issued and the actual opening time of the circuit breaker. When the deviation exceeds the preset threshold, it triggers a mechanical response delay warning, that is, through the deviation detection between the tripping instruction and the actual opening time (such as the delay caused by mechanical jamming), it triggers a mechanical response delay warning in real time to avoid protection failure.

[0167] For the traditional solution that uses linear life prediction and cannot identify non-linear losses, the risk identification layer of this application is used to analyze the non-linear relationship between the number of operations of the protection device and the contact electrical life loss curve, and identify the risk of hidden faults caused by abnormal frequent operations. During implementation, by establishing a non-linear model of the number of operations and contact electrical life loss (such as an exponential decay curve), it identifies the hidden damage of contacts (such as metal fatigue microcracks) caused by abnormal frequent operations.

[0168] For the traditional vibration analysis that only relies on the amplitude threshold and cannot distinguish jamming from normal vibration, the fault location layer of this application is used for coupling analysis based on the vibration spectrum entropy value and the energy storage pressure volatility to locate the fault position of mechanism jamming or spring fatigue. This application accurately locates the fault position (such as connecting rod jamming or spring fatigue) through the coupling analysis of the vibration spectrum entropy value (reflecting mechanical looseness) and the energy storage pressure volatility (reflecting spring stiffness).

[0169] Example 10:

[0170] This application is for the recognition of response actions. Refer to Figure 9 , and input the first judgment result and the second judgment result into a preset signal response device to generate corresponding response actions, including:

[0171] Receive the three-dimensional state encoding of the first judgment result and the quaternion encoding of the second judgment result, that is, integrate the three-dimensional encoding of the physical state of the circuit breaker: switch position, mechanical health, insulation status, and fault root cause, and the quaternion encoding: fault type, location, time, confidence level, to form a full-dimensional decision basis to support the handling of complex faults.

[0172] Construct a multi-modal response decision model to dynamically fuse the encoded data and generate a set of response decisions, that is, through dynamic weight allocation, such as giving priority to handling mechanical faults when the mechanical health weight > insulation status, and fusing multi-modal data to generate a set of response decisions, such as "locking the switch + starting dehumidification".

[0173] Generate a composite response strategy based on the quaternion encoding for fault root cause tracing. According to the fault root cause in the quaternion encoding, such as "contact wear + harmonic overload", generate a multi-action cooperation strategy, such as "derating operation + harmonic filtering + maintenance warning".

[0174] Generate response actions according to the composite response strategy and the triggered response decisions in the set of response actions, that is, based on the policy priority, such as immediately locking a high-confidence fault, and the real-time state of the device, such as prohibiting opening when the energy storage pressure is insufficient, dynamically select safe and feasible response actions.

[0175] In actual implementation, the three-dimensional state encoding in this application includes the binary identification of the switch position, the percentage value of mechanical health, and the insulation status level code; the quaternion encoding includes the electrical protection matching score, the mechanical action reliability index, the probability value of hidden defects, and the maintenance priority weight; use a fuzzy logic controller to analyze the compatibility constraint between the switch position and the insulation status, and when the insulation status level code exceeds the preset risk threshold, trigger a forced shutdown action priority override mechanism; calculate the attenuation coupling coefficient between the mechanical health and the mechanical action reliability index through a spatio-temporal correlation matrix, and generate a preventive opening action instruction when the coefficient is lower than the dynamically adjusted threshold; when the electrical protection matching score is lower than the first critical value and the probability of hidden defects is higher than the second critical value, execute a three-level response to the fault action:

[0176] a) First-level response: Send an adaptive locking signal to the relay protection device to suppress unnecessary tripping;

[0177] b) Second-level response: Activate the pre-pressurization module of the circuit breaker arc extinguishing chamber and start the vibration suppression device;

[0178] c) Third-level response: Generate an equipment isolation instruction with a maintenance priority weight and a topological reconstruction plan;

[0179] When the mechanical health percentage value exceeds the safety corridor for N consecutive cycles, a progressive shutdown action sequence is triggered:

[0180] 1) First stage: Reduce the breaker operating current to the maintenance mode threshold;

[0181] 2) Second stage: Enable the standby mechanical interlock device to take over the operation authority;

[0182] 3) Third stage: Upload an irreversible locking request to the power grid dispatching master station.

[0183] Finally, the response actions are encrypted and signed through the Trusted Execution Environment (TEE) and stored on the blockchain, generating a tamper-proof operation log containing a three-dimensional state snapshot, a response logic chain, and a timestamp.

[0184] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A method for monitoring the position state of a circuit breaker switch based on video recognition technology, characterized in that, It includes: Collecting real-time images of the circuit breaker through a camera device, judging the physical state of the circuit breaker, and determining a first judgment result; Obtaining real-time data of the relay protection device, recording the operating state of the circuit breaker, and determining a second judgment result; Inputting the first judgment result and the second judgment result into a preset signal response device to generate corresponding response actions; among them, the response actions include: opening action, fault action, and closing action.

2. The method for monitoring the position state of a circuit breaker switch based on video recognition technology according to claim 1, characterized in that, The real-time images include visible light images, near-infrared images, depth thermal maps, and images after supplementary lighting.

3. A method for monitoring the position state of a circuit breaker switch based on video recognition technology according to claim 1, characterized in that, The collecting of real-time images of the circuit breaker through the camera device further includes: Pre-constructing a multi-clarity scene database and performing adaptive optical analysis to determine the clarity state of the current scene; Among them, the adaptive optical analysis includes multi-spectral imaging analysis, dynamic polarization analysis, and interference light source array analysis; According to the multi-spectral imaging analysis, calculating the inter-frame error between the visible light image and the near-infrared feature image during alternating acquisition of the real-time image; According to the dynamic polarization analysis, calculating the real-time polarization angle of the rotatable linear polarization filter array and the annular phase retarder of the camera device; According to the interference light source array analysis, generating a structured light distribution array of the current scene and determining the area to be supplemented with light that does not meet the preset clarity; Based on the inter-frame error, polarization angle, and the area to be supplemented with light, building an image enhancement network for the real-time image and performing enhancement processing on the real-time image.

4. The method for monitoring the position state of a circuit breaker switch based on video recognition technology according to claim 3, characterized in that, The image enhancement network includes an input layer, a data alignment sub-network, a physical feature encoder, a conditional modulation enhancement unit, and a cascaded enhancement strategy unit; among them, The input layer includes image data for receiving the real-time image through a signal channel; among them, the signal channel includes a visible light image channel, a near-infrared image channel, a polarization angle tensor channel, and a structured light mask channel; The data alignment sub-network is used to perform a first enhancement process on the image data through a spatio-temporal alignment network to eliminate the inter-frame error; The physical feature encoder is used to extract different modality features in the image data through a multi-branch parallel structure to generate a multi-modal feature map; The conditional modulation enhancement unit is used to use the multi-modal feature map and the inter-frame error as conditional vectors to determine the image area to be enhanced; The cascaded enhancement strategy unit is used to stack multiple dynamic enhancement blocks and perform enhancement processing on image areas of different scales through the multiple dynamic enhancement blocks.

5. The method for monitoring the position state of a circuit breaker switch based on video recognition technology as claimed in claim 4, wherein, The multi-branch parallel structure includes: a spectral branch, a polarization branch, and a supplementary light branch; Among them, the spectral branch is used to extract and separate the specular reflection feature and the texture feature through channel attention; The polarization branch is used to extract the rotation weight feature guided by the polarization angle of the camera device and suppress the specular reflection feature of the metal through directionally separable convolution; The supplementary light branch is used to determine the image gray deviation feature through mask-guided local contrast enhancement.

6. The method for monitoring the position state of a circuit breaker switch based on video recognition technology according to claim 1, wherein, The judging of the physical state of the circuit breaker and determining the first judgment result includes: Constructing a multi-modal feature fusion network, inputting the visible light image, near-infrared image, depth thermal map, and image after supplementary lighting into a spatio-temporal alignment module, and fusing the texture feature, thermodynamic feature, and polarization suppression feature through a cross-modal attention mechanism to generate a high-confidence circuit breaker state feature map; Based on a preset physical constraint state decision tree, perform hierarchical judgment on the key areas in the high-confidence circuit breaker state feature map: First-layer judgment: Identify the contact closing / separation state according to the near-infrared reflectivity and temperature rise gradient in the contact area; Second-layer judgment: Detect mechanical jamming or deformation based on the integrity of the mechanical interlock device contour in the image after polarization suppression; Third-layer judgment: Determine the abnormal surface pollution level of the insulating component through the abnormal ratio of the local contrast between the supplementary light area and the original image; Output a circuit breaker physical state report that fuses the multi-modal judgment results, and determine the first judgment result. The first judgment result includes a three-dimensional state code of the switch position, mechanical health, and insulation state.

7. A method for monitoring the position state of a circuit breaker switch based on video recognition technology according to claim 1, characterized in that Acquire the real-time data of the relay protection device, record the operating state of the circuit breaker, and determine the second judgment result, including: Based on the relay protection device, configure a multi-source data acquisition device, and acquire the real-time electrical parameters of the relay protection device and the multi-dimensional operating state data of the circuit breaker; Construct a multi-source data fusion model with time series alignment, and perform dynamic collaborative analysis on the real-time electrical parameters and multi-dimensional operating state data to obtain a fault-action mapping matrix; Perform multi-dimensional anomaly detection on the fault-action mapping matrix through an autoregressive anomaly propagation model: Generate a second judgment result including the root cause tracing of the fault, and output it in a quadruple coding format.

8. The method for monitoring the position state of a circuit breaker switch based on video recognition technology as claimed in claim 7, wherein, The real-time electrical parameters include the waveform of the protection action signal, the amplitude of the fault current, the impedance of the tripping circuit, and the harmonic distortion rate; The multi-dimensional operating state data includes the mechanical vibration spectrum of the circuit breaker, the closing and opening coil current curves, the pressure value of the energy storage mechanism, and the wear degree index of the switch contacts.

9. A method for monitoring the position state of a circuit breaker switch based on video recognition technology as claimed in claim 7, wherein, The autoregressive anomaly propagation model includes a deviation warning layer, a risk identification layer, and a fault location layer; among them, The deviation warning layer is used to compare the deviation between the moment when the tripping instruction is issued and the actual opening time of the circuit breaker, and trigger a mechanical response delay warning when the deviation exceeds the preset threshold; The risk identification layer is used to analyze the non-linear relationship between the number of actions of the protection device and the contact electrical life loss curve, and identify the hidden fault risk caused by abnormal frequent actions; The fault location layer is used to locate the fault location of mechanical jamming or spring fatigue based on the coupled analysis of the vibration spectrum entropy value and the energy storage pressure volatility.

10. A method for monitoring the position state of a circuit breaker switch based on video recognition technology as described in claim 1, characterized in that, Input the first judgment result and the second judgment result into a preset signal response device to generate corresponding response actions, including: Receive the three-dimensional state code of the first judgment result and the quadruple code of the second judgment result; Construct a multi-modal response decision model, dynamically fuse the coded data, and generate a response decision set; Generate a composite response strategy based on the quadruple code of the root cause tracing of the fault; Generate response actions according to the composite response strategy and the triggered response decisions in the response action set.

Images